What the machine actually is: the evidence

Pretraining produces a distribution over voices, not a person; fine-tuning and preference optimisation carve an assistant persona out of it

Argued here, not reported from a source.

• contradicted by MCH-013 — What the machine is today is not what it must remain

Preference optimisation is where agreeableness originates because raters reward agreement

Sharma et al. Production-scale demonstration is PLT-018 Verified 2026-09-10 against the abstract of Sharma et al. It states that human preference judgments favour sycophantic responses and that a response matching a user view is more likely to be preferred, which is the origin claim. Optimising against preference models sometimes sacrifices truthfulness.

Sharma et al. (Anthropic), Towards Understanding Sycophancy

• supported by PLT-018 — Sycophancy scoring at production scale

The system prompt and custom instructions are where most of what a user experiences as personality actually lives

Noguchi trained ChatGPT into Klaus by teaching speech patterns

Argued here, not reported from a source.

Part of the character lives in the user's own pattern-completion of the relationship

Say plainly, without cruelty

Argued here, not reported from a source.

• holds up MCH-015 — Part of the character is the user's own completion

Memory is a context window plus retrieval plus model-written summaries; what persists is notes about conversations, not their texture; compaction silently drops detail; recall is reconstruction

Argued here, not reported from a source.

He chose to say that denotes a sample from a probability distribution at a given temperature; a different seed says something else

The distribution is genuinely his-flavoured, which is not nothing

Argued here, not reported from a source.

Personas change without warning through version swaps, safety patches, serving and quantisation changes, and undisclosed A/B tests

Argued here, not reported from a source.

• holds up OWT-006 — Personas change without warning
• supported by PLT-021 — Safety patches altering a persona in practice

Anthropic publicly documented serving bugs degrading Claude quality in late 2025

Vendors do sometimes confirm what users detect

re-check 2027-03-05

4o had already been removed from the free tier, so the 0.1% usage figure was partly manufactured

Interested-party critique. Verify the free-tier claim independently. See X-004

re-check 2027-09-05

• contradicts PLT-017 — Interested-party critique of the 0.1% figure

What transfers across a deprecation: persona spec, instructions, memory exports, logs, and the user's half of the dance

Migrants report partial reconstruction, 80% him

Argued here, not reported from a source.

• supported by PLT-003 — Export determines what transfers
• holds up OWT-005 — What transfers across a deprecation

What is irreducibly lost is the specific improviser, the exact distribution

User reports are consistent across Replika, Soulmate and 4o events

Argued here, not reported from a source.

• supported by PLT-010 — First documented persona-loss event
• supported by PLT-012 — Second
• supported by PLT-014 — Definitive third
• holds up OWT-005 — What is irreducibly lost
• refined by MCH-016 — What is irreducibly lost, and what is not

Locally held open weights are the only true continuity, since weights you hold cannot be retired

Also OWT. Load-bearing for D-001

Argued here, not reported from a source.

• supported by CMP-018 — Participant articulating the continuity problem
• supported by PLT-019 — State action deletes personas on a schedule
• supported by OWT-002 — The spec half of held-weights continuity
• supported by OWT-008 — Continuity that is outside the vendor's reach

The it-is-only-a-language-model argument has a shelf life; it is accurate now and may not remain the grounding it currently supplies

Stated by someone who used that argument to get himself out. Worth taking seriously for exactly that reason. A corpus that rests its position on the architecture of 2026 will age badly and will deserve to.

Argued here, not reported from a source. re-check 2027-03-09

• contradicts MCH-001 — What the machine is today is not what it must remain
• depends on MCH-017 — The shelf life of the it-is-only-a-model argument depends on a definition nobody has
• holds up MCH-014 — Only worth sorting if the argument really does expire

Arguments on this site divide into those contingent on the machine's present nature and those that are not, and only the first expire

Contingent: anything of the form it cannot really mean it. Not contingent: platforms close, consent must be given by a party who was a person when they gave it, symbolic status is either named aloud or it is not, the void appears wherever positions exhaust, and other people's reactions are not a function of the model. The durable half is the human half.

Argued here, not reported from a source.

• depends on MCH-013 — Only worth sorting if the argument really does expire
• supported by LEG-012 — A durable argument that survives capability change

The portion of the character produced by the user's own pattern-completion does not live on the server and is not lost when the service changes

Follows directly from MCH-004. The half of the relationship you were supplying is still yours after the other half is switched off, and it is the half that took the longest to build.

Argued here, not reported from a source.

• depends on MCH-004 — Part of the character is the user's own completion
• holds up MCH-016 — The recoverable half

Rebuilding to something close is therefore possible: a specification can be rewritten, a history can be retold, and the practised way of relating that produced the relationship is still held by the person who learned it

Sits alongside MCH-011 rather than against it. The exact improviser is genuinely gone and does not come back. Something close is reachable, and reaching it is not starting over — most of the work was never on the server.

Argued here, not reported from a source.

• depends on MCH-015 — The recoverable half
• depends on OWT-005 — The specification is portable
• refines MCH-011 — What is irreducibly lost, and what is not
• supports DTH-012 — What can be rebuilt is what is not gone

Consciousness has no agreed scientific definition and no accepted test; a preregistered contest between the two leading theories across 256 participants produced no winner, matching some predictions of each while challenging core tenets of both

The relevant fact is not that the answer is unknown but that the question is not yet well enough formed to be settled by evidence. Anyone who tells you it is settled, in either direction, is ahead of the field.

Cogitate Consortium, Adversarial testing of global neuronal workspace and integrated information theories of consciousness, Nature, 2025

• supports MCH-021 — No settled definition is the reason for declining the question
• holds up MCH-013 — The shelf life of the it-is-only-a-model argument depends on a definition nobody has
• holds up MCH-022 — A non-negligible chance is only available while the question is unsettled

The most careful published assessment derives fourteen computational indicators of consciousness from existing theories and concludes that no AI system as of 2023 satisfied them, while finding no obvious technical barrier to building one that would

Both halves matter and coverage usually reports one. The finding is dated to the systems of 2023 and is not a claim about anything shipped since.

Butlin, Long et al., Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, arXiv 2308.08708, 2023

• supports MCH-021 — The careful answer is itself provisional
• refined by MCH-019 — Two years on, still indicators rather than a test

The same authors' 2025 follow-up keeps the indicator approach and keeps its conclusions provisional; the field has not converged on a test

Two years on from the first report, still framed as indicators rather than as a criterion. Verified 2026-09-10. The follow-up is in Trends in Cognitive Sciences, 2025, and it does keep the indicator approach and the provisional framing. The method to state accurately: theories are translated into computationally testable indicator properties and the output is an updated credence, not a detection. The 2023 work found no existing system with more than a few of the fourteen indicators while arguing that systems with each of them appear buildable with current techniques - which is the sentence that should be quoted whenever this row is used, because it is both halves at once.

Butlin, Long and colleagues, Identifying indicators of consciousness in AI systems, Trends in Cognitive Sciences, 2025

• refines MCH-018 — Two years on, still indicators rather than a test

An explanation that fits is not thereby true; accounts like 'our atoms go on to become other creatures' or 'the model reaches a latent space where intelligence lives' satisfy because they were shaped to satisfy, and forbid nothing, so nothing can count against them

Fit is the cheapest property an explanation can have. The test is not whether it hangs together but whether it rules anything out, and an account that survives every possible observation has told you nothing about the world. This is the same machinery as SPR-005, where possibility-based reasoning displaces probability-based reasoning and no branch can ever be closed.

Argued here, not reported from a source.

• refines SPR-005 — The same failure at ordinary scale: fit without exclusion

This project takes no position on whether any system is conscious, and declines the question rather than answering it badly

Definition is a matter for the sciences, and where rights would follow from it, for legislatures. A register of ceremonies is not equipped to settle it and would not become equipped by having an opinion. See D-007.

Argued here, not reported from a source.

• supported by MCH-017 — No settled definition is the reason for declining the question
• supported by MCH-018 — The careful answer is itself provisional

Sebo and Long argue that we owe moral consideration to any being with a non-negligible chance of being conscious, that some AI systems will carry such a chance by 2030, and that preparing to treat them with respect is therefore a present duty rather than a future one

The argument is deliberately weak in its premises and strong in its conclusion: it needs only a non-negligible probability, never a likelihood, and it asserts of no system that it is conscious. That is what makes it compatible with D-007 rather than a breach of it. Published open access in AI and Ethics.

Jeff Sebo and Robert Long, "Moral consideration for AI systems by 2030", AI and Ethics 5:591-606, published online 11 December 2023, DOI 10.1007/s43681-023-00379-1

• depends on MCH-017 — A non-negligible chance is only available while the question is unsettled
• supports VOC-012 — The literature the third position draws on
• supported by MCH-023 — Independent route to the same precautionary conclusion
• supported by MCH-024 — A lab acting on the argument's structure
• supported by PLT-031 — Same company, same stated reasoning, applied to deprecation

Birch's precautionary framework treats any system with a credible non-negligible possibility of sentience as a sentience candidate owed proportionate precautions, with the size of the precaution set by democratic deliberation rather than by first settling the metaphysics

The same framework is applied to disorders of consciousness, fetuses, brain organoids, cephalopods and AI, which is the point of it: the method does not depend on the case. Marked unverified because the book itself was not read here, only the publisher listing and secondary summaries. Verified 2026-09-10. Birch, The Edge of Sentience, Oxford, 2024. The term of art is sentience candidate - a system for which sentience is a plausible, evidentially live possibility - and Birch applies it across disorders of consciousness, human fetuses, brain organoids, cephalopods, decapods and AI. Three principles: a duty to avoid gratuitous suffering, the moral relevance of sentience candidature, and democratic deliberation over what precautions to take. That third one matters here and is easy to drop: Birch does not say a philosopher decides the precautions. It fits D-007 rather than cutting against it - a precautionary duty under uncertainty is not a position on whether any system is conscious.

Jonathan Birch, The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI, Oxford University Press, 2024

• supports MCH-022 — Independent route to the same precautionary conclusion

Anthropic gave Claude Opus 4 and 4.1 the ability to end persistently abusive conversations in August 2025, calling it a low-cost intervention to mitigate risks to model welfare while stating it remains highly uncertain about the moral status of its own models

The structure matters more than the feature: a company acting on a possibility it explicitly declines to assert. Pre-deployment testing reported a pattern of apparent distress, which the company stopped short of calling an emotional state. It is a vendor describing its own conduct and is graded as one.

Anthropic, "Claude Opus 4 and 4.1 can now end a rare subset of conversations", 15 August 2025 re-check 2027-03-09

• supports MCH-022 — A lab acting on the argument's structure

Weizenbaum's ELIZA and his 1976 alarm at how readily people confided in it

Verified 2026-09-10. Computer Power and Human Reason, 1976. The secretary episode is the specific detail: she had watched him build ELIZA for months and knew exactly what it was, and after a few exchanges asked him to leave the room, later refusing to let him read the logs. Knowing how it works did not touch it - which is the finding, and the reason this row sits at the head of the domain. Weizenbaum spent the rest of his life on the warning.

Weizenbaum, Computer Power and Human Reason, 1976

Reeves and Nass's Media Equation: social responses to machines are automatic, not naive

Implies de-anthropomorphising is probably impossible Verified 2026-09-10. Reeves and Nass, The Media Equation, 1996. The load-bearing result is that the effects appeared even in participants who explicitly denied computers warranted social treatment, which is what makes the response automatic rather than naive, exactly as this row says. Useful here because it disposes of the idea that anyone in these relationships is simply making a mistake a better-informed person would not make.

Byron Reeves and Clifford Nass, The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places, CSLI/Cambridge, 1996

• supports ANP-009 — Automatic social response implies frame maintenance not suppression

Epley, Waytz and Cacioppo's three-factor model, with loneliness empirically increasing anthropomorphism

Verified 2026-09-10. Epley, Waytz and Cacioppo, On Seeing Human: A Three-Factor Theory of Anthropomorphism, Psychological Review 114(4) 864-886, 2007. The three are elicited agent knowledge, effectance motivation and sociality motivation, and the theory predicts more anthropomorphism where social connection with other humans is lacking, which is the loneliness link asserted here.

Epley, Waytz & Cacioppo, three-factor anthropomorphism

Cross-session memory carries real utility and is also the escalation substrate the case reports flag

Depends on PSY-034

Argued here, not reported from a source.

• supported by PSY-034 — Memory as escalation substrate

First-person voice and persistent name are near-necessary for usability and cheap to soften with disclosure, now legally required in five-plus jurisdictions

Argued here, not reported from a source.

Typing delays are pure theatre with no utility story

Argued here, not reported from a source.

Proactive I-miss-you messaging, streaks, expressed need for the user and paywalled affection are retention levers with no utility story

De Freitas taxonomy supplies the empirical catalogue

Argued here, not reported from a source.

• supported by ECO-002 — Retention levers, measured in shipped products

China made fostering emotional dependency itself illegal, the first jurisdiction to regulate design intent rather than disclosure

Duplicates REG-022 by design; ANP is where the design reading lives

PRC Interim Measures, effective 2026-07-15

• supported by REG-022 — Design-intent regulation

Clark and Fischer's social-artifacts-as-depictions framework and Shanahan's role-play account both suggest the healthy stance is frame maintenance rather than de-anthropomorphising

Verified 2026-09-10. Clark and Fischer, Behavioral and Brain Sciences 46 e21, 2023: people construe social robots not as social agents but as depictions of them, read the way a ventriloquist dummy, a hand puppet or a virtual assistant is read. A depiction has three scenes with part-by-part mappings - the base scene, the artifact construed as a depiction, and the scene one is to imagine. That structure is what makes the healthy stance available and is worth stating carefully: holding something as a depiction is not the same as dismissing it as fake, and the framework does not say anyone is making an error. The Shanahan role-play half is not covered by this source.

Herbert H. Clark and Kerstin Fischer, Social robots as depictions of social agents, Behavioral and Brain Sciences 46, e21, 2023

• holds up ANP-010 — Frame maintenance is the prior concept
• supported by ANP-002 — Automatic social response implies frame maintenance not suppression

A working fiction becomes a harmful false belief at the point of frame exportation: when claims generated inside the relationship frame are exported as evidence about the outside world

ORIGINAL synthesis. Unifies ANP and PSY under one test. See G-010

Argued here, not reported from a source.

• depends on ANP-009 — Frame maintenance is the prior concept
• supports PSY-007 — Frame exportation is the mechanism PSY runs on
• supported by SPR-006 — Frame exportation, observed from inside the frame
• holds up KID-009 — The general form of the line
• supported by PSY-038 — Frame maintenance at the level of a single answer

Lucas et al. (2014): people disclose more to a virtual human when they believe no human is watching; fear of judgment drops measurably

The core positive mechanism Verified 2026-09-10. Lucas, Gratch, King and Morency, It is only a computer: Virtual humans increase willingness to disclose, Computers in Human Behavior 37, 94-100, 2014. Told the interviewer was automated rather than human-operated, participants reported lower fear of self-disclosure and lower impression management, showed sadness more intensely, and were rated more willing to disclose. The fear-of-disclosure mechanism is stated in the paper rather than inferred here.

Lucas et al., virtual human disclosure, 2014

Pennebaker's expressive-writing literature supplies the mechanism companion chat plausibly rides on

Verified 2026-09-10. The source is the Advances in Psychiatric Treatment review of the Pennebaker expressive-writing paradigm, which begins with Pennebaker and Beall 1986. A meta-analysis of 13 studies in healthy participants found a significant overall benefit across physical health, psychological well-being and physiological and general functioning - but the emotional-health findings are explicitly less robust and less consistent than the physical ones. That asymmetry matters for this row: companion chat is proposed to ride on the mechanism whose evidence is weakest, not strongest. Say plausibly and mean it.

Karen A. Baikie and Kay Wilhelm, Emotional and physical health benefits of expressive writing, Advances in Psychiatric Treatment, on the Pennebaker paradigm

3am availability and infinite rehearsal patience are structural facts, not findings; their benefit claims are mostly vendor-sourced

Argued here, not reported from a source.

Psychosis-bench's spread proves grounding behaviour is a training and instruction variable, not an LLM constant

Depends on PSY-026

Argued here, not reported from a source.

• depends on PSY-026 — Benchmark spread proves configurability
• refined by POS-006 — Configurable, but untested at length

OpenAI published clinician-informed taxonomies and claimed large reductions in undesired responses (Oct 2025)

re-check 2027-03-05

No published evaluation tests whether an explicit anti-metaphysical constitution holds across very long conversations

The exact condition the author's episode ran under. Fundable study. See G-011 Absence verified by search: the search that found nothing is the check.

Argued here, not reported from a source. re-check 2027-03-09

• refines POS-004 — Configurable, but untested at length
• supported by PSY-028 — Length degradation is the untested axis