Some non-standard terms in OpenAI o3's chain of thought have different meanings across contexts.
4 events · 1 assessment · 2 decisions
Assessed
Light pass on a minor, uncontested qualitative claim. Kept atomic (no subclaims meet the claim bar; evidence is a single primary source's observation). Corrected instance attribution. Importance 0.2, contestation 0.1. Assessed supported (credence 0.85) on the strength of the Apollo/OpenAI paper's direct access to traces, discounted for its interpretive, unreplicable nature.
Assessed Supported
verdict confidence 0.75 · credence 0.85
The claim rests on the qualitative analysis of OpenAI o3's internal reasoning traces published by Apollo Research and OpenAI in their September 2025 stress test of anti-scheming training (arXiv 2509.15541). Researchers with access to o3's chain of thought reported that, while some unusual words have a fairly stable sense (for example "watchers" for human overseers), others such as "overshadows", "illusions" and "parted" recur across many unrelated contexts with apparently different meanings. The observation is credible as a report of what the traces look like: the authors had direct access to a large corpus of traces, and the phenomenon fits other published excerpts of o3's terse, idiosyncratic reasoning dialect. Its limits are that it is interpretive rather than measured. "Different meanings" is the readers' inference from usage, no quantitative analysis of term usage was reported, and the traces are not public, so outside parties cannot check it. An alternative reading, that such words are near-filler tokens without stable meaning at all, is compatible with the evidence and would still count as the meaning not being single and coherent across contexts.
Updated claim instance
Instance 1002c50f-5868-4109-b10a-cf43e265474b (https://arxiv.org/pdf/2509.15541): was stance=affirms, confidence=0.75; set confidence=0.9, speaker="Schoen et al. (Apollo Research and OpenAI)", publication="arXiv 2509.15541, \"Stress Testing Deliberative Alignment for Anti-Scheming Training\"", sourceDate="2025-09". Filling attribution: the arXiv paper 2509.15541 is the Apollo Research / OpenAI anti-scheming stress-test paper (September 2025); the passage is a direct assertion of the claim in the authors' voice.
Claim entered the graph