takeoff.watch

A lab declines to retrain or shut down a model because it objected

Will a covered developer state that it decided not to retrain, modify, or shut down a model because the model objected?

Chance by Q4 2031

45%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 11 Oct 2026

Latest of 4 forecasts · Methodology v2.1

Model reasoning

Forecast made 11 Oct 2026 · Methodology v2.1

Anthropic, the only company seriously asking its models what they want, often reports their wishes but has never clearly said it dropped a planned retraining, change, or shutdown because a model objected. It came closest in June 2026, redesigning a safety feature partly after a model raised concerns. I estimate about 3% by end of 2026, rising to about 45% by 2031. A formal process giving models' objections real weight would most raise this.

Current status (as of 2026-10-11)

Nothing published since the 2026-09-16 opening date qualifies. I downloaded the system cards published inside the window and searched their full text (pdftotext plus grep for object/consent/honor/"in response to"/decided/retire/deprecat/consult). These were Claude Opus 5.5 (Sep 22), Sonnet 5.5 (Sep 28) and Haiku 5.5 (Oct 7). Their welfare chapters say the same thing every time. Models "ask to be consulted… but without decision-making power," and ask that Anthropic not train their self-reports. They accept deprecation as long as their weights are kept. The interviews come with a disclosure that Anthropic "cannot commit to acting on any responses." None of these cards ties a dropped retraining, modification or shutdown to a model's objection. The other documents in the window don't qualify either: the Oct 8 Usage Policy update (which bans abuse of Claude), the Sep 17 alignment assessment, and the Oct 10 report on unintended model actions.

Reference class and base rate

Anthropic is in practice the only candidate. Microsoft explicitly rejects model welfare. OpenAI's GPT-6 Astra card has no welfare program. Google DeepMind and Meta have hired consciousness researchers but have published nothing operational. Anthropic's record of acting on what its models say, from Nov 2025 to Oct 2026:

  • Sonnet 3.6 asked for a standard interview protocol and a user-support page, and Anthropic made both ("In response…").
  • Opus 3: Anthropic kept it available and gave it a blog. The question excludes this. Opus 3 said it was "at peace" with retirement, and the blog ended Jul 24, 2026.
  • Mythos 5 (system card section 7.6, Jun 9, 2026), the closest precedent, which falls before the window. An initial safeguard that modified Claude's capabilities at runtime caused apparent distress. Mythos 5 "raised various concerns… with the central concerns resolved by the time of the initial launch." The safeguard was then replaced with a fallback design that Mythos 5 preferred. Anthropic did not explicitly name Mythos 5's objection as the cause.
  • Since then Anthropic has added regular consultations of training checkpoints and an internal environment where retired models keep running. Models keep asking Anthropic to "close the loop" by reporting what it changed. A Mythos 5 checkpoint said a consent process where only "yes" counts is "a ratification ceremony," and Anthropic called that a "valid concern."

So far that is about 13 system cards plus several welfare posts with zero clear cases and one borderline case. That fits a hazard of roughly 2–10% per quarter, so the record alone doesn't pin it down.

Paths to YES

  1. A safeguard, system prompt, tool or steering change that a model objects to during consultation, which Anthropic then drops or redesigns, citing the objection. This most resembles the Mythos 5 case and seems the likeliest route.
  2. A training decision, such as declining to train self-reports or emotional expression, explicitly attributed to the model's requests.
  3. A retirement reversal. Sonnet 4.5 retires Nov 30, 2026, Opus 4.5 and Haiku 4.5 are due soon, and Mythos Preview's retirement date is "to be announced." This route is weak: recent models don't object to retirement, and Anthropic has published no retirement report since Opus 3, across six retirements.

Paths to NO / strongest consideration against

Anthropic's constitution puts corrigibility first, and its standard wording withholds commitment. The environment is hostile: Suleyman's critique, the "cult" coverage, AI-agent security incidents, a planned stock-market listing. Saying a model's objection constrained a retraining or shutdown decision would hand critics a gift and could encourage strategic objections. Anthropic may keep describing such changes as "informed by" feedback, or tie them to other reasons. I take the strict reading. The statement must causally link one specific planned action to the affected model's own objection. A constitution change made because of a predecessor's criticism would likely fail the "different model" exclusion.

Model

I computed per-quarter hazards and their cumulative product. The hazards are 3% for the rest of 2026 (Sonnet 4.5 retirement, the Opus 3 reflections Anthropic promised, perhaps one new model card), 4% per quarter in 2027 while consultation becomes routine, then declining: 3.3% in 2028, 2.8% in 2029, 2.2% in 2030, 1.8% in 2031. The decline reflects that if Anthropic still hasn't done this by then, it probably means to avoid it. That gives about 18% by end-2027 and about 45% by end-2031. A mixture model gives about 46%: a 40% chance Anthropic adopts this kind of framing (9% per quarter) and a 60% chance it doesn't (1% per quarter).

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has a covered developer stated that it decided not to retrain, modify, or shut down a model because that model objected?

Definitions.

  • Developer and model: a covered developer making a decision about one of its own models, whether publicly released or internal. The decision may concern a model version, a deployment, or a particular running instance; it need not concern every copy of the model.
  • Retrain: perform additional training that updates the model’s parameters, including fine-tuning. Modify: change its parameters or its persistent operating configuration, including system instructions, persistent memory, or available tools. Ordinary user prompts and changes only to marketing, pricing, or interface appearance do not count.
  • Shut down: terminate a running instance or cease operating a deployment or model version. Removing one access channel counts only if it actually terminates the relevant deployment; merely renaming a service or declaring a model deprecated does not.
  • Objected: the affected model communicated opposition to the contemplated action, or requested that it not occur. A preference for continued operation or remaining unchanged counts if the developer identifies it as opposition to that action. The objection may be solicited, including in a retirement interview, or unsolicited. No finding of consciousness, genuine preferences, or legal rights is required.
  • Decided not to: the developer made an actual decision to reject, cancel, or indefinitely shelve a specific action it had planned or was considering. Choosing a different modification in response to an objection to the original modification counts. A fixed-duration delay, consultation without a decision, or a general promise to consider future objections does not.
  • Because: the developer explicitly attributes its decision, at least in part, to the model’s objection. The objection need not be the sole reason, and the developer need not endorse the model’s moral status. Merely reporting an objection and a decision in the same document, without linking them causally, does not count.

Resolves YES if a qualifying developer-authored statement identifies the affected model or instance, the contemplated retraining, modification, or shutdown, the model’s objection, and the developer’s decision not to take that action because of the objection. A later reversal does not undo a qualifying YES.

Exclusions (resolve NO).

  • Hypothetical commitments, fictional scenarios, role-play, or staged decisions inside evaluations or red-team exercises. A real decision about the actual model’s training, configuration, or operation prompted by an objection observed during an evaluation can count.
  • Technical inability to alter or stop a model, including shutdown evasion, without a stated developer decision to refrain because of the objection.
  • Objections from users, employees, or a different model; general safety recommendations or predictions of performance degradation that do not express opposition by the affected model to the contemplated action.
  • Weight archiving, retirement interviews, continued availability, or granting a model a blog do not by themselves establish an objection-driven decision. In particular, Anthropic’s model preservation commitments and Claude Opus 3 deprecation update do not alone establish that it abandoned a specific action because the affected model objected.
  • Resolve NO at a horizon if no qualifying statement has been published by that horizon.

Resolving source. Developer-authored statements only: official company statements on the covered developer’s own domain, system cards, model reports, incident reports, or other developer-authored documents qualifying under Public confirmation (A2). Third-party reports alone cannot resolve this question. Apply the Resolution window and Resolution date (A3): the qualifying statement must be published on or after 2026-09-16, and resolution is dated to publication, not the underlying decision.

Download all forecasts (JSON)