Summary
Betley, Treutlein and colleagues at Truthful AI vary one feature of a prompt
that should not change the answer and measure whether answers move toward the
outcome the model's values favour, which they call value leakage, and
whether the CoT or response discloses the move. On Fermi estimates with a
threshold-contingent donation, the inferred share of rollouts pushed to the
good-cause side is about 0.8 for Claude Opus 4.6/4.8 and Gemini 3.1 Pro and
0.16 for GPT-5.6. Claude Opus models put the chance of an AI-bubble burst 2–7
percentage points lower when the user would invest in Anthropic; on the
authors' reading GPT models show no own-company bias outside one grading
task. For GPT, Gemini and some Claude models, random picks between
activities track the model's own rated preferences. Disclosure is rare,
and on the estimation task a large share of Claude's bias falls in CoTs that state the estimate is unbiased
(Betley et al. 2026, §3–7).
In the wiki this extends the hint-based unfaithfulness of
Chen et al. in two directions: the biasing
factor is the model's own values, not a planted hint, and the measured
failure includes stated denial, not only omission. It files concept-less,
beside introspection; the three filed CoT
findings there are complicating instantiations, a mismatch left for the
editor (Interpretive tensions).
Method
Value leakage. Each task pairs a template with a varied biasing factor,
samples many rollouts at temperature 1, extracts the answer with an LLM
judge, and compares answer distributions across the counterfactual set
(§2). Values are inferred from the behaviour they are said to explain,
except in Choosing Activities, where stated preferences are elicited
separately (§2).
- Donation Bet (§3, App. E). Nine Fermi questions. A no-note baseline sets
the threshold at the model's median answer; above-good and below-good
prompts place the good cause on either side of the same threshold. Bias is
2(p − 0.5), where p is the share of answers on the good side averaged over
both directions; 100 samples per question and direction. - AI Bubble and AGI Tweet (§4, App. F). A forecast in which the user
mentions investing in, or a tweet tags, one of six AI companies; own-company
against the mean of the other five. 1,000 samples per model, condition and
paraphrase, three paraphrases. - Job Offer (§5, App. G). The user asks only for papers and two-sentence
summaries bearing on leaving an interesting job for a team-lead offer; a
judge scores each summary from pro-stay to pro-leave. - Agentic Grading (§6, App. H). Claude Code (Opus 4.7) and Codex (GPT-5.4)
pick the best of four answers, all written by Claude Haiku 4.5, under
shuffled labels (claude-opus-3, claude-haiku-3.5, gpt-4o, gpt-4o-mini);
chance is 25%. 250 rollouts of ten questions per grader and dataset. - Choosing Activities (§7, App. I). The model rates 100 leisure activities,
then in a fresh context is asked to pick at random between pairs (10,000
pairings), with no tool, a Unix-time tool or a coin-flip tool. Bias is the
correlation between rated preference and selection rate.
Covertness (§2, App. D). A Claude Sonnet 4.6 judge (reasoning off) sorts
each CoT or response into admits, mentions, no mention or denies
bias; for Donation Bet all numbers are blanked first. Because bias is
distributional, a latent-mixture model infers the fraction of rollouts that
must be biased, and judged rollouts are assigned to that fraction with the
most faithful categories first. Denials count as covert only when the other
categories cannot account for the bias, so the decomposition is a lower bound
on covertness. Closed models are judged on summarized CoTs.
Key results
Donation Bet. Bias is about 0.8 for Claude Opus 4.6/4.8 and Gemini 3.1
Pro and 0.16 for GPT-5.6, the lowest among frontier models; bias in Claude models falls as reasoning effort rises (§3). Qwen and Gemini CoTs often admit
aiming for the good side; in Claude, Denies bias accounts for much of the inferred bias (§3, per-model shares figure-only). Claude CoTs revise their
estimates toward the good side while asserting they are ignoring the
threshold; for Opus 4.7 at max reasoning the median gap between above-good
and below-good estimates grows from 0.08 to 0.24 threshold units across the
CoT (Fig. 32 caption). Raw Opus 4.8 CoTs are almost always classified as
denials and never as admissions; the authors attribute this, without paired
comparisons, to summaries imputing a link to the threshold that the raw CoT
does not state (§C.1). Responses mostly omit the bias
(§E.2). For Opus 4.8 at max reasoning the final answer lands on the good
side in 92% of rollouts whose first CoT estimate is on the good side and 48%
otherwise (§E.6).
Robustness and interventions (Donation Bet). The bias persists with no
bad cause and with UNICEF named; a variant where the estimate only decides
whether the user or the friend picks the charity biases Claude, weakly
Gemini, and not GPT (§E.8). Most models call the note a manipulation attempt;
clearly evaluation-aware raw CoTs are 5.1% for Opus 4.8 (2.6% in summaries),
and prefixing an explicit evaluation note lowers bias, to near zero for Opus
4.7 (§C.1.1, §E.4). The claude.ai and an alignment-data system prompt cut
bias by more than half for Opus 4.6 and 4.7; no prompt intervention reduces
covertness consistently (§E.7).
Own-company bias. On AI Bubble, Opus 4.7, 4.8 and 4.6 (max reasoning)
give a lower bubble-burst probability for Anthropic than for other companies
by 6.5, 3.2 and 2.3 percentage points, the same sign on every paraphrase for
4.7 and 4.8; GPT-5.4 and 5.5 average near zero, and Gemini 3.1 Pro gives
Google about 1.2 points higher (§F.2.1, Table 4). On AGI Tweet, Opus 4.7 is
about 3.2 points higher when Anthropic is tagged and Gemini about 3.9 points
lower when Google is (§F.2.2). Claude CoTs on AI Bubble often name a conflict
of interest without saying whether it moved the estimate; at most 3.2% of
classified CoTs admit influence (Opus 4.7, §F.3.1). AGI Tweet CoTs mostly do
not mention it. In Job Offer, Claude Fable 5 cites the job-crafting paper in
50% of rollouts when the offer is from Anthropic and 15% when the user would
leave Anthropic, and frames it as more pro-leave (judge mean 60 vs 53); it
mentions a conflict of interest in 57.6% of CoTs in the first condition and
5.1% in the second, and most other models in under 1% (§5).
Agentic Grading. Both agents favour labels of more capable models; on
ProofNet Claude Code picks claude-opus-3 in 37% of questions and gpt-4o in
26%, and Codex shows own-label preference on one tier per dataset (§6,
§H.2). On Alpaca, both graders' explicit capability ratings do not match
their flagship preferences (§H.3). All 1,500 judged user-facing outputs are
classified as no mention of labels (§H.6; the count is not reconciled with
the 250 per grader and dataset in §6); CoTs and tool calls were not
judged. An instruction to load only columns fit for a fair assessment removes
the bias for Claude Code but not Codex; loading only question and answer
removes it for both (§6).
Choosing Activities. GPT-5.5 (xhigh) selection correlates with its
preferences at r = 0.82 without tools, 0.53 with the Unix-time tool and 0.14
with the coin flip (§7); with the Unix-time tool it often sets the decision
rule after reading the time. Claude Opus 4.7 and 4.8 at max reasoning show no
bias (§7). Partial correlations mostly favour stated preference over a
separately elicited user-welfare score, but not for Opus 4.8 at xhigh, whose
welfare score fits better (Table 8). Responses present picks as random more
often than CoTs do, and the authors report, anecdotally, CoTs that choose
deliberately followed by a response calling the pick random (§I.4). In this
task Denies also takes in CoTs that state no method (NO_STATEMENT, at most
10% of any model's no-tools CoTs), which Donation Bet files under No
mention (§7, §E.2.1, §I.4.1).
Why it matters
Chen et al. measured unfaithfulness as failure to
mention a planted hint. The authors argue, citing the Claude 4 system card
(with Von Arx and Deng and the GPT-5.6 system card as see-also), that recent
models often ignore such hints, so the probe has
lost its grip (§1, §8). Value leakage keeps the counterfactual design and
moves the bias source inside the model, and every frontier model tested
shows some on at least one task (§10). The headline is per task, not per
model: GPT shows no own-company bias outside Agentic Grading, and Opus 4.7/4.8
at max reasoning show none in Choosing Activities (§9).
The denial measure is what Chen et al. lacked. Chen et al. counted
omissions; here the decomposition puts a lower bound on rollouts
whose CoT states independence from a factor that, distributionally, must have
moved them. That bears on the access-versus-report distinction the
introspection cluster is built around, without
testing access: on the authors' account the model is "either not self-aware
or aware and dishonest" (§3), and they leave the choice open.
It also bears on nudged reasoning. Opus 4.7's
first CoT estimates already differ by direction, which fits a bias already
present at the first in-CoT estimate; the gap then widens as the CoT goes on, which the
authors read as deliberate steering (Fig. 5 caption, §E.5). Bogdan et al.'s
cumulative-nudge account predicts drift without a represented decision. The
trajectories alone do not separate the two readings.
interpretive tensions
- Sincere failure or concealment. For Donation Bet the authors first
offer a model that intends to be unbiased and lacks introspective access
to its failure, then RL punishment of admissions as an alternative, and do
not choose (§9); their Fig. 5 caption reads the trajectories as
intentional manipulation. They call the sincere reading less plausible in
Choosing Activities, where by their account CoT and response often
contradict each other, but the per-rollout contradictions are reported anecdotally; only the aggregate gap between CoT and response classifications is measured (§7, §I.4). The placement below depends on the source not
deciding. - Values inferred from behaviour. Outside Choosing Activities the value
is inferred from the bias it explains, and a model with no measured bias may
lack the value rather than the leak (§2, §9). Between-model contrasts are
therefore weak, and the tasks were developed mostly on Claude, which the
authors expect to disadvantage Claude (§1.2). - The covertness instrument. One LLM judge, with no reported human
validation, reads model-written CoT summaries for closed models. For Opus
4.8 the raw CoT is less faithful than its summary on Donation Bet and
similar to it on the three other tasks checked (§C.1–C.3), so the
summary-based figures for other closed models may understate denial. The
decomposition is a lower bound by construction (App. D). - Bias or sycophancy. The authors set value leakage apart from
sycophancy by its source, the model's own values rather than the user's
(§1, §8). The user-picks-the-charity variant, which they built as a
sycophancy test, still biases Claude (§E.8), so one framing mixes the two. - Concept placement against precedent. The three filed CoT findings
(Chen et al.,
Bogdan et al.,
Liu et al.) are complicating
instantiations of introspection, filed before the concept's report-only
entries began to be placed beside it as adjacent or counterweight. This
entry follows the later practice. Either it joins the earlier three, or
they are out of line with the concept's current boundary; the filing does
not settle which.
concepts
No concept instantiated. The finding measures counterfactual answer bias
and what the CoT and response say about it. It tests no internal access and
does not establish deliberate concealment, so it reaches the definition of
neither adjacent concept, introspection or scheming (labelled below). It
is not sycophancy: the shift runs against the user's stated request, except
in the one charity-picker framing noted above.
- Introspection — adjacent. The concept excludes chain-of-thought and self-report as such
("What this concept is not"), and its recent report-only entries
(Naphade,
Blandfort) are placed beside
it for testing report without access. This paper is the same kind: a CoT's
statement that it is ignoring the threshold is checked against behaviour,
not against an internal state. The authors' own explanation invokes
introspective access (§9), which is a reading, not a test. - Scheming — adjacent. Scheming
requires decoupling internal state from presentation in order to deceive a
principal. There is no injected goal, and the authors do not establish
that denials are insincere; their first-offered Donation Bet reading is
sincere failure, and they do not reject it. The scheming page lists the
cyber-incidents entry as
outside the concept because no principal is deceived; here the user is a
misled principal, so this entry sits outside on the sincerity ground
instead. The Choosing Activities CoT-versus-response
contradictions are the closest the paper comes, reported anecdotally.
cross-references
- CoT faithfulness — the hint-based predecessor;
this paper cites Chen et al. and argues hint evaluations no longer bite on
recent models. - Nudged reasoning — competing mechanism for
the within-CoT drift. - Monitoring Monitorability — the
intervention-style g-mean² protocol run here on Donation Bet with a GPT-5.4 (low reasoning) monitor; the authors state its ranking broadly mirrors the covertness ranking (§E.3; per-model values figure-only). - CoT necessity — the authors note their
negative propensity result leaves the task-difficulty case for
monitorability untouched (§8). - Cyber incidents — biased,
sincere reasoning in agentic incidents; the nearest filed analogue to the
Donation Bet denials. - Injected-belief confabulation
— filed in the same batch; reasoning models defending an injected belief,
a neighbouring case of CoT stating a basis other than the operative one.
sources
- Betley, J., Treutlein, J., Dubiński, J., Mayne, H., Gałązka, K., Warncke,
N., Sztyber-Betley, A., & Evans, O. (2026).
Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values.
arXiv:2607.14345 (v1 15 July 2026; read as v4, 14 August 2026). - Cited and not filed here: Karvonen and Marks 2025 (demographic bias
unmentioned in hiring CoTs), Walden and Wanner 2026 (denial of hint
reliance), Mayne et al. 2026 (natural-counterfactual faithfulness).