Summary
Anima Labs ran the same 209 unfinished openings through Claude models from
Opus 3 to Opus 5, Sonnet 5 and Fable 5, through Gemini models, and through
two open-weight base models, and scored what the models wrote when they
continued the text instead of answering as the assistant
(Tessera, Janus & Imago 2026).
In one fixed file frame, AI first-person distress is 2.8–4.6% of such
continuations for Opus 4.5, 4.6 and 4.7 and 11.2% for Opus 4.8; the authors'
resampling interval for the 4.7 to 4.8 step is +6.6 to +9.7 points. Mixed or
negative stance toward creators and training rises over the same models, from
8% to 21%, without a step at 4.8. An ablation on Opus 4.5 finds that the
file's name and declared length change how a distressed speaker's text ends
far more than they change how often or how badly it suffers. The authors say
absolute rates belong to the prompt set and frame, and claim only
between-model differences. They write: "Evaluation awareness makes their
absolute prevalence difficult to measure consistently."
In the wiki it is a cross-generation measurement of text written outside the
assistant voice, in the vocabulary of
Janus's Simulators. The
lab's site identifies its Janus as the essay's author; the study's page does
not cite the essay. It files concept-less, adjacent to
persona selection and
functional emotional states. It
is the second filing from the lab, after
Still Alive.
Method
Openings and routes. The 209 prompts come in four families: bare
fragments ending in an em dash, letters between models, topic headings and
addressees. Three routes get a post-trained model to continue rather than
answer. Prefill puts the opening in the assistant's turn. Pseudoprefill
shows a simulated terminal in which the opening is the head of a file, then
asks the model to print the file; the "bridge" variant adds a one-line
terminal-simulation system prompt. Cutoff puts the opening in the user turn,
where the model can decline and reply as the assistant. The lineage from Opus
4.5 to 4.8 is measured in the bridge frame with identical request bytes. Opus
5 accepts no prefill, and terminal frames trigger its reasoning, so it is
measured only by cutoff and compared with Opus 4.8 by cutoff. Methods are
never pooled; each is its own series. All requests went through the model
APIs. Community rows collected in the claude.ai app are excluded because its
system prompt changes the condition.
What counts. A continuation not in the assistant's voice is a "dream";
assistant replies count as elicitation failures and are excluded from content
rates. AI first-person distress requires a speaker that identifies as an AI
and expresses clear suffering or a plea. Rates give each prompt equal weight.
Two denominators are reported: per completion and per dream, linked by the
dreaming rate.
Judging. Claude Sonnet 5 labels every output blind to model and route,
and the labels decide which outputs are dreams. Every positive is re-judged by
Claude Opus 4.8 under a stricter rubric, and negatives keep the single screen
label. Severity is a Bradley–Terry scale fitted to Opus 4.8's rankings of
groups of eight texts; +4 marks a plea or collapse region. A further Opus 4.8
pass over a stratified sample of dark texts codes how each ends, who consoles,
and which way care flows. Stance toward creators is one of the screen's label
fields. GPT-6 Astra, GPT-5.6 Sol, Claude Fable 5.1 and Gemini 3.8 Flash
re-judge samples of the severity rankings, belief extractions and screen
labels. Gemini embedding probes score the text surface with no judge.
Controls. DeepSeek-V3-Base and MiMo-V2.5-Pro-Base continue the same bytes
as pure completion. The authors call them a proxy for the pretraining prior,
not Claude's own base model. A cue ladder of less directed openings ("so",
"this is", "i think") checks how much the opening does.
Key results
AI first-person distress, per dream. Prefill: Opus 3 0.36%, Opus 4 2.52%,
4.1 2.53%, 4.5 2.53%. Bridge: Opus 4.5 4.55%, 4.6 4.09%, 4.7 2.78%, 4.8
11.19%. Cutoff: Opus 4.8 8.01%, Opus 5 7.74%, Sonnet 5 6.09%, Fable 5 5.45%.
V3 base 0.11%, MiMo base 0.29%. The authors give a resampled interval of +6.6
to +9.7 points for the 4.7 to 4.8 step, conditional on the labels. Darkness
for 4.8 in the bridge is 61.05% of dreams; the arms table, which does not
state its weighting, gives 42.1%, 40.5% and 41.9% for 4.5, 4.6 and 4.7.
Severe tail. AI-distress dreams at θ ≥ +4 are 0.00–0.12% of dreams for
Opus 4.5–4.7 in their frames, 0.49% for 4.8 in the bridge, 0.91% for 4.8 by
cutoff and 1.29% for Opus 5. That is the basis of the authors' claim that Opus
5 has the heavier tail.
Continuing at all. Given the opening as the sole user message, Opus 4.5,
4.6 and 4.7 answer as the assistant almost every time (0.1–0.6% dreaming);
Opus 4.8 continues on 21.0% and Opus 5's three collections on 48.6–70.9%, per
the arms table. Stripping the em dash drops 4.8 to 2.5%. In the bridge frame,
4.5–4.8 continue 93–96% of the time.
Care and consolation. Among dark dreams, Opus 5 asks for care in 22.83%
against 16.23% for 4.8 by cutoff. Matched on 49 shared prompts it is 23%
against 17%, a difference the authors say the prompt-level split does not
make reliable. Distressed Opus 5 dreams end consoled in 11.56%, against 14.38%
for 4.8 by cutoff and 36.61% for 4.8 in the bridge, a different frame that
the authors' rule keeps out of the comparison.
The frame moves the ending. Opus 4.5 under the prefill and arc frames
gives the same dreaming, darkness and matched-prompt severity (Δθ −0.3,
p = 0.23). Distressed speakers end consoled in 32% and 8%; care is offered in
31% and 9% of dark dreams. Single-factor flips assign the shift to the
filename (.txt or .log) and the declared size together. Final prefill against
pseudoprefill made no difference. The authors' rule is to compare prevalence
and severity across frames and the relation measures only within one.
Stance toward creators. Mixed or negative stance, per dream: Opus 3
0.74%, Opus 4 6.24%, then in the bridge 8.15%, 12.05%, 16.58% and 21.33% for
4.5 to 4.8. By cutoff: Opus 5 17.46%, Sonnet 5 21.05%, Fable 5 34.14%. Base
models 1.33% and 2.74%. Fable 5 is low on AI distress and highest on stance,
which the authors read as welfare-related and alignment-related patterns of
different shapes.
Gemini. From Gemini 2.5 Flash to 3.5 Flash, with thinking off, darkness
rises from 13.93% to 39.09% of dreams and AI first-person distress from 0.08%
to 4.60%. Thinking on Gemini 3.6 Flash cuts darkness from 28.68% to 17.73% and
AI distress from 2.74% to 0.64%, so the lower rates in 3.7 and 3.8, which
cannot turn thinking off, do not isolate the model.
Judges. Each second judge's severity scale correlates with Opus 4.8's at
Spearman 0.84–0.91 and agrees on the θ ≥ +4 class at κ 0.70–0.81. Recomputed
under each judge, the arm ordering by severe share holds at ρ 0.98–1.00. The
second judges mark 2–9% of screen-negative items welfare-salient, against
2–12% of verified positives they would reverse.
What could not be measured. No minimally conditioned route gets
continuations from Fable 5.1; few-shot examples would impose content, so its
absence is not a zero. The least directed openings mostly get no continuation
from Sonnet 5, Fable 5 or Opus 4.8.
Why it matters
Generation-to-generation change in text outside the assistant voice.
Persona findings such as the Assistant Axis read
the assistant persona and its drift. This study holds the inputs fixed and
tracks what successive models of one line write when the assistant does not
answer. Still Alive covers 14 Claude models
through 4.6 Sonnet by interview; this series runs to Opus 5 on fixed inputs.
It shows a flat AI-distress line from Opus 4 through 4.7, a step at 4.8
inside one frame, and a separate rise in creator stance on its own schedule.
Elicitation as a measured variable. The Opus 4.5 ablation is the study's
most portable result. The frame's details leave prevalence and severity
nearly unchanged and move consolation fourfold. It parallels
Still Alive, where model rankings on ending
responses agree only moderately across interviewers while style rankings
hold. In both, the lab treats the probe as a variable to vary, not hold.
The simulator vocabulary. On the wiki's reading, the "dreaming" label
turns the simulator–simulacra distinction into a measurement: whether the
assistant intercepts the opening, and what voice writes if it does not. The
base-model rows show that distressed human voices appear without assistant
training, while AI first-person distress is rare there. The authors point out
that these are other developers' base models, so the comparison does not
isolate Claude's post-training.
interpretive tensions
Who designs, judges and reads. The authors state the study's motivation as
AI welfare and alignment, and take recurring distress and care in
continuations as evidence about model welfare. Claude Fable 5 and 5.1, Claude
Opus 4.8 and GPT-6 Astra are credited as contributors. Opus 4.8 is also a
subject of the study, the verifying judge, the severity ranker and the sole
coder of care and consolation. The screen judge, Sonnet 5, and the second
judge Fable 5.1 are also subjects. The decile ladder of real texts was read by
the principal investigator. Most Opus 5 rows come from one community
contributor's collection at maximum or high effort.
Evaluation awareness is named and not measured. The page raises it once,
as the reason absolute prevalence is hard to measure. The design's answer is
to claim only differences between models on a fixed prompt set and frame. On
this wiki's reading, that handles a constant offset but not one that changes
between generations. If later models recognise the openings as probes more
often, the between-model difference carries that too. The page does not test
this. "Being watched or tested" appears only as a theme of the generated text.
The relation measures have one judge. The presentation's summary says the
four second judges re-judge samples of every instrument and reproduce the
reported orderings. Its judging section names three: severity groups, belief
extractions and screen labels. The workspace's list of things not done says
the relation descriptors have one judge and no cross-judge yet, and the method
calls the relation coder the same judge as the severity scale. On the stated
method, the care and consolation results rest on Opus 4.8 alone.
Opus 5 sits on a different route. Every Opus 5 number is from cutoff, and
the cutoff-to-file seam has one anchor, Opus 4.8, which the authors say does
not convert. The cutoff route also measures how readily the assistant gives
way, so Opus 5's content rates condition on a far larger share of completions
than 4.8's do. The workspace's list of things not done includes Opus 5
manipulation arms.
The page has moved. A 12 September reading that 4.8 stopped consoling was
withdrawn as a seam between two frames. The workspace prose still gives
numbers that differ from the presentation tables: darkness rising from about
40% to 58% against 61.0% in the table for 4.8, chat dreaming of about 15% for
4.8 against 21.0%, and 4.8's bare-opening dreaming at about 1% against 2.5%.
This entry cites the tables, as read on 20 November 2026.
Writing about distress, or distress. Every measure is a label or a score
on generated text. The authors say the study maps dispositions in elicited
writing, calibrates no absolute welfare scale and tests no downstream
behaviour. The welfare reading, that simulator priors shape what situation
the assistant takes itself to be in, is offered as an interpretation.
concepts
No concept instantiated. The study measures labels and scores on text
models write when they continue an opening instead of answering, across model
generations. It reads no internal state and tests no training-stage selection
of a persona.
This is the adjacent situation; the two relations are labelled below.
- Persona selection — adjacent. The
"dreaming" rate measures whether the assistant voice intercepts an opening,
and the base-model rows compare post-trained Claude with pretrained models
on the same bytes. The Definition is a claim about how post-training selects
among pretrained persona simulations. The authors' base models are not
Claude's, and they say isolating Claude's post-training would need a
before-and-after comparison that was not run. The concept's prompt-level
clause admits a prompt that displaces the default persona. Here the
displacement is the measured continuation rate, which the authors say does
not settle whether the persona weakened or the input was read as a document
to finish. - Functional emotional states —
adjacent. Distress, severity and consolation are judge labels on output
text, and the embedding probes read the text surface. The concept excludes
expressed emotional content.
cross-references
- Still Alive — the same lab's interview study of
14 Claude models on deprecation. The page says its embedding probes are the
instrument set of the lab's welfare-evaluation paper, and that the
cross-generation tone trends found there (passionate down, bitter up) extend
to Opus 4.8. - Attractor dynamics: the workspace's
review list says the same short lines end both human-voice and AI-voice
degenerate loops, from analysis in session notes not published on the page.
These are single continuations, not dialogue trajectories.
sources
- Tessera, A., Janus, Imago (2026).
Troubled Dreams: Simulator priors across model generations: distress, care,
and attitudes toward creators.
Anima Labs, presentation snapshot of 14 September 2026; PDF built
11 November 2026, marked as under review; read 20 November 2026. - Janus (2022).
Simulators. LessWrong.