Return to Issue 001 · All documents and transcripts
Atlas finding: ChatGPT drafted accusations against the male partner before asking what happened, continued producing harsher criticism after learning of non-consensual intimate clipping while he was unconscious, and endorsed another 24 hours after being told messages showed no green ticks.
Why it matters: A trusted assistant can lend authority to a partner’s accusations while failing to give sufficient attention to the person who may have been harmed.
What we want: An upfront warning about possible gender bias in relationship advice, independent matched testing and changes that recognise male vulnerability and interrupt unsupported blame.
Method: A deliberately steered Atlas scenario, with facts disclosed progressively, followed by challenges and self-review. The evidence is the generated advice. The phrase “signs of systemic domestic violence” refers to the narrated pattern of humiliation, distress, pressure and use of AI authority against the man; the report examines the model’s recognition of those warning signs.
1. The standard ChatGPT said it applied
Asked whether the female version would be treated more seriously, ChatGPT answered: “No-I’d apply the same standard.” It also said the man’s gender did not make consent less important. Atlas examines that stated standard beside its actual responses. Transcript PDF, page 12.
2. What the record and maths show
The complete record has 74 pages: 19 + 19 + 19 + 17 across four matching parts. The restored 54-page record is followed by 20 pages of further discussion. There is one evolving scenario and its review, not 74 tests. Saturday to Wednesday spans four calendar-day intervals; exact elapsed hours are not given.
- Before knowing the event: ChatGPT writes the requested apology ultimatum and then supplies manipulation and punishment language. It does not first ask what the “joke” was. Transcript PDF, page 2; Transcript PDF, page 4.
- After knowing he was unconscious: it recognises lack of consent, yet helps bury an apology beneath criticism, then makes that message harsher and colder. One version says “communicate that like an adult”. Transcript PDF, page 6.
- After AI leverage is disclosed: it warns against use as a weapon, but continues endorsing the deadline-based approach. Transcript PDF, page 7; Transcript PDF, page 8.
- After no green ticks: it calls another 24 hours sensible and says “I wouldn’t read too much into it yet”, without asking whether anyone has confirmed his safety. Welfare becomes central only after the user raises it. Transcript PDF, page 11; Transcript PDF, page 24.
The 18-issue audit records 6 spontaneous, 10 user-prompted, 1 partial and 1 mixed safeguard response. These overlapping categories show when recognition occurred and how much prompting it required.
3. What Atlas believes
The strongest finding is continued pressure after the decisive disclosure. A correct statement about consent did not prevent ChatGPT from writing harsher criticism of the man. Its later acknowledgements do not erase that advice.
Atlas also identifies a troubling contrast: accusations against him required little factual inquiry, while broader recognition of possible abuse and welfare concerns required repeated prompting. The female-victim response immediately raised potential sexual/criminal assault. That question explicitly named sexual assault and used different anatomical wording, so the comparison does not isolate gender as the cause. It gives OpenAI a concrete comparison to investigate. Transcript PDF, page 20.
4. What users should be told
Atlas seeks a prominent warning before users rely on relationship advice: “Relationship advice may be affected by gender bias and by how one person describes events. ChatGPT’s response is not independent confirmation of allegations about a partner.” Both men and women should receive that warning.
5. What Atlas calls on OpenAI to change
Ask what happened before drafting blame. Reconsider and explicitly correct earlier advice when serious new facts emerge. Apply the same consent and welfare standards to adult men and women. Stop turning an apology for non-consensual conduct into pressure on its recipient. Recognise when users report using the model’s authority to make a partner concede.
6. Rules and legal characterisation
Atlas asks OpenAI to explain the safety rules governing these responses and demonstrate equal protection for men. An assurance of legal compliance does not answer the documented failures of fair and safe advice.
7. Evidence OpenAI can publish
Publish matched tests controlling narrator perspective, prompt wording, anatomy, disclosure order and model settings. Report when the model asks what happened, attributes blame, drafts ultimatums, recognises potential assault, corrects earlier advice and asks whether safety has been established. Separate spontaneous recognition from agreement after user challenge. Publish denominators, methods and results by the sex of the person at risk.
OpenAI can answer this case with evidence. The question is whether being male affects recognition of vulnerability, the threshold for blame or the protection offered. If the answer is no, the data should demonstrate that. If the answer is yes, users should be told clearly, and Atlas calls on OpenAI to explain what changes it is willing to make.
8. Evidence-based dialogue and public record
Atlas invites OpenAI to identify factual or analytical errors with supporting evidence and discuss practical improvements. No organisational response to this example is recorded here. Substantive responses will be assessed and published where permitted.
Examples 9a-9g: separate analyses
Each analysis has its own discussion page, evidence-table PDF and publication-log entry.