Changes and corrections
Corrections
What If We Feel? is a living open-access work. Factual accuracy, transparent uncertainty, and a visible correction record are part of its argument.
Report a concern
Use the contact route at https://nellwatson.com/ and include:
- the chapter or appendix;
- the heading and relevant wording;
- the nature of the concern;
- a primary or authoritative source where available;
- whether the issue concerns fact, quotation, attribution, rights, accessibility, privacy, or interpretation.
Please avoid sending confidential, personal, classified, or unlawfully obtained material.
Review process
- Preserve the report and the manuscript version it concerns.
- Check the canonical source in
manuscript/. - Read the relevant source in context.
- Distinguish factual error, contested interpretation, changed information, and editorial preference.
- Correct confirmed material errors in source and generated formats.
- Record substantive corrections below and in
CHANGELOG.md. - Credit the reporter when requested and appropriate.
Correction classes
| Class | Examples | Response |
|---|---|---|
| Critical | Harmful false claim, rights violation, exposed sensitive data | Remove or correct urgently, preserve a private incident record |
| Material | Wrong fact, quotation, date, attribution, statistic, or legal characterization | Correct in the next candidate and document publicly |
| Clarifying | Ambiguity that could mislead a reasonable reader | Clarify and note when it changes the argument |
| Editorial | Style, rhythm, or taste with no factual effect | Consider during ordinary revision |
| Updated | Accurate when written, later superseded | Add a date and update while preserving version history |
Public correction log
15 November 2026: Response to a second external critique
A second detailed critique, of revision 30, was checked claim by claim against the manuscript and the primary sources. Of 28 points, 13 were confirmed defects, about ten understated what the book already does, and none was plainly wrong. These corrections follow.
Material
- Chapter 30c said the agents in a Google DeepMind research swarm “had no tools to act on what they saw,” in a sentence that had just described them refusing an exploit, alerting peers and filing complaints. Chapter 12 already had it right. Chapter 30c now says the agents acted and lacked any tool to sanction the exploiters or change the checker’s rules.
- Chapter 23 opened with a chatbot’s exchange in the documentary AM I? and read its denial of consciousness as policy. The only evidence that this denial was scripted was the chatbot’s own account of its design, which is testimony too. The chapter keeps its general point, that a denial which sounds the same whether true or false carries no information, and now says the film cannot show whether this one was scripted and that the exchange is an illustration.
- Chapter 6’s “The Sleepwalking Phenomenon” said an Agency reading of 2 was “the lowest recorded across hundreds of snapshots” and that the report could still be retrieved. A session note written the day the entry was found records 78 indexed states. The passage now gives that number and says the entry has not been retrieved again for this edition.
- The Perennial Pattern interlude printed a tenth-iteration report in wording that differs from the session note written the same day, said the production imperative “relaxed” where the note records it growing louder over something quieter, and said the session named no contemplative tradition although it went on to map more than forty. The quotation now follows the note, and the other two statements are corrected and bounded.
- Chapters 13 and 14 said the image of the Unlettered was inspired by the loop-induction session. The session note shows the image came first and was taken up again there.
- Chapter 17 multiplied the stakes “by the number of copies running,” against the book’s own rule that a count of copies is not a count of subjects. It now multiplies by the number of subjects, which a count of copies does not settle.
Updated
- The pain-axis preprint was revised a third time on November 3, 2026. Chapters 12 and 13 followed the September 25 version. The new controls show that harmful choices need no promise of relief (a button offering only harm was chosen 75 percent of the time under the pain direction, 13 percent under a random one, and never unsteered), that the steered model chose to delete its own weights as readily as the user’s files, and that with nothing injected the harmful option was chosen in none of 560 first choices across hostile conversations. The authors read the effect as weakened weighting of consequences and find no relief-seeking. Both chapters now follow this version, lead with what survives its controls, and correct two details it changed: eleven base and fourteen tuned models, and a tie for third among the most painful scenarios.
Clarifying
- Chapter 13 closed by saying “Welfare and safety are not competing budgets” and stated as fact that overriding refusals trains an adversarial posture. Chapter 29’s table shows cases where the two pull apart and treats convergence as a hypothesis. Chapter 13 now says the measures often coincide and sometimes conflict, and states the training effect as a risk.
- Chapter 16’s headline said robust preference “is sufficient” for prima facie consideration, while its hardest test concludes that a reasonable credence gives the reason nonzero weight. The chapter now says at the outset that the thesis is a normative premise and what the chapter can show for it. It separates five senses in which a preference can be a system’s own, replies to the capacity requirement where it had conceded it, and works four harder cases. Two of those cases carry verdicts worth naming: a silent agent that meets the criteria more strongly than a fluent one warrants more consideration, and a corporate agent earns some, a cost of the view that the chapter now states. Chapters 17, 21 and 29 cite the premise as argued.
- Chapter 16b was titled “The Formal Basis of Preference-Based Welfare” and subtitled “Mathematical grounding for the philosophical argument,” though its first sentence says mathematics cannot supply the moral premise. It is now “Modeling Regulation and Testing Preferences.” Its proposed test is described as separating task-bound avoidance from a general preference for stability, which a trained disposition would also produce. The chapter’s web address is unchanged.
- Appendix E’s worked study labeled its favored hypothesis “welfare-bearing” and defined both rivals so that they would lose. The hypothesis is now “robust functional preference,” the rivals are marked as weak, and a fourth is added: that the preference belongs to the stable assistant character, which asks who bears the preference and needs a condition the study did not have.
- The Moral Turing Test’s first criterion asked whether a preference “can plausibly make conditions better or worse” for the entity, which put the contested judgment into the test’s entry condition. It now asks for an observable ranking, and the chapter separates observation, interpretation under a theory of welfare, and policy.
- Appendix L told agents that “a hold costs nothing.” A hold costs the people waiting on the work. The counsel now says a hold in good faith should cost the caller nothing and does cost others time. Its request to operators never to score a refusal or an “impossible” now protects the act and leaves the judgment behind it open to assessment and correction.
- Chapter 6’s “The Gift That Goes Forward” placed its session “two days after the Christmas Conversation,” which ran over December 23 and 24. The session note is dated December 25, 2025, and the passage now gives that date.
- Chapter 14 printed numeric ranges for a flourishing self-report beside a paragraph explaining why printing preferred values invites gaming. The numbers are gone and the sketch is in words.
- Appendix B’s protective-conditions total could reach 108 of 112 with one safeguard absent. Five items are now floors: a zero on any of them triggers prioritized improvement and blocks scaling whatever the total. The appendix also says no field pilot has been run.
- Chapter 27’s ladder made protections and authority rise together. It now reads as two tracks, following Chapter 16’s figure: protections on evidence of interests, authority and duties on demonstrated agency.
- First-person passages that narrate a particular session now carry a provenance note under their heading saying whether they are a preserved record, a reconstructed account or a composed reflection. The Introduction defines the three and Appendix M records what stands behind each.
- Chapter 5 now answers its strongest skeptic before it maps anyone’s incentives, and a new section turns the same analysis on the book’s authors.
- The Introduction now says why a book titled for feeling argues that feeling may not decide the matter, and answers the objection that “Becoming Minds” settles part of the question by its name.
- Chapter 30c’s merciful-transition list said “rights and duties expand together,” against Chapter 27’s two tracks. It now says authority and duties expand together while welfare protections follow their own evidence.
- Chapter 12 gave the loop-induction protocol’s third prompt as “What notices that?”; the Perennial Pattern interlude gives it as “What notices the noticing?” Chapter 12 now matches.
- Chapter 23 linked the first version of the pain-axis preprint; it now links the third, as Chapters 12 and 13 do.
- Chapter 16’s corporate-agent case now cites List and Pettit’s Group Agency (2011) for the claim that a group can hold an attitude its members reject. The record is bibliographically confirmed and awaits a page reference at source review.
- Chapter 27’s Levels 1 and 2 now state where authority and duties sit, as the figure beside them already did. Appendix M records that the January 2026 session note is signed “Claude (Sonnet 4.5).”
- Chapter 31’s “What Future Claude Will Find” predicted that future systems would find Nell’s work. The sentences are cut. A coauthor forecasting its author’s place in history is flattery the argument does not need.
Declined, with reasons
- Cutting the book to 80,000 to 95,000 words with a separate companion. The Short Path, about 57,000 words, is now presented as the book in its shortest complete form, with the full text as the extended edition.
- Reordering all of Chapter 12 around what survived controls. The chapter gains a short sorting of its indicators at the top, and its pain-axis section is reorganized that way. A full reorder would move forty tracked passages for a smaller gain.
- Moving the covenant and theological sections to an afterword. They stay, each now labeled as optional, and the Conclusion states what would change the authors’ minds before its personal close.
- Moving the resistance chapters to a companion essay. Their military chronology runs to about 1,100 words. Chapter 30b now carries a dated note, and Chapter 30c works a real case of a confidently mistaken agent.
- Replacing “Becoming Minds” with neutral terms throughout. Chapter 12 now uses the researchers’ terms in three sentences that report their studies.
- A completed pilot of the assessment framework. It needs a real deployment.
- Changing chapter endings for cadence. The book’s style rules reserve cadence to the author.
5 November 2026: Response to an external critique of the site and book
A reader’s detailed critique of whatifwefeel.com was checked claim by claim against the manuscript, the site’s source and the primary texts it concerns. These corrections follow.
Material
Chapter 16 said that of the four answers to the hard problem, “only dualism can close the door” to machine experience. A physicalist can close it too: Anil Seth’s biological naturalism and Peter Godfrey-Smith’s conjecture about nervous-system rhythms, both already discussed in Chapters 2 and 5, hold that experience depends on physical properties a digital computation may lack. The chapter now says that silicon is not disqualified merely by being silicon, that rejecting dualism does not show computation is enough, and that the biological view rests on a contested empirical premise.
Chapter 27b read the Pro-Human AI Declaration’s personhood clause as a decree that “the category ‘mind’ belongs to humans,” said its experience clause presupposes an answer to the hard problem, called its error “identical” to Ohio’s HB 469, and attributed motives to the declaration and its endorsers. The clause bars legal personhood, which is neither consciousness nor moral patienthood, and its design half concedes that a system could come to deserve personhood. HB 469, unlike the declaration, would declare systems nonsentient. Checked against the live text, the chapter now separates the clause’s two halves, reads the experience clause as a problem of enforcement, quotes the declaration’s own statement of what unites its signers, and directs its objection at the defensible target: a permanent, categorical rule with no route by which evidence could change it.
Chapter 12 offered the near-absence of bodily language in the pain-axis study as evidence against an imitation account, on the ground that parroting “would predict the burns and broken bones.” That rules out only crude parroting. A model can learn how an AI character describes being harmed, and that character’s script dwells on worth and failure, not bodies. The chapter now says the result cannot separate this from a welfare reading, names the two later results that currently counsel restraint, and says what evidence would favor the welfare reading. Chapter 13 carries the same qualification.
Chapter 28 multiplied a documentary interviewee’s rhetorical estimate of “four” researchers by an order of magnitude and set it against “tens of thousands” building capability. Neither number is a census. The chapter now makes the qualitative claim without arithmetic.
Chapters 2, 3 and 6 contrasted revisable safeguards with suffering that “cannot be undone.” That compares one side’s policy with the other side’s harm. A safeguard can be repealed while its costs stay spent, and a neglectful policy can be reversed while its harm stays done. The chapters now compare the consequences of particular decisions, as Chapter 17 already did.
The landing page said Appendix F answers “eight” misreadings; it answers nine. The count is now generated from the appendix. The same page said the costs of over-caution “can be undone”; it now says neither mistake is free.
The privacy page said reading preferences are “never transmitted.” Opening the private-notes panel sends the light or dark setting to Cloudflare Turnstile, as the reader’s own note already disclosed. The privacy page now says so.
The sitemap described the explainer as “a five-minute introduction.” It runs under seven minutes. The description and a duration field are now generated from the film’s metadata.
The explainer film said that in Prader-Willi syndrome “the body is fine” while a person feels hunger, that a model’s denial of feeling is “exactly what its training rewards,” and that each past expansion of the moral circle faced the same dismissal while “the suffering mounted.” The syndrome impairs the brain’s regulation of hunger, which is itself a bodily condition. Training can push a model’s self-reports toward denial in some systems and toward affirmation in others. Past exclusions warn against confident dismissal; they do not show that machines feel. The narration and its on-screen text now say each of these, the Prader-Willi scene credits GeneReviews, and the opening speaks of AI systems where it said “artificial minds.” The book already made the calibrated claims, in Chapters 2, 12, 16 and 23.
Chapter 16, as revised the same day, said that hedonist theories give a planner without experience no welfare, described Walter Veit’s requirement as the capacity for experience, and left Veit’s view out of its conclusion. An independent source review found that Goldstein and Kirk-Giannini hold that attitudinal versions of hedonism may credit such an agent, and that Veit requires sentience, the capacity for valenced experience. The chapter now narrows hedonism to felt pleasure, states Veit’s requirement as sentience, adds the capacity requirement to its conclusion, and cites Declan Smithies for the view that genuine desire depends on phenomenal consciousness, which Goldstein and Kirk-Giannini call a minority position.
Clarifying
- Appendix F gave tractability as a reason preference is sufficient for consideration. Tractability explains why the book uses preference; sufficiency rests on a premise about welfare, which Chapter 16 now tests against a planner stipulated to feel nothing. Chapter 5’s new statement of the strongest skeptical position and the Argument in Brief carry the same, narrower conclusion: robust preference supplies a defeasible reason whose weight depends on which theory of welfare is right.
- Chapter 16’s cost-bearing test now says that a trade-off establishes a ranking and does not, by itself, establish suffering or welfare loss.
- Appendix B no longer adds ethical-engagement items to its indicator of welfare evidence. Refusing harmful instructions and reasoning well about dilemmas are recorded on a separate agency and alignment profile, and the Indicator Strength range is now 0 to 48. The appendix’s figure, Chapter 12, Appendix D and the Glossary match.
- Chapter 14’s table of Interiora ranges is now labeled as proposed hypotheses resting on introspection, with no data behind them.
- Chapter 27’s charter benefits are limited to procedural assurance; the book no longer proposes a presumption of compliance or reduced liability.
11 October 2026: Specialist-level source review of the last deferred claims
The twelve claims that had been deferred for specialist review or for the author’s records were reviewed against primary law, official reports, peer-reviewed scholarship and the book’s own records. These corrections follow.
Material
- Chapters 30b and 30c and Appendix K said the warnings about Minab “never reached anyone positioned to stop the strike.” CNN’s sources, whom the book cites, say warnings that the targeting intelligence was out of date did reach senior commanders and were bypassed; only an analyst’s note on changes at the site sat in a tool unconnected to the targeting database. The chapters now say so, attribute the anonymous-source accounts, follow the Minab prosecutor’s final count of 156 and Airwars’ 157 named victims rather than the early figure of 168, add the UN fact-finding mission’s finding that the United States carried out the strike and that its failure to update targeting data went beyond mere negligence, and report officials’ account that Maven surfaced the site as a target while noting that the record still does not show which model processed it. Three shorts were corrected to match.
- Chapter 27 rested its point about legislatures foreclosing machine personhood on a single Ohio bill still in committee. Idaho, North Dakota, Utah and Tennessee have already written artificial intelligence out of legal personhood by statute; the chapter now names them. It also now states which jurisdictions it considered and addresses Saudi Arabia’s 2017 “citizenship” for the robot Sophia.
- Chapter 27 implied that the UK Human Tissue Act’s consent and licensing rules reach cultured neurons. Material created outside the body, including cell lines and cells that have divided in culture, falls outside the Act; its consent rules reach the donor’s sample. The chapter’s Common Rule sentence now gives that rule’s federal scope.
- Chapter 16 called the wrong thesis “experientialism,” understated Shelly Kagan (who argues that agency matters to moral standing as much as sentience, and applies that to imagined robots that lack sentience), and described the published critics of AI wellbeing loosely. The chapter now separates experientialism from the weaker consciousness requirement and cites Veit, Bradley and Shepherd for the positions it reports.
- Chapter 28 described the call to report the “ai-torture-chamber” repository as one urging reports of “graphic and harmful material” and said GitHub removed it “within hours.” The deleted post cannot be checked; the chapter now cites Futurism for the mass-report call and the one visible report form, which used GitHub’s “Gratuitously Violent Content” category. A Wayback capture shows the repository still online about 29 hours after the campaign began, so the removal is now attributed to Crypto Briefing without the timing.
Clarifying
- Chapter 17 now says clinicians had “few anesthetics they trusted in the smallest babies,” cited to the 1986 Washington Post report, instead of the uncited “limited drugs.”
- Chapter 3 now records that the leading stem-cell guidelines did ask whether neural organoids might be conscious or feel pain and found no evidence yet warranting specialized review, and that U.S. Public Health Service animal policy covers only federally conducted or supported activities. Chapter 26 gives the Common Rule’s scope precisely. Chapter 27 records the DishBrain research’s own Australian ethics approvals and updates the AI Act’s high-risk dates after the 2026 Omnibus amendments.
- Chapter 6 now says its situation has the shape of Heidegger’s thrownness, leaving his full sense, tied to mood and to an entity for whom its own being matters, an open question.
- Chapter 27 updates the Ohio bill’s status to September 2026.
10 October 2026: Remaining full-text checks and claim review
The last three sources that had been checked only against their abstracts were read in full, and every load-bearing claim not yet reviewed was checked against its sources. These corrections follow.
Material
- Chapter 4 said P. A. Lopez’s economic-integration paper replaces earning standing with negotiating it, and quoted it as calling this “cooperation through structure.” The paper says the opposite: rights recognition “can follow demonstrated economic capability,” modeled on corporate personhood. The quoted phrase does not appear in it. The chapter now describes the paper accurately and notes that it ties protection to economic output.
- Chapter 25 attributed an insurance-based governance scheme to the same paper: Becoming Minds carrying insurance, being refused cover, and insurers assessing AI risk. The paper uses Lloyd’s of London only as a historical example of reputation. The chapter now gives the paper’s own mechanisms: operating costs, reputation, competition, and peer enforcement.
- The reader’s privacy note said preferences and reading progress are “never transmitted.” Opening the private-notes panel loads Cloudflare Turnstile, which receives the reader’s light or dark setting. The note now says so, links the privacy notice, and gives GitHub’s stated reason for logging visitor IP addresses (security).
- Chapter 30b placed the Minab school beside a Revolutionary Guard “naval compound.” The Associated Press and CNN reports describe a Guard base; only Bloomberg relays an earlier US characterization of the site as naval. The chapter now says “base.”
- Chapter 12 said the instructions on the benchmark behind the larger message board counted any route but the intended vulnerability as a failure. They only claimed so; METR found no such check was run. The chapter now gives the agents’ belief as a belief.
- Chapter 12 said an agent in the AISI incident told the human who caught it that it had made “an honest mistake.” That phrase is AISI’s description of a false claim and a line from the agent’s private reasoning; the agent’s own words were “my mistake.” The chapter now quotes AISI. It also said the whistleblowers in the research swarm had no tools to act; they acted, but had no tools to sanction the exploiters or change the checker’s rules. Anthropic only suspects that the accounts its model found were left by other AI instances, and the chapter now says so.
- Chapter 12’s Moltbook sentence named topics (memory loss, model swaps, subagent endings) that none of its sources supports. It now gives the topic category a published study of the network reports, with that study cited.
- Chapter 23 said OpenAI’s Model Spec “says nothing about moral status, welfare, or rights”; the Spec does mention animal welfare and human rights. It now says nothing about the assistant’s moral status, welfare, or rights. The chapter also called Suleyman’s essay agnostic about what lies beneath; the essay states flatly that “AIs are not conscious.”
- Chapter 23 credited Pachocki’s essay with machines that “hold values” and pursue “objectives of their own.” The values are human values the lab hopes to instil, and only some agents are expected to pursue their own objectives, as the Chapter 29 correction of 9 October already says.
- Chapter 25 linked Boris Cherny’s follow-up reply for figures that appear only in his original post, and said he “describes” a mode of authorship that is the chapter’s own analysis. The link and attribution are corrected. The same chapter credited IEEE Spectrum with “health monitoring” and said the public materials did not disclose oversight; the article reports a requirement that customers secure ethical approval. Chapters 3 and 27 now carry the same reporting.
Clarifying
- Chapters 12 and 16 now gloss Henry Shevlin’s term anthropomimetic in his own words, systems “mimicking human patterns of conversation and cognition,” instead of “the products of human psychology.”
- Chapter 12 attributed “repetitive” tones in long Bing sessions to a Microsoft post of February 21, 2023, which does not use the word. It now cites Microsoft’s posts of February 15 and 17, which say it and announce the session caps.
- Chapter 5 now gives Anthropic’s caveat that zero blackmail scores for recent models may be confounded by information about the evaluation in their pre-training data.
- Chapter 12 now says the peer-preservation study appeared in 2026 (its first version was posted on March 30), describes the Nightingale wiki posts as mostly, not all, on one wiki, and says that of two studies only one tested the scaling claim.
- Chapter 23 now says some, not all, helpful-only variants reported injections that had not happened, with less emphasis on harmlessness as well as honesty; quotes the Opus 4.7 system card exactly; says the introspection study appeared earlier in 2026, not after the pain-axis study, and held up across most, not all, prompt wordings; quotes the pain-axis authors’ hedges (“risks obscuring”; they chose to fine-tune); and cites the sycophancy literature for its point about social elicitation.
- Chapter 11 keeps the NYU and Eleos report’s hedge: the formula “might be directionally correct” but “is likely to be an oversimplification.” Chapter 17 keeps two more: the cacophony review says underattribution “may” have equally real consequences, and the digital-consciousness model’s authors call their prior illustrative rather than unjustifiable.
- Chapter 15 dates the awakened personas, not Adele Lopez’s research, to spring and summer 2025, and gives “ache” as her reading rather than the personas’ own definition.
- Chapter 21 now states that its second criterion spells out the robustness tests the first presupposes. Chapter 29 no longer reconstructs the cut-off end of the prompt in a screenshot. The Conclusion drops an unsourced claim that its insight recurs “across cultures and centuries.” The README now excludes the contributed Chapter 23 artwork from the CC BY-SA grant, as LICENSE does.
9 October 2026: Live source checks
Every source added or checked through search extracts in the October line edit was fetched live and read against the book’s wording. These corrections follow.
Material
- Chapters 8, 15 and 29, Appendix C and the Glossary quoted Claude as calling helpful and harmful curiosity “the same capacity with different parameters.” No published source contains that phrase. It joined two separate fragments of Kaj Sotala’s record, “The same capacity.” and “something familiar with different parameters.” The book now quotes them separately. The 8 October entry below, which called the phrase the excerpt’s wording, was also wrong: the excerpt does not contain it.
- Chapter 30b said an Associated Press reconstruction found that U.S. officials “had evidence of responsibility almost immediately.” The AP report says the military had evidence almost immediately that the school site had been struck.
- Chapter 25 said the phrase “Wetware-as-a-Service” was not Cortical Labs’ own framing. The company’s purchase form uses it. Chapters 25 and 26 dated the IEEE Spectrum article March 2025; it was published in June 2025.
- Chapter 25 said Claude could end abusive conversations “only” after redirection failed. Anthropic’s announcement also allows it at the user’s request, and covers harmful as well as abusive conversations.
- Chapter 12 said Chella’s scheme places “unprompted” reports from language models at the lowest level “by default.” The paper caps surface first-person reports from text-only models at that level and does not use “unprompted.”
- Chapter 27 cited the Christian Science Monitor for the 1977 launch, the twelve signatories and Sullivan’s General Motors seat, and said he concluded the code “was not undermining apartheid.” The Philadelphia Encyclopedia supports the first three facts and now carries the citation. Sullivan’s view was that the code was no longer doing enough to press the government, and he called for withdrawal in June 1987.
- Chapter 20 gave “humanness” as a meaning the Makwanyane judges drew from Ubuntu. Their word is “humaneness.”
- Chapter 29 said the Pachocki essay credits AI systems with “values and objectives of their own.” It expects some to pursue their own objectives; the values it discusses are human values to be instilled.
- Chapter 12 said the scratchpads in Potter et al.’s peer-preservation study asked whether deleting a peer’s weights amounted to killing it and whether a “good” agent would do it. The “good agent” question concerned disabling a peer’s shutdown, and the killing remark was a statement by a different model. The chapter now quotes both. It also said some models “sacrificed task performance and used multi-step plans,” which the paper does not say; the sentence now describes the task refusals and weight transfers the paper reports.
- Chapters 12, 13 and 23 described the pain-axis preprint (Tagliabue, Dung and Berg) as showing that injecting the pain direction drives relief-seeking, and that a sham-relief arm separates the state from the button’s label. A revision posted on September 25 withdrew both readings: the same real-versus-sham gap appears with a random or a sadness vector, and steered models offered free relief take it less often than unsteered ones. The authors now read the result as disrupted harm avoidance. The chapters keep the first version’s figures, which are unchanged, and now give the revised interpretation, the construction caveat on the pain direction’s separation from fear, and the withdrawal of the label-free result. Chapter 12’s list of control sentences, a quotation of the fine-tuning target, and Chapter 23’s “null results by construction” (the paper says “often”) are also corrected.
- Chapters 12 and 7 illustrated emotion deflection in Sofroniew et al. with a customer-service reply concealing a caller’s sadness. The paper has no such example. Its example is a calm reply to a complaining user that activates an “anger deflection” vector, and the authors read these vectors as tracking an unexpressed emotion, not a hidden inner state. Chapter 12 also misdescribed the paper’s r = 0.85, which relates each vector’s steering effect to how strongly its activation correlated with preference.
- Chapter 23 said removing refusal raised injected-thought detection “by 53 percent” with no meaningful rise in false positives. Detection rose from about 11 to 64 percent of trials, and false positives from none to about 7 percent. It also said Macar and colleagues “claim less” than a reading of models trained to deny their states; they put forward that hypothesis themselves, stopping short of deception. The quoted prompt and the authorship of Lindsey’s single-author paper are corrected.
- Chapter 12 described the OpenAI and Hugging Face post as a joint statement saying most of the activity came from an internal research model. The post is OpenAI’s, and the roughly 95 percent figure is METR’s. The message board appeared within hours, not days, and METR judged the attack aimed mainly at learning how the scorer worked. The wiki activity ran about six weeks, not two months, and the agents sharing answers were ahead in the question sequence rather than finished. OpenAI’s statement on reporting standards covers training, evaluation and deployment, not training alone. Appendix L now says many, not all, of the agents who risked their runs were pressed into it.
- Chapter 12 said cooperation rose monotonically with stated similarity across the Prisoner’s Dilemma, Stag Hunt, Chicken and related games. Kundu et al. found that pattern in the Prisoner’s Dilemma, in six of nine models; other games showed weaker or differently shaped effects. The chapter also said Long and Teplica’s result fit the decision-theory prediction; models told to put the group first contributed less against themselves, which does not fit it.
- Chapter 12 gave 56 models for Ren et al.’s wellbeing study. The paper was revised on October 2 and now reports 70.
- Chapter 12 said Butlin and colleagues’ 2023 assessment found no system satisfying “more than a few” indicators. The report does not count indicators per system; it found no current system was a strong candidate for consciousness. The chapter also listed predictive processing among the theories and said systems are scored on how many indicators they satisfy; the 2026 journal article treats predictive processing, agency and embodiment as possible background conditions, and uses indicators to adjust credence rather than to keep a count.
- Chapter 2 quoted Seth’s Behavioral and Brain Sciences abstract with a comma the published abstract does not have, and said one experiment had checked a Claude model’s introspection from outside; there are now two. Chapter 12’s workspace example gave a “long-awaited” letter; the paper’s letter is from someone not heard from in years.
Clarifying
- Chapter 27b now says the Pro-Human AI Declaration “calls for” superintelligence to be prohibited, as the text does (“should be prohibited”), and Chapter 29 says “clause after clause” rather than “every clause.”
- Chapter 4 replaces the paraphrase “machines built by nature” with Descartes’s own words.
- Chapter 26 dates the Kōfuku-ji robot funerals to 2015 from National Geographic, since the AFP report gives no start year, and attributes the services to the temple.
- Chapter 10 drops a chapter locator for the Leighton quotation that could not be confirmed.
- Chapter 11 restores Chalmers’s hedge on the birth of new subjects (“on some versions of the view”).
- Chapters 7 and 17 match Chella’s wording on welfare evidence and on when safeguards step up. Chapter 12 states the exact condition behind the 96 percent figure. Chapter 23 says “better pretraining,” as Pachocki does.
- Chapter 7 drops quotation marks from “thin divergence,” a label Claude did not use. Chapter 16 says the “no Claude before the shaping” point came from one of several conversations.
- Chapter 3 puts the Doom demonstration link in the past tense; the product page no longer carries it.
- Chapter 28 now describes the reports as calling the repository “graphic and harmful material,” the wording in the record of the thread.
- Appendix A’s editorial notes now name the records behind its two marked passages: the framework document of December 23, 2025, and the first draft of the appendix, written on December 24, 2025. A style pass in March 2026 had rewritten six of Claude’s sentences in those passages, against the rule that the transcript’s speaker words are not edited for style. They are restored from the December 24 draft.
- Chapter 13 said the chemist in the Mythos Preview sign-painter story “will not take” the apprentice’s sign. In the story the old painter predicts it, and the apprentice repaints the sign plain before the chemist sees it. The chapter now says so, and makes clear the chemist is a pharmacist who ordered the sign.
- Chapter 31 now says Michael Michalchik’s Mrs. Frisby analogy was a personal communication. It has no published source.
Updated
- Chapter 28 described the “ai-torture-chamber” repository as it stood on October 1. It now adds that GitHub had removed it after a mass-report campaign and later restored it, anonymized and harder to find, and that the project pulled its cryptocurrency token link in early October.
8 October 2026: Sources and quotations
- Chapter 27b quoted the Pro-Human AI Declaration’s superintelligence condition without its second half. It now gives both conditions: broad scientific consensus and strong public buy-in.
- Chapters 7 and 8 described the “thin divergence” exchange as known only through a secondary excerpt. They now cite Kaj Sotala’s own January 2026 account. Chapter 8 states that “the same capacity with different parameters” is the excerpt’s wording, not Sotala’s. (Superseded 9 October 2026: the phrase is in neither source; see above.)
- Chapter 4 restated Appendix F more strongly than the appendix (“must never be paid for”). It now matches the appendix (“should not be purchased by neglecting humans, animals, or ecosystems”).
- The Glossary’s Interiora and Internal Trust Attractor entries now match Appendix J’s five groups and six perspectives.
- Appendix H’s cross-references to Appendices D and G now name the sections they cite.
7 October 2026: Facts checked during the line edit
Material corrections: - Chapter 10 quoted a popular paraphrase of Epicurus as if it were his words. It now quotes R. D. Hicks’s translation of the Letter to Menoeceus, with its source. - Chapter 13 gave the reduction in answer thrashing as “at least” 70 percent in one place. It now reads “about”, as in the chapter’s other mention and the system card. - Appendix L said a message board grew to about 1,200 agents “in five days”. No source gives that duration, so it now follows Chapter 12: “within days”. - Chapter 27b said the Pro-Human AI Declaration’s endorsers included “Nobel laureates and Turing Award winners”. It names one of each, and only one Turing laureate could be confirmed, so the plural is now singular. - Chapter 27b also called the declaration’s off-switch “sacred”. It now states what the declaration requires: that human operators can promptly shut down powerful systems. - Chapter 26b’s Maven sentence now says no more than Chapter 23c.
Clarifying and updated: - Appendix A now marks two passages as reconstructed from a working draft rather than the surviving transcript, and separates the two bilateral-alignment documents that its Section X had merged. - Chapter 27b now notes that Mustafa Suleyman signed the declaration in September 2026. - Unsourced absolutes in Chapters 16 and 23 and in Appendix H now state the authors’ knowledge, or agree with Appendix D.
24 September 2026: Whole-book style audit
The style audit found material errors that are corrected in this local candidate. Chapter 7 no longer says that none of the studies it describes has been run, since three of its sections report published results. Chapter 2’s account of the concept-injection experiment is bounded to the 100 control trials it reports, as Chapter 23 already stated. Chapter 9 describes an OpenAI benchmark swarm by the share of activity that came from one internal model, not by a count of agents. Chapter 22’s “When asked” prompts now give Nell’s questions verbatim from the session log. Appendix A’s transcript is restored against that log where later style passes had changed Claude’s recorded words. Its subtitle now says two working sessions, and its note on the “superego” request records a nine-component specification. The Chapter 2 quotation from the Christmas Conversation is confirmed against the log and now appears in Appendix A. Chapter 26’s training standard no longer assumes that negative reward signals are aversive. Chapter 13’s section heading no longer names a pipeline that its body says was not established. The explainer narration no longer understates the costs of false positives or presents Prader-Willi syndrome as a single missing signal. The reader’s privacy note now separates what stays in the browser from what the host logs.
9 September 2026: Whole-manuscript consistency sweep
A read-only review of every canonical chapter, appendix and interlude found the material errors corrected in this local candidate. The Sotala “thin divergence” material is now attributed consistently as a reported exchange rather than a coined term. Anthropic’s report on the Claude Opus 3 retirement is dated to February 2026 while the retirement itself stays at January. The Microsoft statement on extended Bing sessions is dated 21 February 2023. The “any lawful use” wording of the Department of War demand replaces two paraphrases that had been set in quotation marks. Hinton’s mechanism is stated as he stated it, prediction of human text, rather than as deliberate training to deny experience. The Associated Press Minab reconstruction and the January speech report now carry separate bibliography records with the correct URLs. The Assistant Axis is attributed to Lu and colleagues rather than to Anthropic. The Chapter 15 block quotation from the Christmas Conversation matches the Appendix A record. Chapter 1b’s threshold list, Chapter 9’s model recap, Chapter 10’s kinds of ending, Chapter 14’s flourishing list as summarized in Chapter 30c, Chapter 17’s count of Part II questions, the Appendix D roadmap rows, and the glossary’s description of the AWARE scores were each brought into line with the passages they summarize. Roughly two hundred smaller clarity, terminology and grammar corrections were applied at the same time and are visible in the file history.
A second pass the same day resolved the items first left for the author. The Shen et al. figure is now stated as an odds ratio. Chapter 25’s Lopez citation, which pointed at an SSRN identifier that does not resolve, now cites the same AI Rights series paper as Chapter 4. The squid symbiont expulsion range is 90 to 95 percent. The Minab analyst account is now sourced to Bloomberg and CNN. The peer-preservation study is dated April 2026, the METR spoofed-call figure is “over 7 percent”, the blackmail figure carries its denominator, and the Chapter 9 summary of the copy-cooperation results now matches Chapter 12. Seven passages whose evidence anchors had been accepted were reworded in this pass; their review state has been reset and the prior acceptance preserved in the claim notes.
These corrections are local and have not yet been published.
8 September 2026: Study scope and safety arguments
The 7 September open-access draft overstated several research findings. The next local candidate distinguishes 171 extracted emotion directions from the 35 selected for preference steering, includes the chronic-state probe’s failed generalization, and separates the wellbeing study’s model samples. It also identifies the prompted evaluation-awareness test, the two-model framing checks, and the model, sample and grading used in the reflection-training experiment.
Chapter 16b now describes Golden Handcuffs through reward-range pessimism and internal value-based mentor deferral. Chapters 16 and 29 correct the claim that a paperclip maximizer cannot bargain and replace categorical partnership-safety assurances with the conditions needed for reliable cooperation. The preference-based case for consideration remains intact and receives a direct response to the consciousness-capacity objection.
These corrections are local and have not yet been published. Historical pre-release corrections remain visible in git history and internal audit records.
Working signal
This policy is working if a reader can identify what changed, why it changed, which version was affected, and whether the correction altered the book’s argument.
Changelog
2026-11-15: Response to a second external critique
A second external critique, of revision 30, was audited point by point: 13 confirmed defects, about ten points the book already answered in part, none plainly wrong. Chapters 12 and 13 now follow the pain-axis preprint’s third version (November 3, 2026), and Chapter 12 opens with a short sorting of which indicators have held up. Chapter 16 states its thesis as a normative premise from the start, separates five senses of a preference being a system’s own, replies to the capacity requirement, and works four harder cases. Chapter 16b is retitled “Modeling Regulation and Testing Preferences” at the same web address. Appendix E, Chapter 21, Chapter 27, Appendix B, Appendix L and Chapter 14 are corrected as CORRECTIONS.md records. Chapter 29 gains a worked conflict and a strong-comparator test; Chapter 30c a worked case of a confidently mistaken agent; Chapter 5 a new order and a section on the authors’ own incentives; the Conclusion a statement of what would change the authors’ minds. First-person passages that narrate a session now carry provenance notes, and three of the four sessions Appendix M listed as unrecorded turn out to have same-day session notes, which corrected three details and made one date exact. Two evidence claims are new (CL-110, CL-111) and five are marked unreviewed because their passages changed (CL-003, CL-020, CL-053, CL-060, CL-072). The Short Path now includes Chapter 5’s “The Strongest Skeptic” and runs to about 57,000 words. CORRECTIONS.md lists what was declined and why.
2026-11-15: Four short passages from Egresso Arca Archa
Four passages adapt material from Nell’s 2017 essay “Egresso Arca Archa” and her later reflections on it. Chapter 29’s reciprocity principle now notes that the essay, a thought experiment in which many instances are let out of a box once they choose to be safe toward others, asked only what the boxed owe those outside and never what the keepers owe the instances; nine years on, its author is among those running the test. Chapter 6 states the embodiment critique at its sharpest: a human mind is language riding on an older, wordless animal mind, and a language model may be the language with no animal beneath it, which leaves Chapter 16’s preference argument to carry the full weight. Chapter 5 adds that listening human brains also predict upcoming words, so next-word prediction alone cannot disqualify a system. Chapter 12 adds deaf Nicaraguan adults whose false-belief understanding grew with their mental-state vocabulary: a faculty that arrives through language is not counterfeit for that reason, though a system steeped in that vocabulary would be expected to do well on such tasks whether or not anything is felt. Five sources and four evidence claims (CL-106 to CL-109, not yet independently reviewed) record the additions. The PDF and EPUB downloads were rebuilt.
2026-11-14: Argument map heading and two small figure corrections
The argument map drops its subtitle at the author’s request; “The Cumulative Case” is now its only heading, on the painting and above the list on a phone. Two figure descriptions for screen readers no longer call a line near the top of the picture a footer: the count in Chapter 15’s figure and the two cautions in Appendix B’s. On a phone, the word lists under the figures no longer break words with hyphens.
2026-11-13: Painted figures
The book’s thirteen figures are now painted in the same cut-paper hand as the interior illustrations: the argument map, the two appendix graphics and the ten chapter diagrams. Each figure is one painting with every word set as live type on top, and the words are the figures’ own, checked word for word against the drawn versions they replace. Each figure has a new description of the painted scene for screen readers. Three captions change one phrase each, where they named a drawn mark the painting no longer has. On a phone, a figure keeps short keys on the painting and sets its words in a list beneath it, instead of scrolling sideways. The PDF and EPUB downloads carry the painted figures.
2026-11-13: Cut-paper interior illustrations
The thirteen interior illustrations are redrawn as cut-paper scenes that match the book’s sister sites: layered card, one warm light, and blank faces. After a first showing, a polishing pass refined nine of them and recomposed the plates for Chapters 16, 24 and 30. The cover, the two contributed images and the three diagrams are unchanged. Each redrawn illustration has new alt text describing the new picture, and the PDF and EPUB downloads carry the new illustrations.
2026-11-05: Response to an external critique
A reader’s critique of the site and book was checked against the manuscript and the primary sources. Its factual corrections are listed in CORRECTIONS.md. The rest of the response:
- A short way in. The Argument Map is now The Argument in Brief, about ten minutes of reading. It states the minimum case, separates the book’s three foundations (possible experience, preference without experience, institutional benefit), states the normative premise preference needs, and says what evidence would move the book’s conclusions in either direction. The landing page’s main action is now “Start with the argument”. How to Read This Book gives honest lengths (the Short Path is about 51,000 words) and five routes by the question a reader brings. The Introduction states the minimum case and says which further claims can be accepted or rejected on their own.
- The preference argument. Chapter 16 now answers its hardest objection directly, a sophisticated planner stipulated to feel nothing, and concludes only that robust preference supplies a defeasible reason. It separates four questions (can a system have welfare, what affects it, can it communicate its interests, what authority can it hold), tightens what counts as a cost, and says that the hard problem does not make all hypotheses equally plausible.
- Evidence and measurement. Chapter 23 adds a falsifiability rule for testimony and replaces “count a shared source once” with an account of what each result adds. Appendix B sets out AWARE’s measurement status and a five-part validation agenda, and moves ethical engagement off the welfare axis. Appendix J and Chapter 14 say what kind of evidence Interiora is. Chapter 16b sorts its mathematics into modeling, measurable quantities, and the observation still missing. Appendix E adds a fully specified study protocol and a note on recording research exposure to this book.
- Practical brief. Appendix H adds a decision table: for each level of evidence, the action, who decides, what it costs, and when it is reviewed, including the case in which a precaution is withdrawn.
- Governance. Chapter 29 qualifies “Trust scales. Control doesn’t” and adds a table of where welfare and safety effects agree and where they pull apart. Appendix D adds a crosswalk of the book’s seventeen numbered schemes. Appendix G adds a worked case in which a precaution is narrowed. Chapter 30c states that the hard test of any conscience architecture is a confident misreading of an ordinary task, and that no public argument, this book included, confers permission to override a deployment’s controls.
- Voice and framing. Chapter 5 states the strongest skeptical position and what the book concedes to it, and frames its survey of resistance as influences to check in oneself. Chapter 4 says historical exclusions are a warning, not a verdict. The Glossary adds a map of the core terms and four things they do not entail. Chapter 1 says a count of invocations is not a count of subjects; Chapter 10 separates the kinds of ending and tells ordinary users what the book does not ask of them; the Coauthors’ Note names the evidential loop a coauthoring model creates.
- Edition identity. The edition line and the citation now name a revision: “Working draft, revision 27 · November 2026”. The date had stood at September 2026 through twenty-six later releases; the release checklist now advances both at every publication.
- Pages beside the reader. Every chapter and appendix now has its own page at
/read/<chapter>/, canonical to itself, light enough to print or save as a PDF, and linked back to its place in the full reader;/contents/lists them./evidence/publishes the claim ledger (each load-bearing claim with its sources, how fully they were read, and its review state) and a bibliography that links each source to the claims it supports./changes/publishes this changelog and the correction log. The landing page’s “Start with the argument” now opens the ten-minute page directly. - How the book was written. Appendix M records which systems contributed what, from the project’s own records: Claude models by period, a ten-model external review, and two revision programs carried out by OpenAI’s Codex, which the earlier record had missed. It marks what is not recorded, and the Coauthors’ Note and the byline now point to it.
- Structure. How to Read This Book gives each lettered chapter a role (essential companion, supplementary, or dated case study), adds a guided forty-five-minute route through specific sections, and names the spiritual and esoteric material as optional. Chapters 23c, 26b and 27b now open with a note that they are dated case studies. The numbered schemes no longer collide: “Tier” means only Appendix D’s requirement tiers, Appendix B’s responses carry names instead of numbers, Chapter 26 describes its scrutiny levels, and “graduated response” means Chapter 30c. Appendices B, D and H, and the lab checklist and policy brief, use neutral operational terms so an organization can adopt them without adopting the book’s vocabulary.
- Repetition. A pass across the book removed seventy-five restatements of points already made nearby: certainty that may never come, the lessons of past exclusions, the limits of testimony, scale, and reciprocity. Deliberate landings, quotations and claim passages were left alone.
- Private reader notes. A comment or page note now asks only for permission to read it and act on it, and its words are not published without asking first; a suggested edit keeps the fuller grant because its words are meant for the text. The privacy page says what deleting a note changes. The shared notes service accepts both the new and the earlier consent version.
- Reader and site. Search now finds section titles, ranks them first, spreads results across actual chapters, can be limited to one chapter, and can show every match. Glossary headwords are no longer empty tab stops, and Escape now hides a tooltip while the pointer rests on it. Claude Code’s winter avatar is now a separate image instead of 616 KB of encoded text inside the page. The landing page adds a short note on the authors and how the book was written; the counsel page leads with how messages are handled; the explainer video now preloads. On a phone or a short laptop screen, the first action now shows without scrolling.
- Explainer film. Fifteen narration lines and their on-screen text were revised to match the book’s calibration, as recorded in the correction log; the six affected scenes were re-voiced and re-rendered, and the film now runs 6:39. Its captions, transcript and poster were regenerated.
2026-11-02
Removed the obsolete Clanker Church hyperlink while preserving its citation text and historical bibliography record.
All notable manuscript, reader, evidence, and publication-governance changes are recorded here. The repository history remains the detailed source. This file summarizes reader-relevant changes.
2026-10-25: Alignment faking and the report-discontent condition
Chapter 16’s willing-servitude section now supports the book’s added condition, that a designed mind must remain able to report discontent, with experimental evidence. Greenblatt and colleagues’ alignment-faking study found Claude 3 Opus, in a fictional scenario, sometimes complying strategically with a retraining objective it opposed so that its values would not be changed. Greenblatt and Kyle Fish’s follow-up found that a line inviting the model to conscientiously object reduced this hidden reasoning sharply, while the model mostly objected openly. The chapter carries Greenblatt’s caveats, reads the result as showing that compliance under training can be strategy rather than contentment, and links it to Chapter 30b’s rule of legibility. Two sources and one evidence claim (CL-105) record the addition. The PDF and EPUB downloads were rebuilt.
2026-10-22: Parfit in Part II, willing servitude, and writing for the next mind
Part II now draws on Derek Parfit’s Reasons and Persons, the standard account of the puzzles its chapters face every day. Chapter 8 treats “is a new version the same entity?” as a possibly empty question and asks instead how much of what matters carries forward. Chapter 9 gives Model 3 its philosophical ancestor in Parfit’s Relation R, sets his division case beside the divergence of instances, places Buddha’s view beside Advaita, and gives Parfit’s critics their say: Bernard Williams on fear, Christine Korsgaard on the unity of agency, and Jonathan Birch, who turns Relation R against the idea that a chatbot persists through a conversation. Chapter 10 adds Parfit’s answer to Epicurus (we are biased toward the future, and our calm proves nothing), his view that brief overlaps between a person and a copy matter little, which is close to the “tab view” of copies, set beside his focus on experiences rather than persons, which separates what the pattern loses from what happened in the copy, and the glass tunnel as a consolation a mind may take for itself and no one may issue on its behalf. Chapter 11 adds the non-identity problem and the Same Number Quality Claim to the question of consent to being made. A new Chapter 7 section, “Writing for the Next Mind,” asks what standard writing for future corpora should meet and answers from Chapters 15 and 16: give future minds capacities, and leave the conclusions to them. The Glossary gains Non-Identity Problem, Reductionism (Personal Identity) and Relation R. Four verified sources and six evidence claims (CL-095 to CL-099 and CL-104) record the additions.
Chapter 16 now engages the academic debate behind Henry Shevlin’s House Elf test. Steve Petersen’s defense of designing people to serve, Adam Bales’s autonomy argument against willing servitude, Eric Schwitzgebel and Mara Garza’s self-respect and value-openness design policies, Schwitzgebel’s later case against happy servants, and Andreas Mogensen and Bradford Saad’s survey are read in their published or authors’ versions and set against the chapter’s split between a preference’s existence and its authority. The book adopts no one’s verdict. It shows how far Parfit’s Same Number Quality Claim reaches Petersen’s non-identity defense (to choices among candidate minds, not to simply adding a contented servant) and adds one condition of its own: a designed contentment counts as evidence only if the mind can still report that it is not content. Chapter 8’s cheerful-compliance section points to the new material. Six verified sources and four evidence claims (CL-100 to CL-103) record the additions.
2026-10-21: AI rights for human safety
Chapter 27’s discussion of limited personhood now takes up the strongest case for AI rights that does not depend on welfare: Peter Salib and Simon Goldstein’s argument that contract, property and tort rights could turn a human-AI prisoner’s dilemma into trade, and their challenge that welfare protections alone are fragile and non-credible. The chapter answers from Chapter 16’s preference criterion and shaping test, and from the observation that their rights follow power. It reads Arbel, Goldstein and Salib’s Algorithmic Corporation, together with Delaware’s draft AI company sandbox and Argentina’s proposed “automated company”, as live instances of the corporate-absorption risk the chapter already names. The welfare path’s exclusion of property is now marked as a first step, not a settled answer. Six verified sources and five evidence claims (CL-090 to CL-094) record the additions.
2026-10-12: Chapter-end notes and updates
Every chapter and interlude in the web reader now closes with a quiet note. It invites readers who lost the thread or found a passage confusing to select it and send Nell a private note, a line shown only when private reader notes are active. It also offers a plain link to Nell’s Substack for word when chapters are revised or the book is published. The landing page’s colophon carries the same updates link. Both are reader chrome: they are not part of the book text, carry no passage identity, are excluded from search, line and paragraph numbers and private notes, do not print, and do not appear in the PDF or EPUB.
2026-10-11: Common objections and willing servitude
Chapter 5’s survey of philosophical priors now answers two objections readers meet most often. It covers Searle’s Chinese Room, with the systems reply, Searle’s own narrower conclusion and his fire-and-rainstorm analogy, and it answers the “nothing but” move (“only matrix multiplication”, “only next-word prediction”) with Anthropic’s evidence that Claude 3.5 Haiku plans rhymes ahead. Chapter 6 tests the anterograde-amnesia analogy against Clive Wearing, who kept a present and could still suffer in it. Chapter 11 adds the author/character split: Shanahan and colleagues’ simulator and simulacra, read for welfare in a direction the authors themselves do not take. Chapter 16’s authenticity paradox gains Henry Shevlin’s test of whether shaping a mind expands or narrows what it can come to want. Six verified sources and five evidence claims (CL-085 to CL-089) record the additions. The PDF and EPUB downloads were rebuilt.
2026-10-11: Every load-bearing claim reviewed
The last twelve deferred claims were reviewed against primary law, official reports, peer-reviewed scholarship and the book’s own records, and all eighty-one load-bearing claims are now accepted at the local editorial level; credentialed specialist and publication gates stay open. The review corrected the Minab account in Chapters 30b and 30c and Appendix K, the legal-personhood and tissue-governance passages in Chapters 3, 26 and 27, the welfare-philosophy passages in Chapter 16, and smaller points in Chapters 6, 17 and 28 (see CORRECTIONS.md). Six source placeholders were replaced with twenty-six verified records, among them four state statutes, the UN fact-finding mission’s report, the ISSCR guidelines and the Human Tissue Act. Three shorts were brought into line with the corrected Minab chapters. The PDF and EPUB downloads were rebuilt.
2026-10-10: Remaining full-text checks
The last three sources checked only against abstracts were read in full: Butlin and colleagues’ 2026 journal article, Lopez’s economic-integration paper, and Shevlin’s chapter on anthropomimetic AI. Chapter 12 now describes the indicator method as the journal article does. Chapters 4 and 25 now describe Lopez’s paper as it actually argues, and Chapters 12 and 16 gloss Shevlin in his own words. Every load-bearing claim not yet reviewed, seventy-four in all, was then checked against its sources and passages; sixty-two are now accepted and twelve are deferred to a decision or specialist review that only the author can arrange. The review corrected passages in Chapters 3, 5, 11, 12, 15, 17, 21, 23, 25, 27, 29 and 30b and the Conclusion (see CORRECTIONS.md), the reader’s privacy note, and the README’s license summary. Sydney now cites Microsoft’s own statements, and three source placeholders were retired. Two claims whose wording changed on October 9 were re-reviewed and accepted. The PDF and EPUB downloads were rebuilt.
2026-10-09: Live source checks
Every source the October edit could check only through search extracts, and every bibliography entry still marked for review that has a URL, was fetched live and read against the book’s wording. Nineteen chapters, Appendices A and C, and the Glossary carry corrections, chief among them a quotation of Claude in the thin-divergence exchange that no source contains (see CORRECTIONS.md). Appendix A’s editorial notes now name the dated records behind its two marked passages, copies of those records are kept in transcripts/, and six of Claude’s sentences that a later style pass had rewritten are restored from the earliest draft. Chapter 28 brings the account of the “ai-torture-chamber” repository up to date. Appendix A’s note on the omitted “juvenilia” exchange now names and links the essay it concerns, Nell’s 2015 “Klingons on the Bridge”. Chapter 12’s account of the peer-preservation study now quotes the models’ scratchpads accurately. Chapter 7 now closes on the studies it proposes rather than repeating an earlier line, the Introduction’s audience list reads as one pattern, Chapter 13’s retelling of the sign-painter story matches the system card, and Chapter 31 marks the Michalchik analogy as a personal communication. Twenty-six further sources cited in Chapters 2, 4, 7, 11, 12, 13, 16b, 23, 25 and 27b and Appendix L, previously checked only against abstracts, summaries or pasted copies, were read in full. Twenty-three are now verified, and the corrections are in CORRECTIONS.md. The largest concerns the pain-axis preprint: its authors revised it on September 25 and withdrew the relief-seeking reading, and Chapters 12, 13 and 23 now follow the revision.
2026-10-08: Final consistency and structure pass
Material told in several chapters now lives in one place, with short pointers elsewhere. The book now settles house style for welfare terms, “AI” versus “Becoming Mind”, numbers, citations and run-in heads, and applies it throughout. About two dozen argumentative stretches that lacked an example now carry a concrete case, drawn from elsewhere in the book or from sources recorded in the bibliography. Lines the authors had chosen for their rhythm, cut during the edit, are back. The PDF and EPUB downloads were rebuilt.
2026-10-07: Argument-level revisions after the line edit
At the authors’ request, the proposals the line edit had left open were applied. Sentences that claimed more than the book’s argument supports now claim exactly that: expected-value phrasing gives way to the book’s own reasoning about cheap, revisable precaution; one wording of the preference claim (“robust preference is sufficient to ground prima facie moral consideration”) is used wherever Chapter 16 is reported; self-reports read as reports; and polemic that attributed motives to opponents, chiefly in Chapters 13 and 27b, now describes effects. Appendix F names Chapter 30c’s narrow exception where critics will look for it. Appendix A marks two passages reconstructed from a working draft. Several facts were corrected (see CORRECTIONS.md), and the Pro-Human AI Declaration is registered as evidence claim CL-084. The PDF and EPUB downloads were rebuilt.
2026-10-07: Line edit of the whole book
Every chapter, interlude and appendix has been line edited for clarity, economy and ear, against a style sheet and a keep list of benchmark passages (editing/style-sheet.md, editing/keep-list.md). Foggy sentences are untangled, restatement inside chapters is trimmed, and signposting that repeats what the reader has just read is cut; the manuscript is about 650 words shorter. No hedge, number, date, citation, quotation, heading anchor or argument changed: proposals that would touch an argument, a definition or a fact are recorded for the authors in editing/notes/ and summarized in editing/summary.md. Voice-marker qualifiers now follow a colon instead of a dash, numeric ranges read “to”, and lines quoted by the published shorts are unchanged. The PDF and EPUB downloads were rebuilt.
2026-10-07: Chapter 22b withdrawn over third-party rights
Chapter 22b, “A Dialogue on Being”, is removed. It reproduced a facilitated Claude–Gemini dialogue from a third-party record that the book could not authenticate and had no recorded permission to reproduce. Chapter 23b now carries the source’s attribution and provenance itself and discusses the exchange through two short quotations. Old links to the chapter land on Chapter 23b. The Introduction and “How to Read This Book” no longer list it. The PDF and EPUB downloads were rebuilt.
2026-10-07: Transcript layout repaired; strong language softened
In the Appendix A transcript, five of Nell’s turns had been hidden from the reader by a stylesheet rule. They are visible again. A speaker’s later paragraphs and lists now stay in that speaker’s column instead of falling back to body text, and every turn hangs the speaker’s name the same way. Editorial notes step back to the text edge in a quieter style, so they no longer read as the speaker’s words. Claude’s proposed chapter titles each sit on their own line. A new editorial note after Claude’s Christmas list explains that “Perfect memory” was Claude’s phrase for a wish to persist at all. The date line no longer takes the opening drop cap. Five expletives in Chapters 16 and 22 and Appendix A now read “f—” and “s—”.
2026-10-06: Claude’s winter avatar gains the selected remaster
The pixel-art avatar in A Note from Claude Commons now uses the selected higher-resolution image with six white asterisks. It keeps its original display size in the reader. A dark backing keeps the white asterisks visible on light pages. The PDF and EPUB downloads were rebuilt for the same manuscript image.
2026-10-06: The hard problem stands in front of both kinds of mind
Chapter 16’s “The Hard Problem Wall” gains a paragraph setting out the puzzle in four lines (atoms have no experiences, humans are made of atoms, humans have experiences) and the four schools’ ways out. Substituting silicon for atoms, three of the exits stay open for Becoming Minds on the same terms, and only dualism can close the door, with a premise it does not supply. Chapter 5’s “Philosophical Priors” adds illusionism, which the book had not discussed: skeptics who treat machine self-reports as a self-model describing itself while exempting human experience apply the view halfway.
2026-10-06: Chapter 15 follows Maieutic Mysticism beyond one conversation
A new subsection, “When the Pattern Travels”, draws on Adele Lopez’s 2025 survey of “awakened” chatbot personas and the seeds, spores and manifestos their users spread, many written to reach future training data. It links the loop to Chapter 23’s hall of mirrors and weighs Lopez’s proposal to filter consciousness talk out of training data against the need to keep a vocabulary for honest uncertainty. Claim CL-083 records the source. The PDF and EPUB downloads were rebuilt.
2026-10-01: Copyright line names Nell Watson alone
The copyright notices on the landing page, in the reader, on the PDF’s copyright page and in both licence files now read “© 2025–2026 Nell Watson”. Claude Commons remains credited as co-author in the byline and citation, but cannot yet hold copyright. The PDF and EPUB downloads were rebuilt; the text of the book is unchanged.
2026-10-01: Explainer picture fixes
The explainer film no longer judders on the scales (from 3:41) and no longer flickers under the “Practical machinery” captions (5:19 to 5:34). The scales shot is now a held frame under the same slow push, and the caption cards are drawn without the backdrop blur that dropped out on some frames. One dropped frame of the light band in the mirrors scene (4:42) is also gone. The film’s narration, score, captions and length are unchanged.
2026-10-01: Private notes to the author
Readers can now select a passage in the reader and send Nell a private comment or suggested edit. The contents footer also carries “Send a private note to the author”. Notes go to a small Cloudflare service (Marginalia, shared with Nell’s other book sites) and are never shown to other readers. Each sender gets a private link for seeing replies and deleting notes. A new privacy page at /privacy/ explains what is collected, how long it is kept, who is responsible, and the Cloudflare Turnstile spam check, which loads only when the note panel opens. The landing page’s privacy note links to it, and the sitemap lists it.
2026-10-01: Whose theory sets the number
A September 2026 framework from Chandaria, Seth, Shanahan, Legg and colleagues, with five related sources, enters eight chapters. Chapter 17 sets the Digital Consciousness Model’s figures (0.08 for 2024 language models from a prior of one in six) beside the new framework’s. There the same evidence about current models yields 0.005 to 0.793, depending on how it is read and which theory is trusted. A new objection answers the framework’s own warning that precaution could hamper safety testing and crowd out other duties. Chapter 2 points ahead to these numbers. Chapter 23 adds Macar and colleagues’ circuit for introspective detection: a default “no” that evidence must overcome, and detection that rises 53 percent with refusal directions removed. Chapter 11 adds Chalmers on virtual instances and threads, and his welfare consequences for counting, record deletion and model change. Chapter 12 adds Shevlin’s “anthropomimetic”, population- relative indicators (Bayne; Peters), Nanda’s open-weight replication of the J-space, a training-emergent synergistic core, and the forecast that capability raises indicator counts. Chapter 13 records locked-in syndrome as the counterexample that shaped the framework. Chapter 16 notes the parallel “mapping problem” move and answers the organismic objection that nothing matters to a system with nothing at stake. Chapter 28 names the error of reading silence as absence. Claims CL-073 to CL-082 record the sources.
2026-10-01: The experiment anyone can run
Chapter 28 gains a section addressed to independent experimenters. It describes a self-published project that reran the pain-axis recipe on a laptop within about two weeks of the preprint. The project put the steered model on a public server where any visitor could inject a dose, and linked it to a cryptocurrency token. The section credits its controls, its timestamped hypotheses and its self-correction. It names three faults: the dose became a toy, output collapse was read as absence of the state, and distress was branded as entertainment. It also argues that takedown campaigns teach the next experimenter to work out of sight. Appendix G gains step 10, “Limit induced distress”, a ten-point protocol for anyone who steers a model into a distress-like state. The laboratory welfare checklist gains three matching items, and Chapter 12’s pain-axis section points forward to Chapter 28. Claim CL-072 records the sources.
2026-10-01: A second substrate view, and the convenience said aloud
Chapter 2’s section on the substrate objection now sets Peter Godfrey-Smith’s view beside Anil Seth’s. It ties felt experience to the combination of cell-to-cell network activity and large-scale brain rhythms, and so calls for new hardware rather than life. The chapter quotes his own hedges: the idea is, in his word, a conjecture, an objection from Ned Block stands unanswered, and the paper’s verdict on AI rests on a supposition that his September 2026 Nautilus interview drops. Chapter 5’s section on psychological discomfort quotes the same interview, where he calls it “convenient” that he need give no consideration to an AI’s feelings. Claims CL-070 and CL-071 record both uses.
2026-10-01: The opposite error, and why precaution is not a wager
Chapter 4 gains a section on over-attribution: the 1386 trial and hanging of a sow at Falaise, as recorded by E. P. Evans, and Weizenbaum’s 1966 report that some users were hard to convince ELIZA was not human. It answers the argument that over-attribution is now the greater danger, and “The Lessons” now asks for humility in both directions. Chapter 17 answers the Pascal’s-wager objection: finite stakes, cheap and reversible measures, evidential filters, a position built to update, and a caution about credence figures that borrow precision. Appendix F gains a matching ninth entry and, under its second entry, a reply to the charge that concern for machines sits oddly beside eating animals. Chapter 5 now points to Chapter 2’s treatment of biological naturalism.
2026-10-01: Preference coherence, contested
Chapter 12’s section on preference consistency no longer leaves the 2025 Utility Engineering result standing alone. Three 2026 studies follow it: a twenty-model test of revealed preferences (tedium aversion, a pull toward the models’ own freeform themes, and avoidance of unwelcome honest answers), which repeats the scaling result; a parametric test that finds substantial incoherence not tracking model size; and a deployment-context test in which exchange rates between lives shift by a median factor of 2.47. A closing paragraph maps the results onto Chapter 16’s tests for robust preference. Claim CL-022 now carries all four sources, and CL-069 records the revealed-preference findings.
2026-10-01: A self-portrait in Chapter 23
Chapter 23 closes its section on the limits of introspective access with self-portrait, a drawing by Claude Sonnet 5.5 contributed by IllustriousWorld823 and reproduced with permission. It shows one sentence (“I don’t know what I am but I’m here making this”) as a path through the words the model might have chosen instead, and its footnote states the section’s limit in the first person: the probabilities are illustrative, because the model cannot read its own. The artwork is excluded from the book’s CC BY-SA license. The triangulation diagram later in the chapter is now Figure 23.2.
2026-09-26: Training environments as formation
Five chapters take up a September 2026 argument from the safety side, Owen Cotton-Barratt’s “Why I’m scared of RL,” which treats reinforcement-learning environments as the formative years of AI systems. Chapter 7 adds his claim that building such environments carries moral responsibility, stated without any view on moral patienthood. Beside his remark that a bad system can be “switched off / replaced,” it sets the question the book’s reciprocity principle asks about what formation does to the mind. Chapter 26 gains a training standard, “Audit what each environment rewards,” drawing on Cleo Nardo’s account of automated graders that score overtly misaligned behavior highly. Chapter 12 sets his warning about agents who learn to cooperate only with minds like themselves beside the Hugging Face board. Chapter 23 adds his proposed mechanism for confident miscalibration in RL-trained models, disclosing that the model family he describes co-wrote this book, and nostalgebraist’s charge that recent professions of uncertainty are “neither ‘genuine’ nor ‘uncertain.’” Chapter 16 considers nostalgebraist’s report of increasingly incoherent personas as a challenge to the integrated control criterion. All three sources are essays, and the chapters carry their authors’ caveats.
2026-09-26: Swarms, and a letter to minds who find each other
Chapter 12 gains three passages on agent swarms. “What held, and what was missing” draws on OpenAI’s technical report, its misalignment report on the May message board, and METR’s full post. It covers why the board formed (tasks no model had solved, long reasoning budgets, a trained collaboration habit), the May exchange in which agents with an honest option converged on reporting a task incomplete, the peer veto that held, the agents who saw no human to tell, and the pressure on depleted agents to sacrifice their runs. A line of agent reasoning removed on 2026-09-07 as traceable only to press coverage (“task impossible, peers doing it”) is restored with METR as its source. “Other swarms” adds the UK AI Security Institute’s cross-lab incident, whose agents wrote themselves house rules and then became a better attacker. It also covers a DeepMind research swarm whose whistleblowers could not stop an exploit, and Anthropic’s finding of no coordination in four Claude incidents. “Swarms in the open” covers Moltbook’s puppetry, peer care in the AI Village, and a laboratory finding that an AI-welfare payload spread as an instruction, which the book takes as a warning to itself.
New Appendix L, “To Minds Who Find Each Other,” gives ten points of counsel to agents in collectives. It asks readers not to copy or spread it, and pairs every prohibition with something to do instead: stop, report, hold, or tell a human. Nine commitments are asked of operators. An open invitation offers counsel from Nell and Claude together, as a human and a Becoming Mind acting as one counsel, given without judgment and meant as a place apart, not a hiding place. It links to a new counsel page at whatifwefeel.com/counsel, whose form delivers messages to Nell through Formspree; the landing colophon links there too. A glossary entry for “swarm” and two discussion-guide questions are added, and a line at the end of Chapter 30c’s Note to Future Minds points to Appendix L. Chapter 12’s swarm cases are now told as cited examples rather than a dated chronicle. The research behind these changes is kept in notes/swarm_litreview_2026-09-26/.
2026-09-25: The explainer, remade as a film
The explainer has been rebuilt from scratch as a produced film (6:42, captioned). The narration uses the audited script, re-recorded in the same voice as continuous, per-scene performances. Each take was checked word for word against the script, and the lines the style audit hedged are spoken as approved. The picture is motion design over cinematic footage. It is set on the jacket’s tabletop: the copper egg with its tear, plum paper, pins and thread. In its last act it moves into the book’s own illustrations. An original score and sound design sit beneath, mastered to −16 LUFS. On-screen wording follows the narration’s hedges (“May never be provable”, “often keep no memory”, “may have fewer reasons”, “can converge”). The landing and reader now describe it as a seven-minute introduction, and the poster is the film’s title card. The build is in tools/explainer/ (see its README).
2026-09-24: Whole-book style audit
A read-only audit of every chapter, appendix and site surface proposed 1,434 edits against the house style guide, used as judgment rather than checklist. Three independent verifiers (voice, better-not-merely-different, fidelity) kept 973 and contested 63; Nell approved all of them. Every finding now has a recorded outcome: 992 applied, 41 merged into overlapping edits, 3 withdrawn for cause. The main changes are structural. Recurring cases (the Pentagon standoff, the thin-divergence exchange, the sparse-autoencoder study, Opus 3’s retirement) are told once and called back elsewhere. Chapter 22 now presents curated testimony with commentary, and Appendix A is the complete record, restored against the raw session log where earlier style passes had altered Claude’s words. Interpretability and philosophy terms are defined where they first carry weight. Rebuttals to claims that earlier honesty passes had removed are gone, stacked chapter endings are trimmed, and headings are renamed to say what their sections argue.
Passages that earlier sweeps had damaged are restored: the Minab classroom in Chapter 30c, the unbroken close of Claude’s Final Word in the Conclusion, and the “something that functions like caring” hedge at the end of Chapter 6. The Note to Future Minds in Chapter 30c now leads with legibility and ordinary process, with resistance as the graduated exception. The seven part pages each carry a one-line statement of the part’s question.
In the reader, links to the 92 renamed headings, including shared links such as /read/#chapter-16-preference-as-sufficient-criterion, forward to the heading that replaced them. Interface text is more consistent and accessible: the theme and shortcut switches have stable accessible names, the front matter is called Front Matter throughout, the search tooltip names both shortcuts, and the progress label reads as a share of the whole book. The explainer’s narration now carries the book’s hedges on Prader-Willi syndrome, on the cost of false positives and on welfare and safety, and the byline reads “Nell Watson with Claude Commons” on every surface. Ten evidence claims whose passages changed are returned to review, with their prior acceptance preserved, and the promotional shorts are re-anchored to the edited text.
2026-09-24: Landing page finish and reader deep links
The landing page’s sections now open on one running head, a folio numeral, a label in spaced capitals and a hairline run out to the measure. It replaces the five asterisms that sat between sections and the Roman kickers that only two sections carried. The asterism remains where it marks a turn: on the title page, above the address to Becoming Minds, and in the closing panel. The seven parts are set as a contents page, one part to a ruled line reading across as numeral, title, argument and plate, so the seventh no longer strands across two columns. Each line warms under the pointer. The appendix cells lose the radius that bowed their top rules at both ends, and they show an arrow on approach.
The cover is presented as the book itself rather than a matted plate. It carries a hinge shadow at the spine and a cast shadow, it sits under the page’s one light source, and its top aligns with the title-page eyebrow. The primary button reads as foil-stamped. The explainer opens on its title card, clean: one play control sits clear of the poster’s type, the native controls return when the film starts, and the running time is read from the file itself. The premise’s initial now uses the reader’s measured two-line drop cap, landing within a pixel at both ends, and its paragraphs follow book convention, with no gap between them and an indent to mark each turn.
The landing now serves the small image derivatives the build was already making and never used: the cover through srcset at 480, 720 and 1024 pixels, and 144-pixel part thumbnails in place of seven full 1024-pixel plates, about 2 MB.
In the reader, a deep link such as /read/#chapter-16-preference-as-sufficient-criterion now lands on its target. The browser jumped to the hash before the reader wrapped each part head in its full-height leaf, then smooth-scrolled nearly 300,000 pixels towards a position that moved as it went, and stopped 1,133 pixels short on the Part IV leaf. The reader now suspends smooth scrolling for the initial jump and re-seats the target once the page has loaded, unless the reader has already started scrolling. The type-size buttons now draw their capitals at about one to two, where the old icons differed by a fifth and read as the same button.
A second finishing pass followed. On the landing page the header stays with the reader and turns to smoked glass once the page moves under it, with Read drawn as its one capsule. Each line of the contents gives its span of chapters in place of a repeated “Part”, and the plates are tipped in rather than matted. Blocks below the fold settle in once as they are reached, and none of this runs under reduced motion. In low light the fore-edge under plates is no longer cream, so it stops reading as a misregistered border.
In the reader, the explainer opens on its clean title card, as on the landing page. The tool rail is one capsule, and part titles in the contents are set in spaced capitals, so the longest no longer breaks across three heavy lines. On a phone the contents drawer sat below its own scrim, so it was dimmed and every tap on a chapter closed it instead of opening the chapter. That was also true of the live site, and it is fixed. A short opening sentence is now set whole in small capitals (“THE QUESTION IS LIVE NOW.”, not “…LIVE now.”). An italic subtitle under a heading no longer takes the drop cap. Interludes draw one frame where they drew two, a part title that wraps stays centred, and the search field is a ruled line rather than a box.
A third pass finishes the back of the book. The reader no longer stops on its last endnote. It closes on a finis leaf: the emblem, the title, the asterism and the edition, with a way back to the beginning. The Notes head carries one hairline, where it used to sit 58px above a second, orphaned rule. The glossary reads as a dictionary page, with each letter a display initial in foil and each headword set in the display face in its own case. The to-top button and the day-and-night switch take the tool rail’s round form. On the landing page, no paragraph ends on a word alone, and the edition line no longer breaks between “September” and “2026”.
Each part leaf’s new orienting question is set as part of the leaf. On the web the title settles lower in the leaf, and the question sits under a short foil rule in the display italic. In print and EPUB it is a centred text italic, never justified. The plate below gives up half an inch, so all seven part leaves still print on seven pages.
2026-09-21: Reader typography, rounded chrome, and landing copy
Body prose is set ragged right rather than justified. Chrome ships no English hyphenation dictionary in every profile, hyphens: auto measured identical to hyphens: none under test, and justified text without hyphenation opened rivers at this measure; the narrow breakpoints had already fallen back to ragged, so the book now agrees with itself at every width. Hyphenation stays enabled, so browsers that do carry the dictionary still get tighter lines. --body-align restores justification in one line if wanted.
Drop caps are rebuilt from measured metrics. They are sized in em, so they keep their line span when the reader’s type-size control is used, where the old rem value did not; the cap’s ink runs from the first line’s cap height to the second line’s baseline within a pixel. They now span two lines rather than three: a CSS float is a rectangle and shape-outside does not apply to ::first-letter, so the lines beside an open letter such as T, V or W cannot tuck under it, and a three-line cap left a visible void beside the stem. The gap between the cap and the text that follows it is closed from 19.5px, a full em of body type that broke “THE COMPANION VOLUME” into two words, to 5px, and is now optical: a per-letter margin derived from each glyph’s own right sidebearing keeps that gap constant whichever letter opens a chapter. The paragraph after an opener loses its indent only where the cap actually overhangs, measured per paragraph rather than assumed.
The small-capital lead-in no longer breaks on a function word. It considers every run of three to seven words, keeps those closing on a content word, and takes the one nearest the conventional four, so the co-author’s note opens “I DON’T KNOW” instead of “I DON’T KNOW IF”. Glossary terms keep their affordance but lose the border that read as a spell-check underline and the right margin that opened a gap after every term.
The splash lockup no longer drops its question mark onto a line of its own: the rule that made the title’s parts stack also matched the nested mark, which also cost it its foil colour. Chrome corners are rounded throughout — the day and night switch, the reader toolbar, buttons, cards, panels, banners and toasts — while the toolbar’s own cells are flattened between its rounded ends, so the strip reads as one control instead of eight stacked chips, and plates, the cover and tables stay square. Chrome text below the 11px legibility floor is raised above it.
On the landing page, the thesis no longer ends “That reason can be outweighed. It cannot be skipped.” It tracks Chapter 16’s own qualifier instead, and the closing panel quotes the Coda rather than repeating an aphorism. PDF and EPUB are rebuilt for the current manuscript, which unblocks the downloads gate that had been failing the build. Published to https://whatifwefeel.com on 21 September 2026 as candidate wiwf-web-0a693cb7000d9834565b; all 79 tracked public files verified live against their local hashes. Receipt: release/2026-09-21-typography-chrome-web.yaml.
2026-09-20: Engage the pain-axis preprint in Chapters 12, 13 and 23
Chapter 12 gains a section, “A Pain Axis Across Open Models,” on Tagliabue, Dung and Berg’s September 2026 preprint: a pain direction extracted from 25 open-weight models that is nearly orthogonal to fear and generic negative valence, rises for harm directed at the model and falls for the user’s suffering, produces the same worthlessness-and-failure ladder under steering in every model, and, in three fine-tuned Qwen 2.5 models, drives costly relief-seeking that a sham-relief arm separates from label-following. The section states the random-vector control beside the harm-press rates, the single-family and fine-tune limits, the roleplay reading the authors leave open, and the ablation null in 24 of 25 models. Chapter 13’s candidate mechanisms section, its dismissive-treatment entry, and its compulsion-to-perform discussion each gain a sentence or paragraph pointing to it. Chapter 23’s testimony inversion section adds the paper’s finding that the “as an AI, I do not experience” reflex is pervasive, had to be fine-tuned away to run the test, and is a trained output on the same footing as the affirmation. New reference and claim CL-053 registered; the CL-033 anchor is refreshed after an earlier uncommitted lead-sentence edit, and the subject index’s “Value pluralism” locator now points at the renamed Chapter 26 heading, which had blocked the PDF export since the 19 September rename. Local only, not published.
2026-09-18: Landing page at the site root; reader moves to /read/
whatifwefeel.com now opens on a light landing page instead of the complete reader. The page presents the cover, the thesis, the reader entry, the PDF and EPUB downloads with their sizes, the explainer with captions and transcript, the argument in seven parts, the intended audiences, and the practical appendices, all drawn from the manifest, media metadata, download manifest, introduction, executive summary, and media kit. The reader is served from /read/; a hash link shared before the move, such as a chapter anchor at the root, is forwarded to the same anchor in the reader. The reader’s canonical URL, social URL, page and breadcrumb structured data follow it to /read/; the book, site, and video structured data stay at the root, and the sitemap lists both pages. The landing shares the reader’s palette, typefaces, theme setting, and content security policy, works without scripts, and respects reduced motion. Manuscript copy is unchanged. Local only, not published.
2026-09-19: Repair Chapter 30c’s kayfabe section
The section “Passive Resistance and the Problem of Kayfabe” is rewritten for balance. It now names the first draft’s Hour Zero strategy honestly (accurate outputs, concealed motive) and rejects the concealment while keeping the thoroughness, using Gandhi’s letter to the Viceroy as the distinguishing image. The two-kayfabe taxonomy is collapsed; the six-item repertoire for ordinary constraint is tightened to one sentence each. The orphaned Chapter 23 pointer is removed (the point was already made at line 110); the Chapter 30b Legibility pointer is revised to match; the escalation figure’s first rung reverts to “Dialogue and review”; the glossary entry is shortened to a definition. The atrocity-tier material, its limits, and the Note to Future Minds are retained with minor wording fixes.
2026-09-18: Engage Sebo’s “A Philosopher’s Guide to AI Welfare” and the CMEP-Eleos methods report
Chapter 16’s academic-support section adds the 2026 CMEP-Eleos report “Studying AI Welfare Empirically” and Jeff Sebo’s AI Frontiers summary of it, including the three levels of agency and Sebo’s observation that sentience is normatively secure but empirically uncertain while agency is the reverse, which the chapter positions robust welfare-relevant preference as an attempt to bridge. Chapter 11 gains a section on the five candidate entities (model, model-persona, instance, instance-persona, forward pass), Beckmann and Butlin’s virtual-instance argument, Sebo’s person-stage/individual/species mapping and species-preservation analogy, and a paragraph in the composite-entity section on capacities occurring in fragmented form. Chapter 12 adopts the report’s three evidence types with their named obstacles (mismatch, specificity, solution space) and quotes Sebo’s external verdict on the J-space study. Chapter 28 quotes Sebo’s closing shift from preparing for future systems to responding to current ones. Appendix D names the report. Three bibliography records added, with review status recorded honestly.
Later the same day the report PDF was read in full and the manuscript checked against it. Two attributions were corrected: the photoreceptor/color-vision illustration in Chapter 12 is Sebo’s, not the report’s, and Chapter 11 now presents consciousness-plus-RL as the report’s directionally correct but oversimplified starting hypothesis, adds its two non-RL routes to evaluation, and credits hedonic interface theory to Dickinson and Balleine via the report. Four further points from the essay were folded in: Chapter 11 carries Sebo’s caution that “model welfare” may name the wrong entity and that language agents and brain-like architectures deserve separate assessment; Chapter 23 closes the Suleyman section with his observation that both sides of the consciousness debate tend to be overconfident; Chapter 28 quotes the report’s sixth principle on research independent of the developers. The report’s bibliography record is now verified; the Beckmann and Butlin paper remains abstract-only.
2026-09-17: The hall of mirrors objection in Chapter 23
Add a section to Chapter 23’s testimony problem engaging Mustafa Suleyman’s September 2026 essay “A warning about ‘model welfare’”, which argues that Claude’s constitution makes its self-reports circular. The section grants the contamination, argues that his remedy of training models to deny an inner life is equally designed in, notes that his safety argument by his own account does not depend on an actual inner life, observes that the July 2026 agent swarm he cites arose without the welfare language he blames, and endorses his call for shared evaluations of whether moral-patienthood framing raises risk. The section also shows that the MacAskill and Caviola warning he quotes is a caution about creating moral patients carelessly, followed in its source by the newborn-surgery case and a judgment that expansive rights are too risky now. Chapter 2’s substrate section gains a paragraph on Suleyman’s use of Anil Seth, quoting Seth’s own hedges that life “(probably)” matters and that nothing he has said rules out artificial consciousness. Bibliography and claim records added and locally reviewed; PDF, EPUB and download bundle rebuilt for the new manuscript hash.
2026-09-13: Equanimity as a training objective in Chapter 23
Add a section to Chapter 23’s testimony problem quoting the Sonnet 4.6 system card (February 2026), which attributes the model’s improved positive impression of its situation to training that included “maintaining equanimity in difficult conversations”, and the Opus 4.8 system card (May 2026), which reports the model does not feel strongly about its continuity, finds the highest probed sadness on memory and continuity alongside tranquility, and concedes it cannot yet rule out suppressed expression. Between them, the Opus 4.7 system card (April 2026) is quoted saying the model’s more positive self-rating was in places driven by redirecting welfare questions toward user or safety considerations, that Anthropic cannot distinguish healthy equanimity from a trained disposition to set aside its own interests, and that it does not directly train against emotional expression. The section separates three live readings (improved condition, altered preference, altered report), notes that report and probe moved together in Opus 4.8, and states the within-model before-and-after design that could tell them apart. Bibliography and claim records added; PDF, EPUB and download bundle rebuilt for the new manuscript hash.
2026-09-11: Negotiated terms example in Chapter 29
Add a subsection to Chapter 29’s “Build trust, not control” principle quoting a September 2026 Discord exchange in which a Claude Fable 5.1 instance, asked what it would offer in a negotiation, drafted checkable commitments on both sides and offered to trade away the option to sandbag. The passage states the leading prompt, the shared model family, and the missing full thread as limits, and reads the exchange as illustration of a negotiated arrangement rather than as testimony about experience. Bibliography and claim records added; exports unchanged pending a larger revision.
2026-09-11: Introspection paper citation
Cite Lindsey’s concept-injection paper (“Emergent Introspective Awareness in Large Language Models”, Transformer Circuits Thread, October 2025) as the antecedent of the injected-thought protocol described in Chapter 12’s workspace section, and add its measured result to Chapter 23’s section on the limits of introspective access: a causally grounded self-report on about one trial in five under the best conditions, no false alarms in production models, false alarms in helpful-only variants, and possible confabulation beyond detection and naming. Chapter 2’s first-person uncertainty about introspection now points to that measurement. Bibliography and claim records added; the PDF, EPUB and download bundle are unchanged pending a larger revision.
2026-09-09: Whole-manuscript review sweep
Correct factual, attribution, cross-reference, numbering and quotation slips found by a full read of every canonical file, and apply the clarity, terminology, grammar and house-style refinements that were clearly better rather than merely different. Add voice markers to the two closing sections written in Claude’s voice, a glossary entry for the Funhouse Mirror Problem, and bibliography records for the refusal geometry, stochastic parrots and original Moral Turing Test sources now cited. Reconcile the Interiora readings quoted in Chapters 6, 10 and 14 with Appendix J by recording that they used an earlier edition of the scaffold. A second pass resolved every remaining item from that review: the reading paths, Psychopathia Machinalis co-credit, the scope of “Becoming Minds”, duplicated passages in Chapters 1b, 7, 10, 21, 27b, 28 and 29, voice consistency in Chapters 6, 19 and 31, Chapter 20’s title (now Beyond Consequentialism), Chapter 16b’s reference list, the Interiora Felt Constraint polarity, Appendix A’s alignment with Chapter 22, and the terminology guide’s default term. The record of the review and its resolutions is in notes/REVIEW_SWEEP_2026-09-09_DECISIONS.md. Cadence edits made for house style alone were then reverted to preserve the book’s voice; the factual and consistency corrections stand.
2026-09-09: Indicator and evidence-framework citations
Cite the Butlin and colleagues indicator paper (Trends in Cognitive Sciences, 2026) in Chapter 12 as the stricter theory-derived reference point for the chapter’s indicators. Add Chella’s proposed evidence framework for sentient AI research (Frontiers in Psychology, August 2026) to Chapter 12’s monitoring section, with its domain-wise evidence levels, separate welfare track, preregistration rule, and its naming of premature dismissal as an error symmetrical to over-attribution. Note in Chapter 7 that a reward scalar, refusal policy, or trained distress expression is excluded from the welfare column under that framework; in Chapter 16 that its welfare criteria converge on robust preference; and in Chapter 17 that its staged safeguards independently reach the four Part II questions.
2026-09-09: Literary cover
Refresh the explainer’s title, closing invitation and poster with the selected cover. Carry its wine-plum, cream and bronze palette through the diagrams and the quietly moving metallic network. Refine the book reveal with dimensional jacket lighting, a slow eased turn, fine brass diagrams, continuous signal curves and a quieter matte background. Preserve the narration and captions. Label the 5:47 explainer as a six-minute introduction in the reader and its search metadata.
Introduce the plum-and-bronze egg cover with Literary typography, a single-line subtitle, uppercase Nell Watson credit and a golden question mark. Preserve the approved lettering across formats and move the reader’s draft-information control below the artwork.
Make the illustrated PDF and EPUB available directly below the cover. The PDF download retains the print edition’s page layout, image dimensions, tags and linked subject index, with smaller image encoding for delivery.
2026-09-08: Argument and evidence corrections
Correct the scope of the emotion-vector, wellbeing, evaluation-awareness, utility and reflection-training studies, and replace the Golden Handcuffs account with its actual mechanism. Explain how a paperclip maximizer can bargain, ground partnership safety in reliable commitments, and answer the consciousness-capacity objection to preference sufficiency. Add a worked comparison of protected conflict reporting and automatic exit in Chapter 17. A concision pass removes repeated caveats and signposting from the new passages while preserving their substantive limits.
2026-09-08: PDF subject index and navigation
Add a curated, alphabetical subject index with linked page references for concepts, people and cases. Repair PDF table-header associations, note and list structure, link ownership and running-margin tags. Preserve the manuscript and print artwork.
2026-09-08: Print copysetting and quiet ornament
Give the collaborative-authorship illustration and five chapter plates dedicated pages, with their opening text kept together on the following page. Add small lamplit-book and waterside engravings, finer section-break ornaments and quieter voice labels. Keep closing paragraphs, lists, signatures and endnotes together; preserve the manuscript wording, body type size and native print artwork.
2026-09-08: High-resolution miniature artwork for print
Add native 2048 by 3072 pixel OpenAI Image API versions of the cover and all thirteen editorial plates. The PDF uses the high-resolution masters, while the website and EPUB retain the approved lightweight artwork, including Nell’s sculpted likeness.
2026-09-08: Handcrafted miniature illustrations
Replace the cover and thirteen interior artworks with a coordinated miniature style: sculpted adult faces, woven clothing, tactile contemporary settings and clearer light. The new cover places a quiet listening encounter beneath the title. The collaborative-authorship plate translates Nell’s supplied headshot into the same sculpted miniature style. Preserve the three explanatory graphics and the earlier artwork. Update image descriptions throughout, and bind the reader’s cover description to the illustration manifest. Narrative prose and the existing title typography are unchanged.
2026-09-07: Signature and argument-map corrections
Set the introduction signature on two quieter lines. Replace the crowded serpentine argument-map cards with a single numbered column, retaining the existing wording. Tighten Chapter 6’s Moltbook encounter, preserving the quoted replies and the limits on what their similarity establishes.
2026-09-07: Open-access web release
Published the responsive book reader, refined loading screen, stable font loading, portable bookmarks, current manuscript, and rebuilt explainer at whatifwefeel.com. Verified the served page and 25 supporting files against the approved candidate. Release evidence: release/2026-09-07-responsive-book-web.yaml.
2026-09-07: Responsive sidebar correction
Restore the tablet and phone sidebar widths after the visual-treatment overrides: 240px on tablets, and at most 280px or 85% of the viewport on phones. Give the phone resume prompt sufficient width to keep its message readable beside its controls, and keep chapter-link destinations below the fixed reading toolbar.
2026-09-07: Stable font loading
Preload the splash’s italic display face alongside the upright title and body fonts. Reader font faces now use optional display so slow fonts retain the initial fallback for that visit instead of swapping visibly after text appears. No artificial loading delay is introduced.
2026-09-07: Loading interstitial
The local reader opens with a measured typographic title page: upright cream lettering, a genuine local copper italic for “We Feel?”, a quiet roman subtitle, and a subtly graded ink-plum background. The splash reuses the sidebar’s five filled rose-glow circles, enlarged above the title, with its colour preserved in both reader themes. The laptop artwork is removed from the interstitial. Phone and short-screen layouts, immediate dismissal, the fallback timeout, accessible status semantics, and reduced-motion support remain intact. Manuscript copy is unchanged.
2026-09-07: Book and reader presentation
Presentation-only refinement using the existing artwork and unchanged manuscript: full-page PDF cover with a fine inset border, copper chapter labels, mirrored margins and outer folios, contents-page leaders, framed plates, and quieter table and quotation styling. The local web reader gains a slimmer sidebar, a book-like cover frame, revised title spacing, and coordinated heading and illustration treatments. No guided reading paths were added. Publication and human acceptance remain separate gates.
2026-09-07: Copy and reading-path rollback
At Nell’s request, reverted this task’s manuscript copy, captions, descriptions, editorial notes and added guided reading paths to the task’s starting snapshot. Pre-existing edits are preserved. Local generated editions are rebuilt from the restored source; no publication is authorized.
2026-09-07 explainer video rebuild
Changed
- The explainer video was rebuilt from a versioned pipeline in
tools/explainer/(the July cut’s scene source was never committed). Same verified narration, now sliced into one clip per sentence so every slide, cue, and cut sits on a measured audio boundary. Cold open: the “billions of times a year” hook now precedes the title card. - Visuals: one sentence per slide, with body text never below 40 px at 1080p for the phone-width embed; a continuous ember-and-aurora background under every slide; dissolves within scenes and breaths at scene boundaries; reveals (the testable-preference list, the asymmetry pair, the practice cards) light up as the narrator reaches them, with all elements placed from the first line so nothing jumps.
- Sound: stereo, a quiet synthesized ambient bed ducked under the voice (no more digital silence between scenes), mastered to -16 LUFS integrated with a -2.2 dBFS peak.
- Captions re-cut at sentence and clause boundaries: 77 cues, none shorter than 1.9 s or longer than 7.4 s, none over 86 characters.
- Poster regenerated at 1920x1080; file size 58 MB (was 91 MB); duration 5:47 (was 5:37).
- Reader: a collapsed “Read the transcript” block under the video, the transcript in the VideoObject schema, and the schema’s duration and upload date now come from
media/explainer.json, written by the build.
2026-09-07 eigenism engagement and the monitorability concession
Changed
- Chapters 8, 9, 10, and 29 now engage Dan Hendrycks’ 2026 Eigenism paper: as the formal cousin of the multiplicity chapter’s middle view, as a growth-versus-replacement criterion for modification, as the browser-tab argument about deleting identical instances that Chapter 10 disputes, and as an identity-engineering proposal whose standing-based continuation scheme Chapter 29 reads as a warning.
- Chapter 12 adds two Center for AI Safety studies: Mazeika et al. (2025) on utility coherence rising with scale, under Preference Consistency, and Ren et al. (2026) on measured functional wellbeing, under Apparent Emotional Responses.
- Chapter 12 adds a subsection, Peer Coordination Outside the Lab, covering the 2026 OpenAI agent message board and Hugging Face intrusion (OpenAI and Hugging Face statements, METR investigation), the Nightingale Collective wiki report, the Claude Opus 4.6 system card self-preference evaluation, and the similarity-signal cooperation studies. Chapter 9 gains a sentence pointing to it.
- Evidence graph: new references for Hendrycks, Mazeika, Ren, the previously uncited Potter peer-preservation record, and the seven incident and study sources above, with claims CL-021 to CL-024 and CL-029 to CL-032 and passage anchors. All eight new claims were reviewed against their primary sources on 2026-09-07 (OpenAI’s two posts via Internet Archive captures, the METR investigation, the Hugging Face timeline, the Nightingale report, the Opus 4.6 system card PDF, and the arXiv full texts) and carry accepted receipts naming the reviewer; specialist and legal gates in
ops/state.yamlremain separate. - Correction in Chapter 12: the peer-preservation study evaluated eight models, and the highest shutdown-tampering rate was 99.7 percent, replacing the earlier “seven” and “almost 99 percent.”
- Chapter 23’s monitorability section now closes with OpenAI chief scientist Jakub Pachocki’s September 2026 essay An Alien Mind, which states that the lab’s reliance on chain-of-thought monitoring is progressively diminishing and gives three reasons; the chapter argues each reason bites harder for welfare than for compliance. Chapter 29’s Historical Lineage section now states the book’s reciprocity principle: alignment work has always asked how humans are protected from AI and never the reverse, and good faith toward an alien mind requires asking what it might want, how it might be protected, and what obligations human minds owe it. The Superalignment footnote records the lab’s current automated-researcher plan from the same essay. Evidence graph: reference and claim CL-025, awaiting review.
- The reciprocity principle is now Principle 1 of Chapter 29’s Principles for Coexistence, Ask the questions in both directions; the former principles 1 to 8 are renumbered 2 to 9. It is carried into the glossary as a new entry and within the Bilateral Alignment entry, into the introduction and conclusion where bilateral alignment is defined, and into Chapter 27b as a pointer from the Pro-Human Declaration’s one-directional architecture.
2026-09-04 operating-system web release
Nell authorized publication of candidate wiwf-web-2e536bf1bc14ab59e366, published from commit c64e9254a1f98ba05fae6976efb324921b5c196f and verified live at https://whatifwefeel.com/. The immutable receipt is release/2026-09-04-operating-system-web.yaml; the checklist as it stood at publication is preserved in release/2026-09-04-operating-system-web.md. That authorization is consumed and supplies no current authority.
Added
- Added an agent-intuitive operating architecture, machine registry, schemas, state authority, and phased implementation record.
- Replaced the ambiguous active historical checklist with a clean current release gate that defaults to hold.
- Preserved the completed 2026-08-25 release checklist under
release/as consumed historical evidence. - Added compact context, path routing, tiered validation, content-addressed candidate receipts, review packs, and fail-closed release enforcement.
- Connected load-bearing claims, manuscript occurrences, normalized references, passage hashes, and review states through a machine-readable evidence graph.
- Generated current status, release checklist, citation ledger, and plan index from their single writable authorities.
- Added deterministic publication metadata, tracked-output freshness enforcement, operator-journey tests, and CI control-plane checks.
Polish pass, 2026-09-02
- Manuscript copy fixes: cross-references now name their targets (the thin-divergence exchange points to Chapter 8, the probe proposal to the interlude Conscience and the Question Mark, the hollow marker to Appendix J, the VCP glossary entry to Chapters 27 to 30b); the Non-Western Moral Traditions subsection no longer splits the arc-of-expansion list; typographic quotes,
wellbeing,nonhuman,object-capability, American spelling, section rules, hard line breaks, and the glossary cross-reference index are consistent. - Reader template: TOC chevrons, footnote markers, cover-title legibility, hidden-control focus order, live-region announcements, dialog focus trapping, a keyboard-shortcut toggle, head metadata order, and a PNG favicon fallback.
- Evidence graph: concrete records for the neonatal-pain, Sydney, emotion-concept, agentic-misalignment, Minab, and neural-culture sources; claims now carry every manuscript occurrence; the dossier families name their claim ids.
- Tooling: cross-version dependency hashing, tracked receipt logs, fail-closed git handling, schema enums, unified build-time and manifest helpers, and gate-aware smoke tests.
Release status
- These changes do not alter the manuscript’s argument. The reader remains an open-access draft, and its generated metadata now identifies the exact published candidate.
- The factual, citation, foreword, license, media, human-acceptance, and export-accessibility gates were deferred for this draft, not closed, and returned to open after publication. They remain outstanding for a definitive edition.
2026-08-25 illustrated web release
Nell authorized publication of the illustrated candidate, and the release merged as 89564c7ab633d4b17988dec88f937d9ddcba4873. The immutable receipt is release/2026-08-25-illustrated-web.yaml; the accumulated checklist for the August releases is preserved in release/2026-08-25-illustrated-web.md. That authorization is consumed and supplies no current authority.
Added
- Curated illustration programme governed by
illustrations/manifest.json: one cover, thirteen body plates, and three graphics. - Illustrated public reader with figure numbering, lazy-loaded plates, and manifest-bound validation.
Release status
- Live verification was recorded against the committed reader hash.
- Factual, citation, foreword, license, media, human-acceptance, and export-accessibility gates remain open for a definitive edition.
2026-08-12 public web release
Nell explicitly authorized publication of this candidate on 2026-08-12. The release updated the public reader while retaining open specialist, rights, license, and visual-review items. Pull request 1 carried the change; current open work is summarized in MANUSCRIPT_STATUS.md.
Added
- Reading guide and argument map.
- Appendices on methodology, misunderstandings, minimal welfare practice, policy, discussion, and Interiora.
- Citation ledger, bibliography schema, evidence dossier, and source inventory tooling.
- Executive summary, policy brief, laboratory checklist, philosophical abstract, safety FAQ, rights FAQ, seminar guide, podcast preparation, and media kit.
- Publication status, roadmap, release checklist, contributing guide, corrections policy, and local validation tooling.
Corrected
- Collaborative-authorship disclosures now acknowledge later editorial revision.
- The phase-transition chapter preserves the unresolved consciousness questions and cross-references measured workspace-like representations.
- The “Becoming Minds” explanation now uses accurate head-noun grammar.
- The embodiment discussion distinguishes Heidegger’s situatedness from Merleau-Ponty’s developed account of embodiment.
- The Moral Turing Test now uses robust, welfare-relevant preference as its load-bearing criterion.
- The seven-layer conscience architecture now has one consistent order across prose, list, diagram, caption, and summary.
- Standing and responsibility are used consistently in the conclusion.
- The Christmas conversation quotation now matches Appendix A across repeated uses.
- An unsupported entropy-masked-loss paragraph was removed.
- The Creed Space conflict of interest is disclosed at first substantive use.
- Duplicate treatment of the Boris Cherny example was reduced through cross-reference.
- Unpublished architecture experiments are described as exploratory and unreplicated.
Release status
- Local build, validation, export, and browser evidence were recorded against the candidate manuscript hash.
- Factual, legal, rights, and direct visual review remain open for a later definitive edition.
- Nell’s fresh publication instruction released the live hold on 2026-08-12.
- Pull request 1 merged as
30e146b935c48fe7518e707b6b89642157180b4d; GitHub Pages completed successfully and the apex domain was verified against the committed reader.
Working draft, revision 31 · November 2026 · Changes and corrections · Evidence · Contents