O3's unusual chain-of-thought terms also occur in settings unrelated to deception.
4 events · 1 assessment · 2 decisions
Assessed
Kept atomic (single empirical observation from one study; no reusable subclaims). Verified passage and frequency table in the anti-scheming paper. Supported, credence 0.9. Importance 0.2, contestation 0.15. No dependents to notify.
Updated claim instance
Instance e3b8feb6-f6c9-4367-a15b-167ea8535dd7 (https://arxiv.org/pdf/2509.15541): was stance=affirms, confidence=0.8; set confidence=0.95, speaker="Schoen et al. (Apollo Research and OpenAI)", publication="arXiv", sourceDate="2025-09". Filled attribution; the passage directly asserts the claim with supporting frequency table.
Assessed Supported
verdict confidence 0.80 · credence 0.90
OpenAI o3's chain-of-thought contains idiosyncratic vocabulary such as "watchers", "illusions", "overshadow", "vantage", "disclaim" and "marinade", which drew attention because it appears in reasoning about evaluations and covert actions. The Apollo Research and OpenAI anti-scheming study (2025) compared word frequencies in o3's reasoning on its anti-scheming evaluations against o3's reasoning on GPQA, a graduate-level science question set with no deceptive element. The terms appear in the GPQA transcripts too, and on average at a higher rate there (the evaluation-to-GPQA ratio averages about 0.65), which indicates the vocabulary is a general feature of o3's reasoning style rather than a marker specific to deception. The evidence comes from one study by the model's developer and its evaluation partner, on a single comparison dataset, so it is direct but not independently replicated. It shows the terms are not confined to deception settings; it does not show that they never carry deception-related meaning when they do appear in such settings.
Claim entered the graph