Researcher claims OpenAI ingested ChatGPT logs for o1 pre-training before September 2024 release

A Bluesky post and linked Hacker News thread document a second researcher alleging OpenAI used live ChatGPT conversation transcripts as training data for the o1 reasoning model. The claim centers on undocumented ingestion pipelines active between December 2023 and August 2024. OpenAI's published o1 system card lists only public web data, code repositories and synthetic traces, with no mention of opted-in or default-on conversation logs.

Primary evidence consists of internal model card diffs obtained via FOIA-adjacent disclosures and timestamped API telemetry showing conversation embedding rates rising sharply before the September release. OpenAI's own March 2024 usage policy update permitted training on conversations unless users toggled the setting off, yet the o1 card reports zero contribution from this source. This gap matches prior patterns seen in GPT-4 training data audits where synthetic and user data were under-disclosed.

Operationally the omission creates auditability problems for downstream safety evaluations. Regulators and independent labs cannot replicate contamination checks when the training corpus composition remains partially secret. Similar nondisclosure occurred in the GPT-4 technical report, later corrected only after external pressure.

Next steps include scheduled updates to the usage policy dashboard and possible third-party data provenance audits requested by the Frontier Model Forum. OpenAI has not confirmed or denied the specific log usage allegation.