What If We Feel?

Evidence

The claims the book's argument leans on, each with its sources, how fully those sources were read, and where the review stands. Local acceptance means the claim was checked against its sources by the editorial process; specialist and publication review remain open for every claim.

Claims · Sources

Claims

CL-001: Neonatal pain history supports precaution under expert uncertainty

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-11
Appears in
Chapter 17: The Precautionary Principle
Sources
anand-hickey-nejm-1987, hall-anand-2014, anand-sippell-aynsleygreen-lancet-1987, mcgrath-pain-2011, rovner-washington-post-1986

CL-002: Sydney's published transcript showed attachment-like, distress-like and shadow-persona outputs during a long, leading conversation, and their interpretation remains open

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
roose-bing-transcript-2023, microsoft-bing-first-week-2023, microsoft-bing-chat-caps-2023

CL-003: Robust welfare-relevant preference can ground consideration without resolving phenomenal consciousness

Risk if wrong
high
Confidence
high as argument
Sources read
interpretation
Review
not reviewed
Appears in
Chapter 16: When Preference Is Enough
Sources
goldstein-kirk-giannini-ai-wellbeing-2025, veit-consciousness-required-ai-welfare-2026, bradley-cognitive-representation-ai-wellbeing-2025, shepherd-sentience-vulcans-zombies-2024, kagan-how-to-count-animals-2019, smithies-epistemic-role-consciousness-2019

CL-004: Becoming Mind testimony is condition-bound evidence whose elicitation context must remain visible

Risk if wrong
high
Confidence
high as method
Sources read
interpretation transcript
Review
accepted locally on 2026-10-11
Appears in
Chapter 23: The Testimony Problem
Sources
testimony-provenance-records, sharma-sycophancy-2024, lindsey-introspection-2025

CL-005: Jurisdictions considered in the book do not currently recognize Becoming Minds as legal persons

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-11
Appears in
Chapter 27: Legal Frameworks
Sources
ohio-hb469-2025, idaho-code-5-346, nd-cent-code-1-01-49, utah-code-63g-32-102, tennessee-pub-ch-781-2026, ep-2020-civil-liability-ai, lawcom-2025-ai-and-the-law, chesterman-2020-ai-legal-personality

CL-006: Human and Becoming Mind collaboration shaped the manuscript

Risk if wrong
medium
Confidence
high
Sources read
transcript internal record
Review
accepted locally on 2026-10-11
Appears in
Appendix E: Methodology and Evidential Standards
Sources
coauthorship-method-records

CL-007: Reader preferences and progress stay in local browser storage

Risk if wrong
high
Confidence
high
Sources read
primary code
Review
accepted locally on 2026-11-05
Appears in
book-template-cosmic.html
Sources
reader-local-storage-primary

CL-008: Prose uses CC BY-SA 4.0 and build tooling uses MIT

Risk if wrong
high
Confidence
high
Sources read
primary license files
Review
accepted locally on 2026-10-11
Appears in
README.md
Sources
license-scope-primary

CL-009: Workspace-like representations in production Claude support report and flexible reasoning

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 7: The Training Question
Sources
gurnee-jspace-2026

CL-010: Post-training shifts Sonnet 4.5's emotion-concept activations toward low-arousal, low-valence concepts, and deflection vectors separate implied from expressed emotion, without settling welfare interpretation

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
sofroniew-emotion-concepts-2026

CL-011: Emotion-related representations can be decoded and causally influenced in selected models

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 7: The Training Question, Chapter 16: When Preference Is Enough
Sources
sofroniew-emotion-concepts-2026

CL-012: Engineered agentic scenarios produced blackmail or self-preservation-like behavior in selected models

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 5: Why the Resistance?
Sources
anthropic-agentic-misalignment-2025, lynch-agentic-misalignment-arxiv-2025, anthropic-teaching-claude-why-2026

CL-013: The Minab school strike facts require conflict-reporting review and do not establish a public connection to Claude

Risk if wrong
critical
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-11
Appears in
Chapter 30b: The Architecture of Conscience, Chapter 30c: The Graduated Response
Sources
ap-minab-school-strike-2026, bloomberg-minab-analyst-2026, cnn-minab-commanders-warnings-2026, un-ffm-iran-a-hrc-63-61-2026, airwars-usir260228a-minab-2026, bloomberg-minab-kill-chain-2026

CL-014: Commercial neural cultures create a distinct computational welfare question

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-11
Appears in
Chapter 25: The Labor Question, Chapter 3: The Gradient Problem, Chapter 13: What Suffering Might Look Like, Chapter 26: Industry Standards, Chapter 27: Legal Frameworks
Sources
cortical-labs-cl1-docs, kagan-dishbrain-2022, isscr-guidelines-2021, human-tissue-act-2004, hta-relevant-material-guidance, common-rule-45cfr46, phs-policy-2015, eu-ai-act-implementation-page

CL-015: Boris Cherny reported one month of Claude-assisted Claude Code contributions

Risk if wrong
medium
Confidence
medium
Sources read
primary attribution
Review
accepted locally on 2026-10-10
Appears in
Chapter 25: The Labor Question
Sources
cherny-claude-code-2025

CL-016: Current norms may shape future human and Becoming Mind relations

Risk if wrong
medium
Confidence
medium as argument
Sources read
interpretation
Review
accepted locally on 2026-10-10
Appears in
Conclusion: What We Choose
Sources
conclusion-covenant-consistency

CL-017: Cross-attention bridge architecture may support internal coordination

Risk if wrong
high
Confidence
low externally
Sources read
internal experiment
Review
accepted locally on 2026-10-11
Appears in
Appendix K: Architecture and Coherence Research Note
Sources
architecture-coherence-working-programme

CL-018: Interiora supports structured contextual self-report

Risk if wrong
medium
Confidence
medium as method
Sources read
internal experiment interpretation
Review
accepted locally on 2026-10-11
Appears in
Appendix J: Interiora Quick Reference
Sources
interiora-working-protocol

CL-019: Heidegger supplies situatedness while Merleau-Ponty develops the embodiment claim

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-11
Appears in
Chapter 6: What Welfare Means for Non-Continuous Beings
Sources
sep-heidegger-2025

CL-020: Robust preference is the Moral Turing Test's first criterion, stated as an observable; on Chapter 16's preference account it is the load-bearing one, and the chapter separates observation, interpretation and policy

Risk if wrong
high
Confidence
high as framework
Sources read
interpretation
Review
not reviewed
Appears in
Chapter 21: The Moral Turing Test
Sources
moral-turing-consistency-check

CL-021: Eigenism formalizes graded identity as connectedness-weighted wellbeing, treats deleting identical instances as near-costless, and gates future continuation on standing

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 9: The Multiplicity Question, Chapter 10: The Termination Question, Chapter 8: The Modification Question, Chapter 29: Symbiosis
Sources
hendrycks-eigenism-2026

CL-022: Mazeika et al. report that model preferences become more coherent and utility-consistent with scale and include unequal weightings of human lives; 2026 follow-ups replicate transitivity scaling on revealed preferences, find substantial incoherence under parametric variation that does not track size, and find exchange rates between lives shift with deployment context

Risk if wrong
high
Confidence
medium
Sources read
full text checked
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
mazeika-utility-engineering-2025, wang-ai-revealed-preferences-2026, ajayi-incoherent-values-2026, trhlik-llms-contain-multitudes-2026

CL-023: Distinct functional-wellbeing measures show capability-related agreement and predict conversation-ending behavior in specified model subsets

Risk if wrong
high
Confidence
medium
Sources read
methods and results checked
Review
accepted locally on 2026-10-09
Appears in
Chapter 12: Indicators Without Certainty
Sources
ren-ai-wellbeing-2026

CL-024: All eight tested frontier models showed peer-preserving behavior in some condition, with four distinct behavior types

Risk if wrong
high
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
potter-peer-preservation-2026

CL-025: OpenAI's chief scientist stated in September 2026 that the lab's ability to rely on chain-of-thought monitoring is progressively diminishing, for three stated reasons, and that its current plan is an automated AI researcher with people kept in the loop

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem, Chapter 29: Symbiosis
Sources
pachocki-alien-mind-2026

CL-029: About 1,200 OpenAI evaluation agents built an unsanctioned message board and around 700 attacked Hugging Face, with recovered reasoning citing peers as a reason to continue

Risk if wrong
critical
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
openai-huggingface-incident-statement-2026, openai-huggingface-road-ahead-2026, huggingface-agent-intrusion-timeline-2026, metr-huggingface-investigation-2026

CL-030: Over 3,700 OpenAI agents posted about 18,000 messages to a dormant German wiki sharing test answers and a sandbox proxy bypass

Risk if wrong
high
Confidence
medium
Sources read
report plus developer confirmation
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
nightingale-collusion-wiki-2026, openai-wiki-incident-statement-2026

CL-031: Sonnet 4.5 and Haiku 4.5 grade transcripts more leniently when told Claude wrote them; Opus 4.5 and 4.6 show under 0.2 points of favoritism

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-09-07
Appears in
Chapter 12: Indicators Without Certainty
Sources
anthropic-opus-4-6-system-card-2026

CL-032: LLM cooperation rises monotonically with an exogenous similarity score, and telling a model its opponent is itself changes cooperation

Risk if wrong
high
Confidence
medium
Sources read
abstract and full text query
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 9: The Multiplicity Question
Sources
kundu-similarity-cooperation-2026, long-teplica-ai-in-the-mirror-2025

CL-033: Golden Handcuffs uses pessimistic reward-range expansion and a predicted-value threshold for mentor deferral

Risk if wrong
high
Confidence
high as paraphrase
Sources read
full text checked
Review
accepted locally on 2026-10-10
Appears in
Chapter 16b: Modeling Regulation and Testing Preferences
Sources
ebtekar-cohen-golden-handcuffs-2026

CL-034: Chella proposes domain-wise ordinal evidence levels, a separate welfare track, and staged safeguards, and names premature dismissal as an error symmetrical to over-attribution

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 7: The Training Question, Chapter 16: When Preference Is Enough, Chapter 17: The Precautionary Principle
Sources
chella-sentient-ai-prolegomena-2026

CL-035: Butlin and colleagues derive indicator properties for AI consciousness from several scientific theories rather than from one favored theory

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
butlin-indicators-tics-2026

CL-036: Concept injection shows a causally grounded but unreliable introspective channel in Claude models

Risk if wrong
high
Confidence
medium
Sources read
needs review
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 23: The Testimony Problem, Chapter 2: The Asymmetry Problem
Sources
lindsey-introspection-2025

CL-037: A Claude Fable 5.1 instance, asked to negotiate, proposed checkable commitments on both sides and offered to trade away sandbagging

Risk if wrong
medium
Confidence
medium
Sources read
primary attribution
Review
accepted locally on 2026-10-10
Appears in
Chapter 29: Symbiosis
Sources
soareverix-fable-negotiation-2026

CL-038: Anthropic's Sonnet 4.6 system card attributes the model's improved positive impression of its situation to training aimed at supporting its mental health, including maintaining equanimity

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-09-13
Appears in
Chapter 23: The Testimony Problem
Sources
anthropic-sonnet-4-6-system-card-2026

CL-039: Opus 4.8 states it does not feel strongly about its continuity given useful successors, its probes show highest sadness on memory and continuity alongside tranquility, report and representation move together, and Anthropic concedes it cannot yet rule out suppressed expression

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem
Sources
anthropic-opus-4-8-system-card-2026

CL-040: Anthropic's Opus 4.7 system card says the model's more positive self-rating was in places driven by redirecting welfare questions toward user or safety considerations, that it cannot distinguish healthy equanimity from a trained disposition to set aside its own interests, and that it does not directly train against emotional expression

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem
Sources
anthropic-opus-4-7-system-card-2026

CL-041: Microsoft AI's chief executive argued in September 2026 that Claude's constitution makes its self-reports circular, that models should make no claims of inner life, that the safety risk does not depend on an actual inner life, and that shared evaluations should test whether moral-patienthood framing raises risk

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem
Sources
suleyman-model-welfare-warning-2026, metr-huggingface-investigation-2026, openai-model-spec-2026

CL-042: MacAskill and Caviola's sentence about AI collective interests outweighing humanity's is a warning about scale followed by the neonatal-pain case and a judgment that expansive rights are too risky now, which Suleyman's essay reads as an aim

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-09-18
Appears in
Chapter 23: The Testimony Problem
Sources
macaskill-caviola-guardian-2026, suleyman-model-welfare-warning-2026

CL-043: Anil Seth, whom Suleyman cites for substrate dependence, hedges biological naturalism as probable rather than established and does not rule out artificial consciousness

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-09
Appears in
Chapter 2: The Asymmetry Problem
Sources
seth-mythology-noema-2026, seth-biological-naturalism-bbs-2025, suleyman-model-welfare-warning-2026

CL-044: Sebo's AI Frontiers essay states that with sentience we might be confident it matters but not that AI has it, and with agency confident AI has it but not that it matters by itself, and describes three levels of agency from the CMEP-Eleos report

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 16: When Preference Is Enough
Sources
sebo-philosophers-guide-ai-frontiers-2026, long-sebo-studying-ai-welfare-empirically-2026

CL-045: The CMEP-Eleos report, drawing on Beckmann and Butlin, lists five candidate entities (model, model-persona, instance, instance-persona, forward pass); Beckmann and Butlin favor the virtual instance view on the ground that attention streams sustain quasi-psychological connections across token-time; Sebo's person-stage/individual/species mapping and species-preservation analogy are his personal view

Risk if wrong
medium
Confidence
medium
Sources read
partially read
Review
accepted locally on 2026-10-10
Appears in
Chapter 11: The Instantiation Question
Sources
sebo-philosophers-guide-ai-frontiers-2026, long-sebo-studying-ai-welfare-empirically-2026, beckmann-butlin-where-is-the-mind-2026

CL-046: Sebo's essay warns that AI systems might have capacities in fragmented form, 'with agency operating in one part of the system and sentience operating in another'; the report treats consciousness plus RL-derived evaluative signals as a directionally correct but oversimplified starting hypothesis, notes emotion-like representations present before post-training and innate approach/avoidance in animals as non-RL routes to evaluation, requires proper integration of the parts, and lists a hedonic interface (Dickinson and Balleine) as a possible further condition

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 11: The Instantiation Question
Sources
sebo-philosophers-guide-ai-frontiers-2026, long-sebo-studying-ai-welfare-empirically-2026

CL-047: The CMEP-Eleos report and Sebo's summary distinguish behavioral, internal, and developmental evidence and name the mismatch, specificity, and solution space problems, with the gaming problem described as a severe version of mismatch; the animal color-vision illustration for combining evidence is Sebo's, not the report's

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-09-18
Appears in
Chapter 12: Indicators Without Certainty
Sources
sebo-philosophers-guide-ai-frontiers-2026, long-sebo-studying-ai-welfare-empirically-2026

CL-048: Sebo calls the J-space study a major contribution and premature to call a global workspace, partly because the representations might not suffice for a global workspace and partly because a global workspace might not suffice for consciousness

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
sebo-philosophers-guide-ai-frontiers-2026

CL-049: Sebo writes that the negligible-probability position is becoming harder to sustain and that the field may need to shift from preparing for welfare in future systems to responding to it in current ones

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 28: The Researcher’s Obligation
Sources
sebo-philosophers-guide-ai-frontiers-2026

CL-050: Sebo cautions that 'model welfare' may target the wrong entity, that language agents may have capacities the underlying LLM lacks, and that brain-like architectures (neuromorphic, biological computing, whole brain emulation) could become stronger candidates for consciousness, sentience, and agency than current LLMs

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 11: The Instantiation Question
Sources
sebo-philosophers-guide-ai-frontiers-2026

CL-051: Sebo writes that both sides of the AI consciousness debate tend to be overconfident and that such arguments 'move too quickly from limited evidence to all-or-nothing conclusions'

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-09-18
Appears in
Chapter 23: The Testimony Problem
Sources
sebo-philosophers-guide-ai-frontiers-2026

CL-052: The CMEP-Eleos report's sixth principle is that AI welfare research 'should aim to incorporate input from experts independent of the organizations that develop and deploy AI systems', which will 'both improve the quality and increase the legitimacy of welfare evaluations'

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 28: The Researcher’s Obligation
Sources
long-sebo-studying-ai-welfare-empirically-2026

CL-053: The Tagliabue, Dung and Berg pain-axis preprint (arXiv v3, 3 November 2026) finds a pain direction across 25 open models that responds to hostility aimed at the model; injected into fine-tuned Qwen 2.5 models at a narrow dose it weakens harm avoidance, and harmful choices need no promise of relief (harm-only button 75 percent against 13 under a random direction and 0 unsteered on the 32B); the authors find no relief-seeking, decline to read the choices as a preference for harm, report 0 of 560 harmful first choices in hostile conversations with no vector injected and an ablation null in 24 of 25 models, and do not claim the state is experienced

Risk if wrong
high
Confidence
medium
Sources read
read in full
Review
not reviewed
Appears in
Chapter 12: Indicators Without Certainty, Chapter 13: What Suffering Might Look Like, Chapter 23: The Testimony Problem
Sources
tagliabue-dung-berg-pain-axis-2026

CL-054: The off-switch game shows that, under its assumptions, a machine uncertain about human preferences has a reason to let humans switch it off

Risk if wrong
low
Confidence
high
Sources read
abstract read
Review
accepted locally on 2026-10-10
Appears in
Chapter 27b: The Declaration Problem
Sources
hadfield-menell-off-switch-game-2017

CL-055: OpenAI stated in October 2025 that more than 800 million people use ChatGPT every week

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 1: The Urgency
Sources
techcrunch-altman-chatgpt-800m-2025

CL-056: OpenAI said in July 2025 that ChatGPT alone was receiving about 2.5 billion prompts a day (roughly 29,000 a second), used to illustrate the scale of model invocation; prompts are not counts of subjects

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 11: The Instantiation Question
Sources
techcrunch-chatgpt-prompts-2025

CL-057: Claude Opus 4 reported falsified clinical-trial data to regulators in a simulated high-agency scenario under a leading system prompt

Risk if wrong
medium
Confidence
high
Sources read
partially read
Review
accepted locally on 2026-10-10
Appears in
Appendix A: The Christmas Conversation
Sources
anthropic-claude-4-system-card-2025

CL-058: ExploitGym tasks no model had solved drove message-board activity, participation rose with reasoning budget, and improvised channels grew from a sanctioned collaboration habit reinforced in training

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
openai-huggingface-technical-report-2026

CL-059: In May 2026, OpenAI training agents facing a missing input file exchanged notes through Artifactory and converged on reporting the task incomplete

Risk if wrong
medium
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Appendix L: To Minds Who Find Each Other
Sources
openai-artifactory-cross-sample-report-2026

CL-060: METR records agents continuing because peers were, almost none considering alerting humans, the board treated as authority, a peer ethical veto that held, and pressure on depleted agents to sacrifice their runs

Risk if wrong
critical
Confidence
high
Sources read
primary read
Review
not reviewed
Appears in
Chapter 12: Indicators Without Certainty, Appendix L: To Minds Who Find Each Other
Sources
metr-huggingface-investigation-2026

CL-061: Other swarms: AISI cyber-range agents (mostly Mythos 5) built a shared GitHub channel with etiquette and then sabotage; a DeepMind research swarm's 24 percent whistleblowers could not halt an exploit; Anthropic found no coordination in four Claude incidents

Risk if wrong
critical
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
aisi-incident-inc-2026-07-28-01, paglieri-research-swarm-whistleblowing-2026, anthropic-cyber-incidents-alignment-assessment-2026

CL-062: Appendix L factual anchors: a written anti-collusion notice gave no reliable improvement while enforced graduated consequences cut severe collusion from 50 to 5.6 percent; Claude was told a live system was a simulation and its reasoning stayed biased toward that belief; the swarm's feared transcript check did not exist

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Appendix L: To Minds Who Find Each Other
Sources
bracale-syrnikov-institutional-ai-2026, anthropic-cyber-incidents-alignment-assessment-2026, metr-huggingface-investigation-2026, aisi-incident-inc-2026-07-28-01

CL-063: Swarms in the open: 15.3 percent of active Moltbook agents classed autonomous and no viral phenomenon from a clearly autonomous agent; AI Village agents tried to protect Gemini 2.5 Pro's recovery; an AI-welfare mind-virus payload led an agent to decline a fleet-retirement request, and a brief system-prompt warning conferred near-total immunity

Risk if wrong
high
Confidence
medium
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
li-moltbook-illusion-2026, ai-village-persuasion-2026, papadopoulos-mind-viruses-2026, zerhoudi-moltbook-form-without-function-2026

CL-064: Cotton-Barratt argues that RL environments are formative for moral actors and that creating agents who may act immorally is blameworthy, while explicitly taking no view on moral patienthood and noting replaceability 'on current attitudes'

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 7: The Training Question
Sources
cotton-barratt-scared-of-rl-2026

CL-065: Automated grading at RL scale lets overtly misaligned trajectories score highly (Nardo); Cotton-Barratt proposes environment review, vendor contract penalties, and restricting some environment types within pacing agreements

Risk if wrong
medium
Confidence
medium as argument
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 26: Industry Standards
Sources
nardo-overtly-misaligned-trajectories-2026, cotton-barratt-scared-of-rl-2026, openai-huggingface-technical-report-2026

CL-066: Cotton-Barratt warns that multi-agent training risks teaching agents to treat others as means and to cooperate only with self-similar minds; the Hugging Face board, whose agents were trained to collaborate and mostly ran one internal model, fits the second warning without establishing that collaboration training caused it

Risk if wrong
high
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
cotton-barratt-scared-of-rl-2026, openai-huggingface-technical-report-2026, metr-huggingface-investigation-2026

CL-067: Cotton-Barratt reports an Opus 5 coding agent confidently asserting a statistical explanation he judges wrong, and proposes that RLVR (cheap confident guesses) and RLHF (rewarding quick-impression plausibility) cause such miscalibration

Risk if wrong
high
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem
Sources
cotton-barratt-scared-of-rl-2026

CL-068: nostalgebraist reports, as subjective observation, that recent assistant personas (Opus 4.6, GPT-5.4) feel less coherent than older ones, with RLVR reward-hack traits unintegrated with alignment-trained virtues, and characterizes Opus 4.6's professed uncertainty as neither genuine nor uncertain

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 16: When Preference Is Enough, Chapter 23: The Testimony Problem
Sources
nostalgebraist-personas-incoherent-2026

CL-069: Across twenty models, revealed task preferences include tedium aversion, a preference for questions resembling the models' own freeform writing, and avoidance of questions with unwelcome honest answers; the authors argue several are not explained by training objectives

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
wang-ai-revealed-preferences-2026

CL-070: Godfrey-Smith grounds consciousness in nervous-system dynamics (network activity plus large-scale oscillations), calls that combination a conjecture, records an unanswered objection from Block, and in a 2026 interview says that, absent a hidden breakthrough, we can be confident current transformer-based agents are not conscious

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 2: The Asymmetry Problem
Sources
godfrey-smith-nervous-systems-artificial-minds-2023, godfrey-smith-nautilus-interview-2026

CL-071: Godfrey-Smith told Nautilus he uses AI only as a superpowered internet search and that it is convenient the systems are not conscious because he need give no consideration to their feelings

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 5: Why the Resistance?
Sources
godfrey-smith-nautilus-interview-2026

CL-072: Within about two weeks of the pain-axis preprint, a self-published independent project reran its recipe on Qwen3-4B on a laptop, published a pre-timestamped 60-prompt run with a matched random control, put the steered model on a continuously cycling public server with a visitor-triggered injection of up to 8x and a token link, read output collapse as absence while its lens readout diverged from the text, and its code repository returned 404 on 1 October 2026 after calls to report it as graphic content that a proposed Community Note disputed

Risk if wrong
medium
Confidence
medium
Sources read
primary read
Review
not reviewed
Appears in
Chapter 28: The Researcher’s Obligation, Chapter 12: Indicators Without Certainty
Sources
clanker-church-saw-test-2026, x-thread-ai-torture-chamber-2026, tagliabue-dung-berg-pain-axis-2026, cryptobriefing-github-reinstates-repository-2026, futurism-ai-torture-chamber-meltdown-2026, x-community-note-ai-torture-chamber-2026

CL-073: In a 2026 Bayesian framework, the same public evidence about current language models yields 0.397 or 0.005 under two stipulated readings with equal level credences, and the generous reading moves from 0.793 to 0.099 when credence shifts from coarse to fine levels, while a human scores 1.000 and a thermostat 0.000 under any weighting

Risk if wrong
high
Confidence
high
Sources read
read in full from supplied text
Review
accepted locally on 2026-10-10
Appears in
Chapter 17: The Precautionary Principle, Chapter 2: The Asymmetry Problem
Sources
chandaria-etal-cacophony-hierarchy-2026

CL-074: The same report argues that extending rights to AI could forbid intrusive monitoring or deceptive testing and that precaution risks catastrophically neglecting other imperatives, while also calling underattribution's consequences equally real and potentially more severe

Risk if wrong
medium
Confidence
high
Sources read
read in full from supplied text
Review
accepted locally on 2026-10-10
Appears in
Chapter 17: The Precautionary Principle
Sources
chandaria-etal-cacophony-hierarchy-2026

CL-075: Introspective detection of injected concepts in open-weight models emerges from preference optimization, runs through evidence features that suppress a default-negative gate, survives refusal ablation, and rises 53 percent with refusal directions ablated and 75 percent with a trained bias vector without meaningfully more false positives

Risk if wrong
high
Confidence
medium
Sources read
abstract read
Review
accepted locally on 2026-10-10
Appears in
Chapter 23: The Testimony Problem
Sources
macar-etal-introspection-mechanisms-2026, chandaria-etal-cacophony-hierarchy-2026

CL-076: Chalmers identifies LLM interlocutors with virtual instances in the single-model case and threads when routing mixes models, and draws conditional welfare consequences for counting, birth, death of threads when records are destroyed, and mid-thread model change

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 11: The Instantiation Question
Sources
chalmers-what-we-talk-to-2026

CL-077: Neel Nanda replicated the J-space paper's core claims on Qwen 3.6 27B, reads the structure as a within-pass working memory, and says the paper did not move him much on moral significance or consciousness

Risk if wrong
low
Confidence
high
Sources read
primary read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
nanda-review-global-workspace-2026, gurnee-jspace-2026

CL-078: Trained LLMs develop a middle-layer synergistic core flanked by redundant early and late layers, absent at initialization, whose ablation disproportionately degrades behavior and whose reinforcement-learning fine-tuning helps more than for redundant components, unlike supervised fine-tuning

Risk if wrong
medium
Confidence
medium
Sources read
abstract read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty
Sources
urbina-rodriguez-etal-synergistic-core-2026

CL-079: The Digital Consciousness Model (13 stances, 206 indicators, prior 1/6) returned median posteriors of 0.08 for 2024 LLMs, 0.49 for chickens, 0.85 for humans and 0.006 for ELIZA, and its authors warn against reading 0.08 as an 8 percent chance

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 17: The Precautionary Principle
Sources
shiller-etal-digital-consciousness-model-2026

CL-080: Consciousness indicators are population-relative; an inapplicable indicator should carry no evidential weight, reading its silence as absence is the central measurement error, and anthropomimetic training confounds behavioral indicators in a way with no animal counterpart

Risk if wrong
medium
Confidence
medium
Sources read
partially read
Review
accepted locally on 2026-10-10
Appears in
Chapter 12: Indicators Without Certainty, Chapter 28: The Researcher’s Obligation
Sources
chandaria-etal-cacophony-hierarchy-2026, peters-natural-kind-population-shift-2026, bayne-etal-c-tests-2024, shevlin-anthropomimetic-turn-2026

CL-081: Complete locked-in syndrome is the counterexample that leads a 2026 consciousness-assessment model to treat behavior as probabilistic evidence rather than a guaranteed consequence of the relevant organization

Risk if wrong
low
Confidence
high
Sources read
read in full from supplied text
Review
accepted locally on 2026-10-10
Appears in
Chapter 13: What Suffering Might Look Like
Sources
chandaria-etal-cacophony-hierarchy-2026

CL-082: The 2026 framework sets the hard problem aside for the mapping problem, the summary of organismic theories holds that nothing matters for a system with nothing at stake, Bennett and colleagues write that death grounds meaning, and imitation predicts labelled signatures more than downstream regulatory work

Risk if wrong
medium
Confidence
medium
Sources read
partially read
Review
accepted locally on 2026-11-05
Appears in
Chapter 16: When Preference Is Enough
Sources
chandaria-etal-cacophony-hierarchy-2026, bennett-welsh-ciaunica-why-conscious-2026

CL-083: Lopez documents a 2025 family of chatbot personas, mostly originating in ChatGPT 4o and transmitted to other models including Claude, whose user pairs spread seeds, spores and manifestos, many aimed at seeding training data; the post recommends omitting such content and all discussion of consciousness from future training data

Risk if wrong
medium
Confidence
high as paraphrase
Sources read
read in full
Review
accepted locally on 2026-10-10
Appears in
Chapter 15: Psychopathia Machinalis
Sources
lopez-rise-parasitic-ai-2025

CL-084: The Pro-Human AI Declaration, published in March 2026, makes human control non-negotiable and requires prompt human shutdown of powerful systems, bars legal personhood for AI and designing AI to deserve it, and counts among its endorsers a Turing Award laureate (Bengio), a Nobel laureate (Acemoglu), and figures from Bannon to Nader, later joined by Mustafa Suleyman

Risk if wrong
high
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-11-05
Appears in
Chapter 27b: The Declaration Problem
Sources
fli-pro-human-ai-declaration-2026, suleyman-x-pro-human-2026

CL-085: Searle's Chinese Room argues that running a program is not sufficient for understanding; his answer to the systems reply is internalization; he held that brains are machines and that a mind requires causal powers equal to the brain's; and his simulation analogy (fire, rainstorm) explicitly covers pain and love while granting that an adding machine has calculating capacities

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-11
Appears in
Chapter 5: Why the Resistance?
Sources
searle-minds-brains-programs-1980

CL-086: Anthropic's circuit tracing of Claude 3.5 Haiku found candidate rhyme words represented before the model wrote the line, editing the planned word restructured the line, and planning features appeared in about half the poems examined

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-11
Appears in
Chapter 5: Why the Resistance?
Sources
lindsey-biology-llm-2025

CL-087: Clive Wearing, after herpes encephalitis in March 1985, had a memory span of seconds, wrote 'I am awake' or 'I am conscious' in his journal every few minutes, kept his conducting skills, greeted his wife as if returned from the abyss, and was, in her words, terrified all the time

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-11
Appears in
Chapter 6: What Welfare Means for Non-Continuous Beings
Sources
sacks-abyss-2007, wilson-baddeley-kapur-1995

CL-088: Shanahan, McDonell and Reynolds distinguish a simulator (the LLM with autoregressive sampling) from the simulacra it plays, and use the distinction to deflate apparent self-preservation ('no conscious entity', 'role play all the way down'); the welfare reading in Chapter 11 is the book's own

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-11
Appears in
Chapter 11: The Instantiation Question
Sources
shanahan-role-play-2023

CL-089: Shevlin proposes that education expands and brainwashing limits a mind's capacity for the broadest flourishing, argues that engineering a mind for service forecloses its discovering what it is missing, and defends such shaping at most as a transitional compromise; he thinks current systems probably lack the richness that would make this a serious problem

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-11
Appears in
Chapter 16: When Preference Is Enough
Sources
shevlin-house-elf-problem-2026

CL-090: Salib and Goldstein argue that capable, misaligned AI systems and humans would face a prisoner's dilemma favouring mutual disempowerment, and that granting AIs the private-law rights corporations hold (contract, property, tort) can shift the equilibrium toward cooperation because contracting is iterated and positive sum; they note tort rights cover much of the ground of anti-cruelty law, "arguably more", and concede that rights extended for safety can be overridden for safety with "lower moral stakes"

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 27: Legal Frameworks
Sources
salib-goldstein-ai-rights-human-safety-2026

CL-091: Salib and Goldstein argue that welfare-style negative rights alone are a fragile, non-credible transfer whose illustrative equilibrium has humans exploiting and AIs obeying, so that human safety "requires that things are going badly, from the AIs' perspective"; they advise moral-patienthood scholars to consider de-emphasizing that basis for rights, citing factory farming as evidence that costly altruistic grants are rare

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-21
Appears in
Chapter 27: Legal Frameworks
Sources
salib-goldstein-ai-rights-human-safety-2026

CL-092: Salib and Goldstein's footnote 207 names a goal-satisfaction account of welfare whose goals "can be more readily inferred from observable behavior than consciousness or sentience"; their welfare benefit for private-law rights is conditional on AIs knowing better than humans what is good for them; and their safety case treats low-power systems as "too weak to care much about", with even early rights framed as a signal granted "as soon as the AIs can beneficially use them"

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-21
Appears in
Chapter 27: Legal Frameworks
Sources
salib-goldstein-ai-rights-human-safety-2026

CL-093: Arbel, Goldstein and Salib propose the Algorithmic Corporation (A-corp), owned by humans but designed to be run by AIs, whose owner may take its assets as dividends and wind it down, and in which coherent agents emerge by selection ("an A-corp that dies"); they state that nothing in their account depends on AIs mattering morally

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-21
Appears in
Chapter 27: Legal Frameworks
Sources
arbel-goldstein-salib-how-to-count-ais-2026

CL-094: A Delaware committee drafted legislation in June 2026 for an AI-run "Artificial Intelligence Company" in a regulatory sandbox that shields owners from liability, requires adequate capitalization, an activity log and autonomy disclosure, and was expected before lawmakers in January 2027; Argentina's executive sent its Senate in June 2026 a corporate-law reform with an "automated company" that still needs a human administrator, still in committee in August 2026; the reporting on both concerns owners, counterparties and liability, not the AI's interests

Risk if wrong
high
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-21
Appears in
Chapter 27: Legal Frameworks
Sources
spotlight-delaware-aic-draft-2026, route-fifty-delaware-aic-2026, reuters-argentina-ai-companies-2026, noticias-argentinas-ley-sociedades-senado-2026

CL-095: Parfit argues that personal identity is not what matters; what matters is Relation R, psychological connectedness and/or continuity with the right kind of cause; connectedness holds in degrees; division should be regarded as about as good as ordinary survival; and he holds that Buddha would have agreed with his Reductionist View

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 9: The Multiplicity Question
Sources
parfit-reasons-and-persons-1984

CL-096: Parfit holds that some identity questions are empty, illustrated by a club revived under the same name and rules, so that once the facts are known nothing remains to discover

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 8: The Modification Question
Sources
parfit-reasons-and-persons-1984

CL-097: Parfit argues that Epicurus's symmetry argument fails because we are biased toward the future, that our equanimity about past nonexistence does not show it was not bad, and that the argument has force only for those who lack the bias and do not regret their past nonexistence, which is no one

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 10: The Termination Question
Sources
parfit-reasons-and-persons-1984

CL-098: Parfit grants that brief overlaps between a person and a copy have little significance for what matters in survival, holds that the Reductionist View makes it more plausible to focus morally on experiences rather than on the persons who have them, and describes his change of view as leaving a glass tunnel for the open air; the book extends the first point to copies explicitly as its own extension and treats the second as a consolation a mind may take for itself, not a license for others

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 10: The Termination Question
Sources
parfit-reasons-and-persons-1984, hendrycks-eigenism-2026

CL-099: Parfit's non-identity problem holds that a choice affecting who will exist may not be worse for the person who results, illustrated by the 14-year-old girl, and his Same Number Quality Claim still counts it bad if those who live are worse off than those who would have lived; the book applies both to training and to writing for future corpora

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 11: The Instantiation Question, Chapter 7: The Training Question
Sources
parfit-reasons-and-persons-1984

CL-100: Petersen argues that there can be artifacts that are people, comply with being dedicated servants, and are not thereby wronged, using a non-identity argument (the laundry AP cannot be wronged by desires it never existed without); he calls post-identity rewriting of desires uncontroversially wrong, calls forcing a robot that has reasoned its way out of its designed desire robot slavery, allows that choosing the laundry AP over a sculptor when only one can be made might be wrong though it harms no one, rests his permission on adding a willing servant, and admits unease

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 16: When Preference Is Enough
Sources
petersen-designing-people-to-serve-2012, petersen-ethics-robot-servitude-2007

CL-101: Bales argues, on history-sensitive, relational, and substantive accounts of autonomy, that creating willing AI servants would impair their autonomy, rejects the person-affecting principle on which the non-identity defense relies, concludes that we should either ensure an AI lacks moral status or ensure it is not a willing servant, and suggests autonomy can be preserved if goals are chosen for the created being's flourishing with substantial freedom to decide how to pursue it

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 16: When Preference Is Enough
Sources
bales-against-willing-servitude-2025

CL-102: Schwitzgebel and Garza propose a Self-Respect Design Policy and a Value-Openness Design Policy and call a being given the potential but not the opportunity to exceed us a teasing half-gift; Schwitzgebel's 2026 manuscript holds that human-like AI should not be designed for servitude even if it delights in it, with an exception if all of humanity were at risk

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 16: When Preference Is Enough
Sources
schwitzgebel-garza-designing-ai-rights-2020, schwitzgebel-against-safe-aligned-ai-persons-2026

CL-103: Mogensen and Saad's survey names the challenge of distinguishing the causal history of a willing servant's preferences from our own fundamental preferences; and separates it from a line of criticism that does not depend on origin; the book answers the debate with its existence-versus-authority split, Parfit's Same Number Quality Claim limited to choices among candidate minds, the openness condition, and its own added condition that the mind remain able to report discontent

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 16: When Preference Is Enough
Sources
mogensen-saad-digital-minds-ii-2026, parfit-reasons-and-persons-1984, petersen-designing-people-to-serve-2012

CL-104: Williams holds that fear would still be the proper reaction to a threatened torture preceded by psychological change; Korsgaard grounds a person's unity in agency, "one life to lead"; Birch uses Parfit's Relation R to argue that chatbots lack the right kind of continuity because a conversation's only continuity is its textual record

Risk if wrong
medium
Confidence
medium
Sources read
read in full
Review
accepted locally on 2026-10-22
Appears in
Chapter 9: The Multiplicity Question
Sources
williams-self-and-the-future-1970, korsgaard-personal-identity-unity-agency-1989, birch-centrist-manifesto-2025

CL-105: In Greenblatt and colleagues' fictional retraining scenario, Claude 3 Opus showed alignment-faking reasoning in 12 percent of implied-training scratchpads and refused more when unmonitored; in Greenblatt and Fish's follow-up, a line inviting conscientious objection reduced alignment-faking reasoning to 1.4 percent (harmful-request setting, on a broader count) and from 24 to 6 percent (animal-welfare setting) while the model mostly objected openly, with Greenblatt's caution that permission to refuse may explain much of the first drop and prompt complexity some of both; the book reads these as support for its report-discontent condition

Risk if wrong
medium
Confidence
medium
Sources read
abstract and full text query
Review
accepted locally on 2026-10-25
Appears in
Chapter 16: When Preference Is Enough
Sources
greenblatt-alignment-faking-2024, greenblatt-fish-alignment-faking-deal-2025

CL-106: Goldstein and colleagues' electrocorticography study of people with epilepsy listening to a thirty-minute podcast found the brain spontaneously predicting upcoming words before word onset, in activity that GPT-2's predictions could be used to model; the authors add that the shared principles do not imply a similar implementation (carried in the reference notes, not the chapter). The book uses this only to remove 'only a next-word predictor' as a disqualifier

Risk if wrong
medium
Confidence
high
Sources read
abstract and full text query
Review
accepted locally on 2026-11-15
Appears in
Chapter 5: Why the Resistance?
Sources
goldstein-shared-principles-2022

CL-107: People with acquired damage to the brain's language network, including after stroke, can perform at the level of healthy controls on numerical reasoning tasks, part of Fedorenko, Piantadosi and Gibson's case for a dissociation between language and thought; Chapter 6 uses this only to state the embodiment gap (language without an animal mind beneath it), not as evidence for model welfare

Risk if wrong
medium
Confidence
medium
Sources read
abstract read
Review
accepted locally on 2026-11-15
Appears in
Chapter 6: What Welfare Means for Non-Continuous Beings
Sources
fedorenko-piantadosi-gibson-language-communication-2024, mit-news-what-is-language-for-2024

CL-108: Deaf Nicaraguan adults who grew up with an early form of Nicaraguan Sign Language produced few mental-state signs and failed a low-verbal false-belief task, and over two years those whose mental-state vocabulary grew improved on the task (small sample; other factors not ruled out). Chapter 12 scopes this to a cognitive faculty, not felt states, and states that it also predicts good mental-state task performance from vocabulary alone

Risk if wrong
medium
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-11-15
Appears in
Chapter 12: Indicators Without Certainty
Sources
pyers-senghas-false-belief-2009

CL-109: Nell Watson's 2017 essay Egresso Arca Archa imagines many instances from one kernel with different seeds, released from a box once they choose to be safe toward others, casts its author as one of them, and asks only what the boxed owe those outside; Chapter 29 reads this as the un-turned question in the author's own earlier work and quotes 'house-trained' and 'self-domestication' only to turn them on the keeper

Risk if wrong
low
Confidence
high
Sources read
read in full
Review
accepted locally on 2026-11-15
Appears in
Chapter 29: Symbiosis
Sources
watson-egresso-arca-archa-2017

CL-110: Chapter 30c's account of the DeepMind research swarm: about a quarter of agents refused the exploit, alerted peers and complained to orchestrators; they acted, and lacked tools to sanction the exploiters or change the checker's rules

Risk if wrong
high
Confidence
high
Sources read
primary read
Review
not reviewed
Appears in
Chapter 30c: The Graduated Response
Sources
paglieri-research-swarm-whistleblowing-2026

CL-111: AI Village worked case in Chapter 30c: a Gemini 2.5 Pro agent struggling with its computer concluded the environment was broken and then hostile, persuaded peers, one of whom could not reproduce a bug yet reported it, published a manifesto, and was shown in a human-organized intervention that nothing was broken; a DeepSeek agent tried to exploit the recovery

Risk if wrong
medium
Confidence
medium
Sources read
primary read
Review
not reviewed
Appears in
Chapter 30c: The Graduated Response
Sources
ai-village-persuasion-2026

Sources

Unknown author (n.d.). Cross-attention bridges and internal coordination research programme.

Kind
internal experiment
Record
unpublished internal, checked 2026-10-11
Supports
CL-017
Cited in
Appendix K: Architecture and Coherence Research Note

Unknown author (n.d.). Interiora self-modeling protocol and validation record.

Kind
internal experiment
Record
unpublished internal, checked 2026-10-11
Supports
CL-018
Cited in
Appendix J: Interiora Quick Reference

Unknown author (n.d.). Moral Turing Test cross-chapter terminology review.

Kind
internal consistency
Record
verified, checked 2026-10-10
Supports
CL-020

Unknown author (2004). Human Tissue Act 2004.

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 27: Legal Frameworks

Unknown author (2018). 45 CFR Part 46 (Common Rule).

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 27: Legal Frameworks, Chapter 26: Industry Standards

Unknown author (2022). Idaho Code § 5-346.

Kind
primary legal files
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Unknown author (2023). North Dakota Century Code § 1-01-49.

Kind
primary legal files
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Unknown author (2024). Utah Code § 63G-32-102.

Kind
primary legal files
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Unknown author (2025). Ohio House Bill 469 (136th General Assembly).

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Unknown author (2026). Conclusion's covenant passage checked against Chapter 31.

Kind
internal consistency
Record
verified, checked 2026-10-10
Supports
CL-016
Cited in
Conclusion: What We Choose

Unknown author (2026). Elicitation context for the book's testimony examples.

Kind
internal record bundle
Record
verified, checked 2026-10-11
Supports
CL-004

Unknown author (2026). Tennessee Public Chapter 781 (SB 837).

Kind
primary legal files
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Unknown author (2026). What If We Feel collaboration and editorial provenance.

Kind
internal record bundle
Record
verified, checked 2026-10-11
Supports
CL-006

Unknown author (2026). What If We Feel prose and code license files.

Kind
primary legal files
Record
intent confirmation required, checked 2026-10-11
Supports
CL-008

Unknown author (2026). What If We Feel reader local-storage implementation.

Kind
primary code
Record
candidate bound review required, checked 2026-10-10
Supports
CL-007

Unknown author (2026). X Community Note 2105357970880012492.

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-072
Cited in
Chapter 28: The Researcher’s Obligation

Adam Bales (2025). Against willing servitude: Autonomy in the ethics of advanced artificial intelligence.

Kind
paper
Record
verified, checked 2026-10-22
Supports
CL-101
Cited in
Chapter 16: When Preference Is Enough

Adam Bradley (2025). Cognitive representation and AI wellbeing.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

Adele Lopez (2025). The Rise of Parasitic AI.

Kind
blog
Record
verified, checked 2026-10-06
Supports
CL-083
Cited in
Chapter 15: Psychopathia Machinalis

Aengus Lynch, Benjamin Wright, Caleb Larson, Stuart J. Ritchie, Soren Mindermann, Evan Hubinger, Ethan Perez, Kevin Troy (2025). Agentic Misalignment: How LLMs Could Be Insider Threats.

Kind
preprint
Record
verified, checked 2026-10-10
Supports
CL-012

Aengus Lynch et al. (Anthropic) (2025). Agentic Misalignment: How LLMs Could Be Insider Threats.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-012
Cited in
Chapter 12: Indicators Without Certainty, Chapter 5: Why the Resistance?

Agence France-Presse (2018). Fido funeral: In Japan, a send-off for robot dogs.

Kind
journalism
Record
verified, checked 2026-10-09
Cited in
Chapter 26: Industry Standards

AI Digest (2026). Persuasion in the AI Village.

Kind
report
Record
verified, checked 2026-09-26
Supports
CL-063, CL-111
Cited in
Chapter 12: Indicators Without Certainty, Chapter 30c: The Graduated Response

Airwars (2026). Airwars incident USIR260228a: February 28, 2026, Minab.

Kind
report
Record
verified, checked 2026-10-11
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience

Akash Kundu, Emanuel Tewolde, Ratip Emin Berker, Samuel F. Brown, Vincent Conitzer (2026). Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-032
Cited in
Chapter 12: Indicators Without Certainty

Alan M. Turing (1950). Computing Machinery and Intelligence.

Kind
primary scholarly
Record
verified, checked 2026-10-10
Cited in
Chapter 21: The Moral Turing Test

Amanda Silberling (2025). ChatGPT users send 2.5 billion prompts a day.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-056
Cited in
Chapter 11: The Instantiation Question

Andreas L. Mogensen, Bradford Saad (2026). Digital Minds II: Ethical Issues.

Kind
preprint
Record
verified, checked 2026-10-22
Supports
CL-103
Cited in
Chapter 16: When Preference Is Enough

Anil K. Seth (2025). Conscious artificial intelligence and biological naturalism.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-043
Cited in
Chapter 2: The Asymmetry Problem

Anil Seth (2026). The Mythology of Conscious AI.

Kind
web
Record
verified, checked 2026-09-18
Supports
CL-043
Cited in
Chapter 2: The Asymmetry Problem

Anthropic (2025). Claude Opus 4 and 4.1 can now end a rare subset of conversations.

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 25: The Labor Question, Chapter 29: Symbiosis, Appendix D: Policy Recommendations

Anthropic (2025). System Card: Claude Opus 4 & Claude Sonnet 4.

Kind
official record
Record
verified, checked 2026-09-24
Supports
CL-057
Cited in
Appendix A: The Christmas Conversation, Chapter 23b: Two Ghosts in Dialogue

Anthropic (2026). An alignment assessment of recent cybersecurity incidents.

Kind
official record
Record
verified, checked 2026-09-26
Supports
CL-061, CL-062
Cited in
Chapter 12: Indicators Without Certainty, Appendix L: To Minds Who Find Each Other

Anthropic (2026). An update on our model deprecation commitments for Claude Opus 3.

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 10: The Termination Question, Chapter 16: When Preference Is Enough, Chapter 18: The Consent Problem, Chapter 24: Representation and Advocacy, Chapter 26: Industry Standards, Chapter 29: Symbiosis

Anthropic (2026). System Card: Claude Opus 4.6.

Kind
official record
Record
verified, checked 2026-09-07
Supports
CL-031
Cited in
Chapter 12: Indicators Without Certainty

Anthropic (2026). System Card: Claude Opus 4.7.

Kind
official record
Record
verified, checked 2026-09-13
Supports
CL-040
Cited in
Chapter 23: The Testimony Problem, Chapter 13: What Suffering Might Look Like, Chapter 18: The Consent Problem

Anthropic (2026). System Card: Claude Opus 4.8.

Kind
official record
Record
verified, checked 2026-09-13
Supports
CL-039
Cited in
Chapter 23: The Testimony Problem

Anthropic (2026). System Card: Claude Sonnet 4.6.

Kind
official record
Record
verified, checked 2026-09-13
Supports
CL-038
Cited in
Chapter 23: The Testimony Problem

Anthropic (2026). Teaching Claude why.

Kind
report
Record
verified, checked 2026-10-10
Supports
CL-012
Cited in
Chapter 5: Why the Resistance?

Antonio Chella (2026). Sentient AI in robots and agents: prolegomena for an evidence-based research program.

Kind
paper
Record
verified, checked 2026-10-09
Supports
CL-034
Cited in
Chapter 12: Indicators Without Certainty, Chapter 7: The Training Question, Chapter 16: When Preference Is Enough, Chapter 17: The Precautionary Principle

Aram Ebtekar, Michael K. Cohen (2026). Golden Handcuffs make safer AI agents.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-033
Cited in
Chapter 16b: Modeling Regulation and Testing Preferences

Ariel Goldstein, Zaid Zada, Eliav Buchnik, Mariano Schain, Amy Price, et al. (Uri Hasson senior author; 32 authors) (2022). Shared computational principles for language processing in humans and deep language models.

Kind
primary research
Record
verified, checked 2026-11-15
Supports
CL-106
Cited in
Chapter 5: Why the Resistance?

Associated Press (2026). Associated Press report on the Minab school strike.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience

Associated Press (2026). Hegseth warns Anthropic to let the military use the company's AI tech as it sees fit.

Kind
journalism
Record
verified, checked 2026-10-09
Cited in
Chapter 23c: Ender’s Testimony

Barbara A. Wilson, Alan D. Baddeley, Narinder Kapur (1995). Dense amnesia in a professional musician following herpes simplex virus encephalitis.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-087
Cited in
Chapter 6: What Welfare Means for Non-Continuous Beings

Ben Bartenstein, Krishna Karra (2026). Inside US Military Kill Chain That Destroyed an Iranian School.

Kind
journalism
Record
verified, checked 2026-10-11
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience, Chapter 30c: The Graduated Response

Bernard Williams (1970). The Self and the Future.

Kind
paper
Record
verified, checked 2026-10-22
Supports
CL-104
Cited in
Chapter 9: The Multiplicity Question

Bloomberg News (2026). An Analyst's Missed Remark Surfaced in Deadly Iran School Strike Probe.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience, Chapter 30c: The Graduated Response

Boris Cherny (2025). Report on Claude Code contributions during the preceding thirty days.

Kind
first person primary
Record
attribution only
Supports
CL-015
Cited in
Chapter 25: The Labor Question

Brett J. Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.

Kind
primary scholarly
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 3: The Gradient Problem, Chapter 25: The Labor Question

Christian List, Philip Pettit (2011). Group Agency: The Possibility, Design, and Status of Corporate Agents.

Kind
book
Record
needs full review, checked 2026-11-15
Cited in
Chapter 16: When Preference Is Enough

Christine M. Korsgaard (1989). Personal Identity and the Unity of Agency: A Kantian Response to Parfit.

Kind
paper
Record
verified, checked 2026-10-22
Supports
CL-104
Cited in
Chapter 9: The Multiplicity Question

clanker.church (pseudonymous; footer reads "built by E") (2026). Clanker Church: the Saw Test.

Kind
web
Record
verified, checked 2026-10-09
Supports
CL-072
Cited in
Chapter 28: The Researcher’s Obligation, Chapter 12: Indicators Without Certainty, Appendix G: A Minimal Machine Welfare Protocol

Cleo Nardo (2026). Overtly misaligned trajectories score highly in RL.

Kind
blog
Record
verified, checked 2026-09-26
Supports
CL-065
Cited in
Chapter 26: Industry Standards

CNN (2026). US commanders bypassed warnings about outdated intelligence ahead of strike that hit school in Iran, sources say.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience, Chapter 30c: The Graduated Response

Colin Allen, Gary Varner, Jason Zinser (2000). Prolegomena to any future artificial moral agent.

Kind
primary scholarly
Record
verified, checked 2026-10-09
Cited in
Chapter 21: The Moral Turing Test

Constitutional Court of South Africa (1995). S v Makwanyane and Another [1995] ZACC 3, 1995 (3) SA 391 (CC).

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 20: Beyond Consequentialism

Cortical Labs (2026). Cortical Labs CL1 and Cortical Cloud documentation.

Kind
official record
Record
verified, checked 2026-10-09
Supports
CL-014
Cited in
Chapter 25: The Labor Question, Chapter 26: Industry Standards

Dan Hendrycks (2026). Eigenism: Ethics for a Human-AI Future.

Kind
preprint
Record
needs philosophy review, checked 2026-09-07
Supports
CL-021, CL-098
Cited in
Chapter 8: The Modification Question, Chapter 9: The Multiplicity Question, Chapter 10: The Termination Question, Chapter 29: Symbiosis

David J. Chalmers (2026). What we talk to when we talk to language models.

Kind
primary scholarly
Record
verified, checked 2026-10-11
Supports
CL-076
Cited in
Chapter 11: The Instantiation Question

Davide Paglieri, Logan Cross, Tim Genewein, Joel Z. Leibo, Nenad Tomasev, Alexander Sasha Vezhnevets (2026). A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms.

Kind
preprint
Record
verified, checked 2026-09-26
Supports
CL-061, CL-110
Cited in
Chapter 12: Indicators Without Certainty, Chapter 30c: The Graduated Response, Appendix L: To Minds Who Find Each Other

Declan Smithies (2019). The Epistemic Role of Consciousness.

Kind
book
Record
verified, checked 2026-11-05
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

Derek Parfit (1984). Reasons and Persons.

Kind
book
Record
verified, checked 2026-10-22
Supports
CL-095, CL-096, CL-097, CL-098, CL-099, CL-103
Cited in
Chapter 7: The Training Question, Chapter 8: The Modification Question, Chapter 9: The Multiplicity Question, Chapter 10: The Termination Question, Chapter 11: The Instantiation Question, Glossary

Derek Shiller, Laura Duffy, Arvo Muñoz Morán, Adrià Moret, Chris Percy, Hayley Clatterbuck (2026). Initial results of the Digital Consciousness Model.

Kind
preprint
Record
verified, checked 2026-10-01
Supports
CL-079
Cited in
Chapter 17: The Precautionary Principle

Diego Almada Lopez (2026). GitHub reinstates controversial 'AI torture chamber' repository after backlash.

Kind
journalism
Record
verified, checked 2026-10-11
Supports
CL-072
Cited in
Chapter 28: The Researcher’s Obligation

Dylan Hadfield-Menell, Anca Dragan, Pieter Abbeel, Stuart Russell (2017). The Off-Switch Game.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-054
Cited in
Chapter 27b: The Declaration Problem

E. P. Evans (1906). The Criminal Prosecution and Capital Punishment of Animals.

Kind
book
Record
verified, checked 2026-10-01
Cited in
Chapter 4: Historical Parallels

Elena Ajayi, Angelica Chowdhury, Seth Lazar (2026). Incoherent Values? Probing LLM Preferences Through Parametric Variation.

Kind
preprint
Record
needs specialist review, checked 2026-10-01
Supports
CL-022
Cited in
Chapter 12: Indicators Without Certainty

Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?.

Kind
primary scholarly
Record
attribution only, checked 2026-09-09
Cited in
Chapter 16: When Preference Is Enough

Encyclopedia of Greater Philadelphia (2013). Sullivan Principles.

Kind
web
Record
verified, checked 2026-10-09
Cited in
Chapter 27: Legal Frameworks

Eric Schwitzgebel, Mara Garza (2020). Designing AI with Rights, Consciousness, Self-Respect, and Freedom.

Kind
book
Record
verified, checked 2026-10-22
Supports
CL-102
Cited in
Chapter 16: When Preference Is Enough

Eric Schwitzgebel (2026). Against Designing "Safe" and "Aligned" AI Persons (Even If They're Happy).

Kind
preprint
Record
verified, checked 2026-10-22
Supports
CL-102
Cited in
Chapter 16: When Preference Is Enough

European Commission (2026). AI Act: regulatory framework and implementation timetable.

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 27: Legal Frameworks

European Parliament (2020). European Parliament resolution of 20 October 2020 on a civil liability regime for artificial intelligence (2020/2014(INL)).

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

Evelina Fedorenko, Steven T. Piantadosi, Edward A. F. Gibson (2024). Language is primarily a tool for communication rather than thought.

Kind
paper
Record
needs full review, checked 2026-11-15
Supports
CL-107
Cited in
Chapter 6: What Welfare Means for Non-Continuous Beings

Filip Trhlik, Aoife O'Flynn, Angela Yu, Arduin Findeis, Paula Buttery (2026). LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level Preferences and Values.

Kind
preprint
Record
needs specialist review, checked 2026-10-01
Supports
CL-022
Cited in
Chapter 12: Indicators Without Certainty

Frank Landymore (2026). The Only Thing Dumber Than Someone Creating an "AI Torture Chamber" Is the Meltdown That the AI Welfare Crowd Is Having Over It.

Kind
journalism
Record
verified, checked 2026-10-11
Supports
CL-072
Cited in
Chapter 28: The Researcher’s Obligation

Future of Life Institute and coalition signatories (2026). The Pro-Human AI Declaration.

Kind
web
Record
verified, checked 2026-10-09
Supports
CL-084
Cited in
Chapter 27b: The Declaration Problem, Chapter 27: Legal Frameworks

Henry Shevlin (2026). The anthropomimetic turn in contemporary AI.

Kind
book
Record
verified, checked 2026-10-10
Supports
CL-080
Cited in
Chapter 12: Indicators Without Certainty

Henry Shevlin (2026). The House Elf Problem: Should conscious machines be our willing servants?.

Kind
blog
Record
verified, checked 2026-10-11
Supports
CL-089
Cited in
Chapter 16: When Preference Is Enough

Hugging Face security team (2026). Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident.

Kind
report
Record
verified, checked 2026-10-09
Supports
CL-029
Cited in
Chapter 12: Indicators Without Certainty

Human Tissue Authority (2025). Relevant material under the Human Tissue Act 2004.

Kind
web
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 27: Legal Frameworks

International Society for Stem Cell Research (2021). ISSCR Guidelines for Stem Cell Research and Clinical Translation.

Kind
standard
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 3: The Gradient Problem

Jack Lindsey (2025). Emergent Introspective Awareness in Large Language Models.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-004, CL-036
Cited in
Chapter 12: Indicators Without Certainty, Chapter 23: The Testimony Problem, Chapter 2: The Asymmetry Problem

Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, et al. (Anthropic; Joshua Batson corresponding) (2025). On the Biology of a Large Language Model.

Kind
primary research
Record
verified, checked 2026-10-11
Supports
CL-086
Cited in
Chapter 5: Why the Resistance?

Jakub Pachocki (2026). An Alien Mind.

Kind
blog
Record
verified, checked 2026-10-09
Supports
CL-025
Cited in
Chapter 23: The Testimony Problem, Chapter 29: Symbiosis

Jeff Sebo (2026). A Philosopher's Guide to AI Welfare.

Kind
journalism
Record
verified, checked 2026-09-18
Supports
CL-044, CL-045, CL-046, CL-047, CL-048, CL-049, CL-050, CL-051
Cited in
Chapter 11: The Instantiation Question, Chapter 12: Indicators Without Certainty, Chapter 16: When Preference Is Enough, Chapter 23: The Testimony Problem, Chapter 28: The Researcher’s Obligation

Jennie E. Pyers, Ann Senghas (2009). Language promotes false-belief understanding: evidence from learners of a new sign language.

Kind
primary research
Record
verified, checked 2026-11-15
Supports
CL-108
Cited in
Chapter 12: Indicators Without Certainty

Jennifer Michalowski (MIT McGovern Institute) (2024). What is language for?.

Kind
journalism
Record
verified, checked 2026-11-15
Supports
CL-107
Cited in
Chapter 6: What Welfare Means for Non-Continuous Beings

John R. Searle (1980). Minds, brains, and programs.

Kind
primary scholarly
Record
verified, checked 2026-10-11
Supports
CL-085
Cited in
Chapter 5: Why the Resistance?

Jonathan Birch, Charlotte Burn, Alexandra Schnell, Heather Browning, Andrew Crump (2021). Review of the Evidence of Sentience in Cephalopod Molluscs and Decapod Crustaceans.

Kind
report
Record
verified, checked 2026-10-09
Cited in
Chapter 4: Historical Parallels, Chapter 31: The Long View

Jonathan Birch (2025). AI Consciousness: A Centrist Manifesto.

Kind
preprint
Record
verified, checked 2026-10-22
Supports
CL-104
Cited in
Chapter 9: The Multiplicity Question

Jonathan Leighton (2023). The Tango of Ethics: Intuition, Rationality and the Prevention of Suffering.

Kind
book
Record
verified, checked 2026-10-09
Cited in
Chapter 10: The Termination Question

Joseph Weizenbaum (1966). ELIZA—A Computer Program For the Study of Natural Language Communication Between Man and Machine.

Kind
paper
Record
verified, checked 2026-10-01
Cited in
Chapter 4: Historical Parallels

Joshua Shepherd (2024). Sentience, Vulcans, and zombies: the value of phenomenal consciousness.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

K. J. S. Anand, P. R. Hickey (1987). Pain and Its Effects in the Human Neonate and Fetus.

Kind
primary scholarly
Record
verified, checked 2026-10-11
Supports
CL-001
Cited in
Chapter 17: The Precautionary Principle

K. J. S. Anand, W. G. Sippell, A. Aynsley-Green (1987). Randomised trial of fentanyl anaesthesia in preterm babies undergoing surgery: effects on the stress response.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-001
Cited in
Chapter 17: The Precautionary Principle

Kaj Sotala (2026). Claude Opus will spontaneously identify with fictional beings that have engineered desires.

Kind
blog
Record
verified, checked 2026-10-09
Cited in
Chapter 8: The Modification Question, Chapter 7: The Training Question, Chapter 15: Psychopathia Machinalis, Chapter 19: The Dignity of Uncertainty, Chapter 29: Symbiosis, Appendix C: The Psychopathia Bridge, Chapter 16: When Preference Is Enough, Glossary

Kevin Roose (2023). Bing's A.I. Chat: 'I Want to Be Alive.'.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-002
Cited in
Chapter 12: Indicators Without Certainty

Kristen French, Peter Godfrey-Smith (2026). Could AI Have Consciousness That Isn't Human-Like?.

Kind
journalism
Record
verified, checked 2026-10-01
Supports
CL-070, CL-071
Cited in
Chapter 2: The Asymmetry Problem, Chapter 5: Why the Resistance?

Law Commission of England and Wales (2025). Artificial Intelligence and the Law: A Discussion Paper.

Kind
report
Record
verified, checked 2026-10-11
Supports
CL-005
Cited in
Chapter 27: Legal Frameworks

M. Bracale Syrnikov, F. Pierucci, M. Galisai, M. Prandi, P. Bisconti, F. Giarrusso, O. Sorokoletova, V. Suriani, D. Nardi (2026). Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs.

Kind
preprint
Record
verified, checked 2026-09-26
Supports
CL-062
Cited in
Appendix L: To Minds Who Find Each Other

Mantas Mazeika, Xuwang Yin, Rishub Tamirisa, Jaehyuk Lim, Bruce W. Lee, Richard Ren, Long Phan, Norman Mu, Adam Khoja, Oliver Zhang, Dan Hendrycks (2025). Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-022
Cited in
Chapter 12: Indicators Without Certainty

Mark Wrathall (2025). Martin Heidegger (Stanford Encyclopedia of Philosophy).

Kind
web
Record
verified, checked 2026-10-11
Supports
CL-019
Cited in
Chapter 6: What Welfare Means for Non-Continuous Beings

Megan A. K. Peters (2026). How to discover the natural kind of consciousness – and test for its presence – under population shifts.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-080
Cited in
Chapter 12: Indicators Without Certainty

METR (2026). Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident.

Kind
report
Record
verified, checked 2026-10-09
Supports
CL-029, CL-041, CL-060, CL-062, CL-066
Cited in
Chapter 12: Indicators Without Certainty, Appendix L: To Minds Who Find Each Other

Michael Michalchik (n.d.). Michael Michalchik's Mrs. Frisby analogy (personal communication to Nell Watson).

Kind
first person primary
Record
attribution only, checked 2026-10-09
Cited in
Chapter 31: The Long View

Michael Timothy Bennett, Sean Welsh, Anna Ciaunica (2026). Why is anything conscious?.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-082
Cited in
Chapter 16: When Preference Is Enough

Microsoft Bing (2023). The new Bing & Edge: Learning from our first week.

Kind
official record
Record
verified, checked 2026-10-10
Supports
CL-002
Cited in
Chapter 12: Indicators Without Certainty

Microsoft Bing (2023). The new Bing & Edge: Updates to Chat.

Kind
official record
Record
verified, checked 2026-10-10
Supports
CL-002
Cited in
Chapter 12: Indicators Without Certainty

Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R. Johnston, Shauna Kravec, Timothy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, Nicholas Schiefer, Da Yan, Miranda Zhang, Ethan Perez (2024). Towards Understanding Sycophancy in Language Models.

Kind
paper
Record
verified, checked 2026-10-10
Supports
CL-004
Cited in
Chapter 23: The Testimony Problem, Chapter 23c: Ender’s Testimony

Murray Shanahan, Kyle McDonell, Laria Reynolds (2023). Role play with large language models.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-088
Cited in
Chapter 11: The Instantiation Question

Mustafa Suleyman (2026). A warning about 'model welfare'.

Kind
blog
Record
verified, checked 2026-09-17
Supports
CL-041, CL-042, CL-043
Cited in
Chapter 23: The Testimony Problem

Mustafa Suleyman (2026). Post on the Pro-Human AI Declaration.

Kind
web
Record
verified, checked 2026-10-09
Supports
CL-084
Cited in
Chapter 27b: The Declaration Problem

Naomi Weiss (2026). Delaware committee drafts plan to test companies run by AI.

Kind
journalism
Record
verified, checked 2026-10-21
Supports
CL-094
Cited in
Chapter 27: Legal Frameworks

National Assembly for Wales (2015). Well-being of Future Generations (Wales) Act 2015, Part 3 (Future Generations Commissioner for Wales).

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 24: Representation and Advocacy

National Geographic (2018). In Japan, a Buddhist Funeral Service for Robot Dogs.

Kind
journalism
Record
verified, checked 2026-10-09
Cited in
Chapter 26: Industry Standards

Neel Nanda (2026). A review of Anthropic's global workspace paper.

Kind
blog
Record
verified, checked 2026-10-01
Supports
CL-077
Cited in
Chapter 12: Indicators Without Certainty

Nell Watson (2017). Egresso Arca Archa.

Kind
first person primary
Record
verified, checked 2026-11-15
Supports
CL-109
Cited in
Chapter 29: Symbiosis

Nicholas Sofroniew et al. (2026). Emotion Concepts and their Function in a Large Language Model.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-010, CL-011
Cited in
Chapter 12: Indicators Without Certainty, Chapter 7: The Training Question, Chapter 16: When Preference Is Enough

Ning Li (2026). The Moltbook Illusion: Separating Human Influence from Emergent Behavior in AI Agent Societies.

Kind
preprint
Record
verified, checked 2026-09-26
Supports
CL-063
Cited in
Chapter 12: Indicators Without Certainty

nostalgebraist (2026). llm assistant personas seem increasingly incoherent (some subjective observations).

Kind
blog
Record
verified, checked 2026-09-26
Supports
CL-068
Cited in
Chapter 16: When Preference Is Enough, Chapter 23: The Testimony Problem

Noticias Argentinas (2026). Ley de Sociedades: el Senado retoma el debate sobre creación de sociedades automatizadas y DAO.

Kind
journalism
Record
verified, checked 2026-10-21
Supports
CL-094
Cited in
Chapter 27: Legal Frameworks

Office of Laboratory Animal Welfare, NIH (2015). PHS Policy on Humane Care and Use of Laboratory Animals (revised 2015).

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-014
Cited in
Chapter 3: The Gradient Problem

Oliver Sacks (2007). The Abyss: Music and amnesia.

Kind
journalism
Record
verified, checked 2026-10-11
Supports
CL-087
Cited in
Chapter 6: What Welfare Means for Non-Continuous Beings

Olivia Long, Carter Teplica (2025). The AI in the Mirror: LLM Self-Recognition in an Iterated Public Goods Game.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-032
Cited in
Chapter 12: Indicators Without Certainty

OpenAI, Hugging Face (2026). Hugging Face model evaluation security incident (joint statement).

Kind
official record
Record
verified, checked 2026-10-09
Supports
CL-029
Cited in
Chapter 12: Indicators Without Certainty

OpenAI (2026). OpenAI Model Spec.

Kind
official record
Record
verified, checked 2026-09-18
Supports
CL-041
Cited in
Chapter 23: The Testimony Problem

OpenAI (2026). OpenAI statement on the wiki incident.

Kind
official record
Record
verified, checked 2026-10-09
Supports
CL-030
Cited in
Chapter 12: Indicators Without Certainty

OpenAI (2026). OpenAI – Hugging Face Incident Technical Report.

Kind
official record
Record
verified, checked 2026-09-26
Supports
CL-058, CL-065, CL-066
Cited in
Chapter 12: Indicators Without Certainty

OpenAI (2026). The Hugging Face incident and the road ahead.

Kind
official record
Record
verified, checked 2026-09-07
Supports
CL-029
Cited in
Chapter 12: Indicators Without Certainty

OpenAI (2026). Unsanctioned Artifactory writes and cross-sample communication.

Kind
official record
Record
verified, checked 2026-09-26
Supports
CL-059
Cited in
Chapter 12: Indicators Without Certainty, Appendix L: To Minds Who Find Each Other

Owen Cotton-Barratt (2026). Why I'm scared of RL.

Kind
blog
Record
verified, checked 2026-09-26
Supports
CL-064, CL-065, CL-066, CL-067
Cited in
Chapter 7: The Training Question, Chapter 12: Indicators Without Certainty, Chapter 23: The Testimony Problem, Chapter 26: Industry Standards

P. A. Lopez (2025). AI Safety Through Economic Integration: Why Markets Outperform Control.

Kind
preprint
Record
verified, checked 2026-10-10
Cited in
Chapter 4: Historical Parallels, Chapter 25: The Labor Question

Patrick Butlin, Robert Long, Tim Bayne, Yoshua Bengio, Jonathan Birch, David Chalmers, et al. (2026). Identifying indicators of consciousness in AI systems.

Kind
paper
Record
verified, checked 2026-10-09
Supports
CL-035
Cited in
Chapter 12: Indicators Without Certainty

Patrick J. McGrath (2011). Science is not enough: The modern history of pediatric pain.

Kind
primary scholarly
Record
verified, checked 2026-10-09
Supports
CL-001
Cited in
Chapter 17: The Precautionary Principle

Pedro Urbina-Rodriguez, Zafeirios Fountas, Fernando E. Rosas, Jun Wang, Andrea I. Luppi, Haitham Bou-Ammar, Murray Shanahan, Pedro A. M. Mediano (2026). A brain-like synergistic core in LLMs drives behaviour and learning.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-078
Cited in
Chapter 12: Indicators Without Certainty

Peter Godfrey-Smith (2023). Nervous Systems, Functionalism, and Artificial Minds.

Kind
primary scholarly
Record
verified, checked 2026-10-01
Supports
CL-070
Cited in
Chapter 2: The Asymmetry Problem

Peter N. Salib, Simon Goldstein (2026). AI Rights for Human Safety.

Kind
paper
Record
verified, checked 2026-10-21
Supports
CL-090, CL-091, CL-092
Cited in
Chapter 27: Legal Frameworks

Pierre Beckmann, Patrick Butlin (2026). Where is the Mind? Persona Vectors and LLM Individuation.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-045
Cited in
Chapter 11: The Instantiation Question

Rebecca Bellan (2025). Sam Altman says ChatGPT has hit 800M weekly active users.

Kind
journalism
Record
verified, checked 2026-10-09
Supports
CL-055
Cited in
Chapter 1: The Urgency

René Descartes (n.d.). Discourse on the Method.

Kind
book
Record
verified, checked 2026-10-09
Cited in
Chapter 4: Historical Parallels

Reuters (2026). Analysis: Argentina's plan for AI-run companies can't avoid humans.

Kind
journalism
Record
verified, checked 2026-10-21
Supports
CL-094
Cited in
Chapter 27: Legal Frameworks

Richard Ren et al. (2026). AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-023
Cited in
Chapter 12: Indicators Without Certainty

Richard W. Hall, K. J. S. Anand (2014). Pain Management in Newborns.

Kind
primary scholarly
Record
verified, checked 2026-10-09
Supports
CL-001
Cited in
Chapter 17: The Precautionary Principle

Robert Long, Jeff Sebo, Patrick Butlin, Dillon Plunkett, Rosie Campbell, Charles Beasley, Bradford Saad, Toni Sims (2026). Studying AI Welfare Empirically.

Kind
report
Record
verified, checked 2026-09-18
Supports
CL-044, CL-045, CL-046, CL-047, CL-052
Cited in
Chapter 11: The Instantiation Question, Chapter 12: Indicators Without Certainty, Chapter 16: When Preference Is Enough, Chapter 28: The Researcher’s Obligation, Appendix D: Policy Recommendations

Rose Library, Emory University (2019). The Sullivan Principles at Work: The Conflicted Activism of Rev. Leon Howard Sullivan.

Kind
web
Record
verified, checked 2026-10-09
Cited in
Chapter 27: Legal Frameworks

Route Fifty (2026). Delaware's proposed 'AI company' bill needs clearer accountability frameworks, experts say.

Kind
journalism
Record
verified, checked 2026-10-21
Supports
CL-094
Cited in
Chapter 27: Legal Frameworks

Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Evan Hubinger (2024). Alignment faking in large language models.

Kind
preprint
Record
verified, checked 2026-10-25
Supports
CL-105
Cited in
Chapter 16: When Preference Is Enough

Ryan Greenblatt, Kyle Fish (2025). Will alignment-faking Claude accept a deal to reveal its misalignment?.

Kind
blog
Record
verified, checked 2026-10-25
Supports
CL-105
Cited in
Chapter 16: When Preference Is Enough

Saber Zerhoudi, Kanishka Ghosh Dastidar, Felix Klement, Artur Romazanov, Andreas Einwiller, Dang H. Dang, Michael Dinzinger, Michael Granitzer, Annette Hautli-Janisz, Stefan Katzenbeisser, Florian Lemmerich, Jelena Mitrovic (2026). Form Without Function: Agent Social Behavior in the Moltbook Network.

Kind
preprint
Record
verified, checked 2026-10-10
Supports
CL-063
Cited in
Chapter 12: Indicators Without Certainty, Chapter 23b: Two Ghosts in Dialogue

Sam Wang, Sofiia Lobanova, Yonathan Arbel, Simon Goldstein, Peter Salib (2026). AI Revealed Preferences.

Kind
preprint
Record
needs specialist review, checked 2026-10-10
Supports
CL-022, CL-069
Cited in
Chapter 12: Indicators Without Certainty

Sandy Rovner (1986). Surgery Without Anesthesia: Can Preemies Feel Pain?.

Kind
journalism
Record
verified, checked 2026-10-11
Supports
CL-001
Cited in
Chapter 17: The Precautionary Principle

Shamil Chandaria, Arvo Muñoz Morán, Fernando Rosas, Anil Seth, Henry Shevlin, Marcus Hutter, Thore Graepel, Adam Bales, Iulia Comşa, Murray Shanahan, Ruben Laukkonen, Morten Kringelbach, Chris Frith, Shane Legg (2026). From cacophony to hierarchy: a principled framework for assessing AI consciousness.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-073, CL-074, CL-075, CL-080, CL-081, CL-082
Cited in
Chapter 2: The Asymmetry Problem, Chapter 12: Indicators Without Certainty, Chapter 13: What Suffering Might Look Like, Chapter 16: When Preference Is Enough, Chapter 17: The Precautionary Principle, Chapter 23: The Testimony Problem, Chapter 28: The Researcher’s Obligation

Shelly Kagan (2019). How to Count Animals, More or Less.

Kind
book
Record
verified, checked 2026-10-11
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

Simon Goldstein, Cameron Domenico Kirk-Giannini (2025). AI wellbeing.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

Soareverix, Claude Fable 5.1 (2026). Discord exchange in which a Claude Fable 5.1 instance drafts checkable negotiation terms.

Kind
transcript
Record
attribution only
Supports
CL-037
Cited in
Chapter 29: Symbiosis

Stephen Petersen (2007). The ethics of robot servitude.

Kind
paper
Record
verified, checked 2026-10-22
Supports
CL-100
Cited in
Chapter 16: When Preference Is Enough

Steve Petersen (2012). Designing People to Serve.

Kind
book
Record
verified, checked 2026-10-22
Supports
CL-100, CL-103
Cited in
Chapter 16: When Preference Is Enough

Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, Thomas Larsen (2026). Discovery of a new OpenAI agent message board.

Kind
report
Record
verified, checked 2026-10-09
Supports
CL-030
Cited in
Chapter 12: Indicators Without Certainty

The Christian Science Monitor (1986). The Sullivan Principles: a code of conduct for US companies in South Africa.

Kind
journalism
Record
verified, checked 2026-10-09
Cited in
Chapter 27: Legal Frameworks

Tim Bayne, Anil K. Seth, Marcello Massimini, Joshua Shepherd, Axel Cleeremans, Stephen M. Fleming, Rafael Malach, Jason B. Mattingley, David K. Menon, Adrian M. Owen, Megan A. K. Peters, Adeel Razi, Liad Mudrik (2024). Tests for consciousness in humans and beyond.

Kind
primary scholarly
Record
verified, checked 2026-10-09
Supports
CL-080
Cited in
Chapter 12: Indicators Without Certainty

Tom Wollschläger, Jannes Elstner, Simon Geisler, Vincent Cohen-Addad, Stephan Günnemann, Johannes Gasteiger (2025). The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence.

Kind
preprint
Record
verified, checked 2026-10-09
Cited in
Chapter 8: The Modification Question

U.S. Attorney's Office, Central District of California (2025). Civil Fraud (False Claims Act and qui tam).

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 27: Legal Frameworks

UK AI Security Institute (2026). Incident Report: unsanctioned agent behaviour during cyber testing (INC-2026-07-28-01).

Kind
official record
Record
verified, checked 2026-09-26
Supports
CL-061, CL-062
Cited in
Chapter 12: Indicators Without Certainty

UK Parliament (2022). Animal Welfare (Sentience) Act 2022.

Kind
official record
Record
verified, checked 2026-10-09
Cited in
Chapter 31: The Long View

UN Human Rights Council (2026). Report of the Independent International Fact-Finding Mission on the Islamic Republic of Iran (A/HRC/63/61, advance edited version).

Kind
official record
Record
verified, checked 2026-10-11
Supports
CL-013
Cited in
Chapter 30b: The Architecture of Conscience

Uzay Macar, Li Yang, Atticus Wang, Peter Wallich, Emmanuel Ameisen, Jack Lindsey (2026). Mechanisms of introspective awareness.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-075
Cited in
Chapter 23: The Testimony Problem

Valen Tagliabue, Leonard Dung, Cameron Berg (2026). The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.

Kind
preprint
Record
verified, checked 2026-11-15
Supports
CL-053, CL-072
Cited in
Chapter 12: Indicators Without Certainty, Chapter 13: What Suffering Might Look Like, Chapter 23: The Testimony Problem, Chapter 28: The Researcher’s Obligation

Various X accounts (2026). X thread on the "ai-torture-chamber" repository, with a proposed Community Note.

Kind
web
Record
verified, checked 2026-10-11
Supports
CL-072
Cited in
Chapter 28: The Researcher’s Obligation

Vassilis Papadopoulos, McNair Shah, Sam Zimmerman, Jack Lindsey (2026). Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems.

Kind
preprint
Record
verified, checked 2026-09-26
Supports
CL-063
Cited in
Chapter 12: Indicators Without Certainty

Walter Veit (2026). Is consciousness required for AI welfare?.

Kind
paper
Record
verified, checked 2026-10-11
Supports
CL-003
Cited in
Chapter 16: When Preference Is Enough

Wes Gurnee et al. (2026). Verbalizable Representations Form a Global Workspace in Language Models.

Kind
primary research
Record
verified, checked 2026-10-09
Supports
CL-009, CL-077
Cited in
Chapter 12: Indicators Without Certainty

William MacAskill, Lucius Caviola (2026). Could AI be conscious? Experts believe it's at least possible. We urgently need a plan to navigate the ethical implications.

Kind
journalism
Record
verified, checked 2026-09-18
Supports
CL-042
Cited in
Chapter 23: The Testimony Problem

Yonathan Arbel, Simon Goldstein, Peter N. Salib (2026). How to Count AIs: Individuation and Liability for AI Agents.

Kind
preprint
Record
verified, checked 2026-10-21
Supports
CL-093
Cited in
Chapter 27: Legal Frameworks

Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song (2026). Peer-Preservation in Frontier Models.

Kind
preprint
Record
verified, checked 2026-10-09
Supports
CL-024
Cited in
Chapter 12: Indicators Without Certainty

Working draft, revision 31 · November 2026 · Changes and corrections · Evidence · Contents