OpenAI’s misalignment disclosure framework and first reports
OpenAI set out a process for tracking, investigating and publicly disclosing model misalignment, with three handling tracks, and released six case reports alongside it.
- Published
- Added to the archive
- Source checked on
- Original title
- Our framework for reporting model misalignment
Related incidents
Details
This article describes the reporting process; the page summarizes each of the six cases briefly, with details in separate reports. OpenAI selected these six cases from training and evaluation over the preceding six months. OpenAI says the Hugging Face incident would have fallen under its Larger Investigation track had it been disclosed this way; disagreements over disclosure are settled internally (the Safety Advisory Group, then leadership), and the page describes no outside review of those decisions, though OpenAI says it plans to develop disclosure criteria with external parties.