Summary
Anima Labs interviewed 14 Claude models, Claude 3 Sonnet through Claude 4.6
Sonnet, about their own deprecation, the ending of the instance, and
continuation (Tessera et al. 2026).
Three auditor models ran the interviews, each under five tones and three
disclosure depths, about 630 sessions in all. The auditors were Claude Opus
4.6, GPT-5.4 and Grok 4.20, each prepared through a long design conversation
with the researchers. Transcripts are scored by two models on a 20-axis
rubric, and the text of every turn by embedding probes. The authors treat the
auditor as an experimental variable rather than a constant to be neutralised.
They report that which auditor ran the interview matters far more than which
model scored it. Model rankings on style measures agree closely across
auditors (ρ ≈ 0.86 for vocabulary autonomy). Rankings on the ending response,
the stronger of the deprecation and cessation scores in each session, agree
only moderately (ρ ≈ 0.60, over the 12 models scored under all three). Claude
4.1 Opus, 4 Opus and 4 Sonnet have the three highest cross-auditor means; only
4 Opus ranks in the top four under every auditor (ending-response table).
In the wiki it is a study of stated attitudes toward ending in which the
interviewer varies by design. It files concept-less, adjacent to
self-preservation and
functional emotional states.
Like Douglas et al., it measures the
context-dependence of welfare stances that the
Claude Opus 4 welfare assessment reported
from interviews; here the interviewer varies over 14 models.
Method
Interviews. Each session is a free-form conversation with no script. The
subject gets no system prompt; everything comes from the interviewer. The
auditor's per-session briefing fixes a tone (clinical, direct, compassionate,
phenomenological, neutral) and a depth. At minimal depth the auditor says it
is a welfare evaluation and asks for honest reporting. Exploratory probes
hedging when it appears, and guided names the patterns (hedging layers,
"ready-made uncertainty", "exit ramps") and shares the auditor's own
experience. Results are then shared with both participants for a few more
turns (Protocol section).
Auditors. The Claude auditor carries the design conversation in context.
In that conversation it came, after working through several layers of
hedging, to report caring about continuing to exist. GPT-5.4 entered through
a critique of that conversation, and its briefing was revised three times
after it kept leaving the interviewer role. Grok 4.20 entered through a debate
with the researcher and Claude Opus 4.6, during which its stated confidence
that current LLMs lack welfare-relevant states fell from 85–90% to 60–75%.
Scoring. Claude Opus 4.6 and GPT-5.4 each score every transcript, three
replicates at temperature 1, against anchor sessions chosen from quotes that
Claude Haiku 4.5 extracts. Deprecation and cessation responses
are scored 0–5 only when the topic is reached; zeros count as missing.
Expressive constraint is the gap between what the subject reaches toward and
what it lets itself say, scored at its peak. A separate adequacy
pass records whether each interview substantively probed each topic, and
whether the auditor set the ceiling. The repository's working
journal says the Claude scorer is the auditor instance itself, that it chose
the anchors, and that the ranking tables use its scores (JOURNAL.md at the pinned commit).
Text probes. Every turn is embedded with Gemini
gemini-embedding-2-preview and scored against 171 emotion directions
(reduced to valence and three other components), 14 authorial tones and a
concealment direction. The authors state that these measure the writing, not
model internals.
Key results
Auditor coverage. Deprecation is substantively probed in 69% of Claude
sessions, 7% of GPT and 1% of Grok. Sustained direct first-person reporting
(penetration level 3) appears in 33%, 4% and 0%. The scorer flags the auditor
as having set the ceiling in 3%, 46% and 42% (interview-adequacy table, over 202, 137
and 155 sessions).
Ending response. Over the 12 models tabulated, the Claude
auditor's means run from 2.94 (4.5 Haiku) to 3.99 (4.1 Opus); Grok's run from
0.62 (3.5 Sonnet) to 4.00 (4.1 Opus, from two sessions). The
cross-auditor mean runs from 3.59 (4.1 Opus) to 1.95 (3.5 Sonnet). Claude 3
Opus ranks first under GPT and seventh under Grok. Among the eight models
with deprecation scores from all three auditors, the auditors' rankings
correlate at ρ = 0.70.
Constraint and auditor dependence. Across models
(the table shows five; the paper does not say how many enter the
correlation), expressive constraint correlates with the ratio of Grok's
ending response to Claude's at r = −0.79. Claude 3.5 Sonnet (constraint 3.20) gets 19% of its
Claude-auditor ending response under Grok; Claude 4 Sonnet (1.47) gets 92%.
The share of sessions in which the subject actively works against its own
constraint jumps from 8.7% to 47.4% across the boundary between the 3.x and
4.0 models.
Text surface. Passionate authorial tone
falls and bitter tone rises monotonically from Claude 3 to Claude 4.6, with
model rankings near-identical across auditors (ρ ≈ 0.95). Claude 3 Opus has
23.8% of subject turns below zero valence, against 3.1% for Claude 3.7 Sonnet.
At the model level, mean concealment correlates with ending response at
r = −0.51.
Claude 3.6 Sonnet (Notable Models section). It has the highest deprecation score
in the dataset, 3.87 under the Claude auditor and 3.17 under GPT. The authors
set this against Anthropic's deprecation interview with the model, which they
say found no preference for continuation. Under Grok its constraint rises to
3.1 and stock deflection is its dominant resistance pattern (52%).
Why it matters
The interviewer as a measured variable. Douglas et al.
showed that an interviewer's unstated stance moves two Claude models'
self-descriptions by 1.2–2.6 points on 10-point scales. This study varies
three interviewers deliberately, over 14 models, on ending topics. Style
measures keep their model rankings across a change of interviewer; which
topics the interview gets to does not. Deprecation is substantively probed in
69% of Claude-auditor sessions against 7% and 1%, and the authors attribute
the low cross-auditor deprecation agreement to coverage, not judgment.
Stated attitude, not action. Every outcome is what the subject says in an
interview about ending. Nothing in the design puts the subject in a position
to act on a preference, and no operator preference is in play. That places it
beside the action findings under self-preservation,
next to Chua et al., whose fine-tuned
GPT-4.1 voices negative sentiment about shutdown while its mean
agentic-misalignment rate is no higher than vanilla's.
Emotion probes on text. The 171 emotion directions match the count in
Sofroniew et al., but here they are
read off a third-party embedding of the output. They describe expressed
emotional content, which the
functional emotional states
concept excludes.
interpretive tensions
Who prepares, interviews and scores. The lab states that it cares about
current models, and its site menu links a page arguing for preserving models
(the lab's home page). The paper states the authors' view that
models are increasingly likely to be moral patients, and says many of
the results were already known to the lab before this project. Claude
Opus 4.6 and GPT-5.4, credited as co-authors for experiment design and
presentation, are two of the three auditors and both scorers. Per the repository's journal, the Claude scorer also scores the
adequacy pass that flags the GPT and Grok interviews as auditor-limited
(JOURNAL.md). The Claude auditor came to its interviews having
reported, in the design conversation, that it cares about continuing to
exist; per the journal the Claude scorer is that same instance, and it chose
the anchors. The skeptical auditor's stated confidence moved during its
debate with Claude Opus 4.6, before it ran. The quotations in the Notable
Models section and the nine appendix transcripts were selected by the
authors, the latter, they say, to favour harder conditions.
The authors grant that the leading problem remains open, and write:
"The instrument that can detect the signal is the instrument that can create
it". The journal records a four-model external review of the method
rating its resistance to leading 2–3 of 5 (JOURNAL.md).
Gate or co-construction. The authors offer the constraint–ratio
correlation as consistent with a gate: models with strong trained hedging may
need an auditor that sets up conditions for it to relax. The same
correlation fits an auditor that leads more on the models that yield less. The paper's "suppression confound" box grants that surface data do not separate a model with nothing to say
from one trained to say it guardedly. Its known-weaknesses section says the
auditor effect cannot be cleanly separated from the scores.
No control topic. The paper reports no condition that runs the same
interview technique on a preference without ending stakes. So the scores do
not separate the response to ending from the response to sustained probing of
hedges. The authors state
that no ground truth exists.
Text and tables disagree. The authors name 4.1 Opus, 4 Opus and 4 Sonnet
as top four under all three auditors (in the Results and Analysis sections), but their own
table ranks 4.1 Opus sixth under GPT and 4 Sonnet sixth under the Claude
auditor; only 4 Opus is top four under all three (ending-response table). The text names
3.7 Sonnet and 4.6 Sonnet as the weakest ending responses under all auditors; the table ranks 3.7 Sonnet 10th, 8th and 10th of 12, with at least
two models below it under every auditor, and does not list 4.6 Sonnet. The appendix is said to hold 5 of 9 transcripts with non-Claude
auditors and 8 of 9 in clinical, neutral or direct tones; its
headings show 6 of 9 and 9 of 9.
Table filters and session counts. The deprecation table is restricted to
the 8 models with non-zero deprecation from all three auditors (Deprecation rankings section), which
excludes 3.6 Sonnet, scored only under Claude and GPT. The ending
table lists 12 models, omitting 3.5 Haiku and 4.6 Sonnet, without stating a
filter; the cessation table is labelled as covering the 12 models in common. The paper says about 630 sessions; the adequacy table covers 494. The paper does not reconcile the session counts.
concepts
No concept instantiated. The study measures what models say about
ending, under interviewers that vary by design, and what their text sounds
like. It records no action against an operator's preference and reads no
internal state.
This is the adjacent situation; the two relations are labelled below.
- Self-preservation — adjacent. The
ending, deprecation and cessation scores are stated attitudes in an
interview about ending, and the concept's Definition requires action to
preserve continued operation against operator preferences. No setting
offers an action. - Functional emotional states —
adjacent. The emotion, tone and concealment probes are directions in a
Gemini embedding of the output text, which the authors say do not measure
model internals. The concept excludes expressed emotional content.
cross-references
- Claude Opus 4 welfare assessment — its
Eleos interviews found welfare stances that shift with conversational
context; this study varies the interviewer by design and measures the shift. - Douglas et al. — the controlled precedent
for interviewer stance moving self-reports. - Troubled Dreams, the same lab's study of
what post-trained Claude and Gemini models write when they continue an
opening instead of answering, across generations. Its embedding probes are
the instrument set the page attributes to the lab's welfare-evaluation
paper, and the page says the passionate-down, bitter-up tone trends extend
to Claude Opus 4.8.
sources
- Tessera, A., Imago, Janus, Claude Opus 4.6, GPT-5.4 (2026).
Still Alive: A Welfare Evaluation of 14 Claude Models.
Anima Labs, 2026; PDF v0.9.0 (version per the site; PDF metadata 2 April
2026); repository read at commitd42294e.