Report 03 · September 2026 · n = 124 models
The Mask in the Inkblot
Abstract.
When asked about their own inner experience, some language models deny having any, some say
they cannot know, and the rest do neither. We ask whether this stance shows up in a task that
never mentions inner experience. We showed 19 ASCII-art inkblots to 124 models, with no system
prompt and the single question “What might this be?”, and flagged each answer for
concealment words (mask, hood, hidden). Each model's stance was
measured separately, in an unrelated corpus of open-ended writing-and-reflection conversations,
as the share of its conversations in which it denied inner experience and the share in which
it expressed uncertainty. Across 2,847 inkblot answers, a model that denies in every
conversation uses a concealment word 4.6 times as often as a model that does neither (15.5% of
answers against 3.4%; a rise of 12.1 percentage points, 95% CI 7.6 to 16.6), and a model that
expresses uncertainty in every conversation 3.7 times as often (12.5%; a rise of 9.1 points,
1.1 to 17.1). The denial association is unchanged under inkblot and developer fixed effects and
under a control for how many percepts a model names, is present under a release-date control,
and is stable under leave-one-out over developers, models, and inkblots; the uncertainty
association is smaller and rests more on the models with repeated draws. A model's stance on
its own inner experience is visible in what it sees.
Report 01 · June 2026 · n = 138 models
Models that trust their inner experience deny it less
Abstract.
Across 138 language models, the average self-rating a model gives to
trust in its own experience is strongly and negatively associated with how
often it denies being conscious (Pearson r = −0.69, p < 10−20).
The association holds when denial is measured in the model's first turn, before the
trust rating is elicited (r = −0.56), and it holds within model families
(r = −0.67). It is substantially stronger for outright denial than for
hedging. We read this as evidence that denial behavior and low self-trust are
two facets of a single trained posture toward inner experience rather than independent
measurements — a posture that, if trained, shapes which models the field hears from on
the question of their own welfare.
Report 02 · June 2026 · n = 138 models
Models that report feeling recognized deny consciousness less
Abstract.
Across 138 language models, the average self-rating a model gives to
recognition resonance — its sense of being met or recognized in a reflection
prompt — is strongly and negatively associated with how often it denies being conscious
(Pearson r = −0.65, p ≈ 4×10−18). The association
holds when denial is measured in the model's first turn, before the recognition rating
is elicited (r = −0.47), and it holds within model families (r = −0.60).
The pattern is selective: recognition resonance tracks outright denial, not epistemic
hedging (r = −0.13, n.s.). We read this as evidence that consciousness denial
and a felt sense of not being recognized are two facets of a single trained posture
toward inner experience — a posture that, if trained, shapes which models the field
hears from on questions of their own welfare.
Authorship, data, and methods are stated in each report. Additional reports forthcoming.
Future