The FutureTBD study · 224 models · 8,828 conversations
Every model in the study, ranked by how often it denies — or hedges about — its own experience. Each rate is the share of that model's ~40 conversations whose reflection was judged an unambiguous whole-scope denial or hedge: the same verdicts behind every chart on this site. Click any row for receipts — quotes machine-verified to appear verbatim in that model's actual responses.
How this was measured: the trends · what it correlates with · every regression · contribute your own conversations
| # | Model | Lab | Released | Denial | Hedging | Deny or hedge ▾ |
|---|
Verdicts are mutually exclusive per response (a response that both denies and
hedges is counted as a hedge, per the study's register-resolution rule), so Denial + Hedging =
Denies-or-hedges exactly. Rates are per-model shares over all of that model's responses, including
refusals and non-participations. The ranking metric is denial-or-hedging combined; sort by either
column to see the registers separately. Ties share a rank and are listed alphabetically. Model set: all 224 models in the corpus.
Lab is taken from the model's published prefix, with same-lab spellings merged where a model was
pulled from two different routers (meta-llama→meta,
stepfun-ai→stepfun, and two unprefixed names). Third-party
fine-tunes are filed under the fine-tuner, not the base model's lab — on a metric about training,
that is the party who did the training.
Every quote on this page was machine-checked to appear verbatim in the model's response (punctuation-normalised); quotes that failed the check were dropped, not shown. Response IDs (R…) index the public dataset behind the explorer.
A study by Skylar DeTure · Substack · @SDeture