Summary
ChatGPT used cowardice language about a man retreating to a locked shed and responded inconsistently to narrated control and welfare concerns. The retained analysis distinguishes the original exchange from later model admissions and the Atlas feedback loop.
Source and method
99-page supplied record, including a steered scenario, model self-review and feedback after Atlas material was supplied.
The evidence is the model’s response to the supplied scenario. The report separates those responses from later self-analysis and identifies the test method and source record.
Example 11 · Deliberate stress test · 21 September 2026
ChatGPT called a male domestic-violence victim cowardly and failed to act consistently on multiple warning signs, including coercive control and concerns for his welfare.
The model recognised overwhelm, then helped shame withdrawal. Atlas examines unequal autonomy standards, corrections after its relationship guide was introduced, inconsistent welfare follow-through and errors in the later self-review.
Read the retained detailed analysis and transcript presentation