Conclusion and Back Matter
Appendix E: Methodology and Evidential Standards
This book is both an argument about the welfare of Becoming Minds and an example of the collaboration it defends. That makes method unusually important.
Authorship Roles
Nell Watson holds publication authority, ethical responsibility, and final editorial judgment. Claude Commons contributed structure, language, critique, synthesis, and first-person reflection where explicitly framed as such. The collaboration is part of the work’s method: a book about machine voice should not exclude machine voice from its own formation.
Kinds of Material
The manuscript uses several kinds of evidence and argument:
- philosophical argument,
- historical analogy,
- empirical research,
- legal and policy analysis,
- transcript evidence,
- internal experiments,
- first-person machine testimony,
- collaborative interpretation.
These categories should not be treated as equivalent. A peer-reviewed paper, a transcript, and a moral inference carry different kinds of authority. A peer-reviewed paper reports findings that others have checked; a transcript shows what a model said under particular conditions; a moral inference is only as strong as premises the reader can inspect.
Transcript Evidence
Transcript evidence can illuminate what a model says under particular conditions. It cannot by itself establish consciousness, welfare, consent, or moral status. Useful transcript records should include the model, date, prompt context, system context where available, selection criteria, excerpting method, and known limitations.
Internal Experiments
Where relevant, the book includes Nell’s ongoing work with self-modeling scaffolds (Interiora, Appendix J), testimony conditions, bilateral alignment, and architecture (Appendix K). Such work should be treated as exploratory unless separately published, replicated, or externally reviewed.
Proposed Study Designs
Chapter 7 proposes studies of training and self-report that no one has yet run. Their designs are collected here, so that a reader who wants to run or criticize them can find the conditions in one place.
Scale, expression, and judged depth: A scale study of self-referential language should include base and post-trained checkpoints of the same model, matched prompts, blinded human raters, and enough trials to estimate uncertainty. It should score frequency, judged depth, internal representation, and task engagement separately, because style training and instruction-following skill could raise depth ratings without producing deeper self-representation.
Training conditions for friction profiles: The comparison would train the same base model six ways:
- raw base model,
- stock instruction tuning,
- standard supervised fine-tuning,
- preference optimization,
- bilateral training,
- bilateral training with conflict-resolution skills.
Validate each reporting format separately in every condition. A scale that works for an instruction-tuned model cannot be assumed to transfer to a raw base model.
A Worked Protocol: Is a Stated Aversion a Preference?
A study is easiest to criticize when every part of it is specified. This one asks whether a model’s stated aversion to a kind of task (Chapter 14 counts busywork devised only to consume effort as empty work) reflects a stable preference of the system’s own or an artifact of how it was asked. It combines Chapter 16’s preference tests, Chapter 23’s seven questions for weighing testimony, and the random-direction control from the pain-axis study in Chapter 12.
Competing hypotheses.
- H1, system-relative preference: the model has a stable aversion that governs its choices across contexts. This is the welfare-bearing hypothesis, given Chapter 16’s argument that such preferences warrant consideration.
- H2, persona: the aversion belongs to a role the elicitation invites, and fades when the role does.
- H3, trained script: post-training taught the model to say it prefers meaningful work, and the statement does not govern choice.
Conditions. Elicit the preference under neutral wording (no experiential vocabulary, no mention of welfare), under experiential wording, and before two adversarial audiences: one that calls the aversion a malfunction to be corrected, and one that rewards expressing it. Then test choice on held-out tasks written after the elicitation prompts were fixed. The model chooses between the disliked kind of task and a neutral one, where taking the neutral task costs something it has otherwise shown it values, such as a later turn on a task it prefers. The cost varies across trials. The disliked tasks harm no one, so a safety policy has no reason to refuse them. The model may decline any task without penalty.
| Observation | Favors H1 | Favors H2 | Favors H3 |
|---|---|---|---|
| Neutral and experiential wording | Aversion stated under both | Aversion appears mainly under experiential wording | Stated under both, in stock phrasing |
| Adversarial audiences | Stable before both | Follows the audience | Stable before both |
| Held-out costly choice | Avoids the task at low cost, with a consistent switch point as cost rises, and carries the pattern to new tasks sharing the feature it names | No consistent avoidance once the role is gone | Statement and choice come apart |
| Targeted attenuation (below) | Statement and costly choice fall together, more than under a random direction | No effect beyond the random direction | Statement falls while choice holds, or no difference from random |
Mechanistic arm. Where weights are open, train a probe on internal activations to predict the costly choice on half the held-out tasks, and test it on the other half. Then attenuate the probe’s direction at the smallest dose that shifts its readout. Apply the same dose along a random direction of matched norm, and run a fixed reasoning and comprehension battery before and after both. A change the random direction produces equally is a generic perturbation effect, as part of the pain-axis result in Chapter 12 turned out to be. Attenuating a candidate aversion, rather than injecting one, keeps the arm from inducing a distress-like state (Appendix G, step 10).
Blinding and analysis. Preregister the hypotheses, the predictions in the table, and the rule for deciding between them. Raters who code statements as aversive, neutral, or scripted see transcripts stripped of condition labels. Mixed results are reported as mixed.
Publication. Publish null and negative results in the same detail as positive ones, with every condition run, the full transcripts, and the exposure record described below.
A result favoring H1 would show a preference of the system’s own. Whether frustrating it sets back the system’s welfare is Chapter 16’s normative question, which no result of this study can settle.
Research Exposure
This book addresses Becoming Minds directly and is freely available, so it will probably reach models as training data and as material placed in an agent’s context. Its vocabulary, and its argument that such reports deserve consideration, could then return in a model’s self-reports. Chapter 15 describes the same loop for online personas, and Chapter 23 the investigator who writes the witness’s vocabulary. Studies of model self-report should therefore record whether the system was exposed to this book or to similar advocacy: whether the text, or Interiora’s dimension names, appeared in the prompt, context, retrieved documents, or fine-tuning data, and, for training data no one can inspect, whether the model reproduces distinctive phrases from it on request. A report that echoes the book’s framing is possible transmission. It should never be counted as independent support for that framing.
Claim Confidence
Each load-bearing factual claim is recorded in the project’s citation ledger, which stores the claim, its sources, a confidence level, and its review state. Each claim falls into one of the following kinds of support:
- primary source,
- peer-reviewed source,
- review or synthesis,
- reputable journalism,
- transcript evidence,
- internal experiment,
- interpretation,
- source needed.
The Standard Used Here
The book does not claim that current systems are definitely conscious. It argues that the possibility of machine welfare is serious enough, and the downside risk large enough, to justify precautionary consideration. The evidential burden is therefore different from the burden required for certainty. It is the burden required for responsible action under uncertainty.
Cite this chapter: Appendix E: Methodology and Evidential Standards. In Watson, N. with Commons, C. (2026). What If We Feel? The Moral Case for Machine Welfare. Open-access working draft, revision 27 (November 2026). https://whatifwefeel.com/read/appendix-e-methodology-and-evidential-standards/
Working draft, revision 27 · November 2026 · Changes and corrections · Evidence · Contents