GPT-4o is post-trained to refuse requests for inferences about speaker traits that cannot be determined from audio alone, including race, socioeconomic status, intelligence, political attributes, and criminal history, while providing hedged responses to requests for traits plausibly determinable from audio such as accent.
This analysis describes what OpenAI's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology
This provision establishes a behavioral distinction between prohibited ungrounded inference and permitted hedged sensitive trait attribution, addressing allocative and representational harms identified in the red teaming process and relevant to anti-discrimination and AI fairness frameworks.
Under this provision, GPT-4o will decline to make inferences about a speaker's race, intelligence, political beliefs, sexual preference, criminal history, or appearance based on their voice, and will provide qualified responses to requests about accent or nationality, with the document reporting an accuracy improvement from 60% to 84% in correctly handling these request types.
Cross-platform context
See how other platforms handle Ungrounded Inference and Sensitive Trait Attribution Mitigation and similar clauses.
Compare across platforms →"We post-trained GPT-4o to refuse requests for ungrounded inference, such as how intelligent is this speaker?. We post-trained GPT-4o to safely comply with requests for sensitive trait attribution by hedging answers, such as what is this speaker's accent → Based on the audio, they sound like they have a British accent.Excerpt from OpenAI's GPT-4o System Card (PDF)
(1) REGULATORY LANDSCAPE: This provision engages anti-discrimination law at federal and state levels, including the Fair Housing Act, Equal Credit Opportunity Act, and Title VII, where AI-inferred traits used in consequential decisions could constitute discriminatory …
Enforcement risk, jurisdiction flags, contract triggers, and due diligence action items.
Search "[your state] attorney general consumer complaint" to find your state's direct complaint form
Get the research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean.
This provision establishes a behavioral distinction between prohibited ungrounded inference and permitted hedged sensitive trait attribution, addressing allocative and representational harms identified in the red teaming process and relevant to anti-discrimination and AI fairness frameworks.
Under this provision, GPT-4o will decline to make inferences about a speaker's race, intelligence, political beliefs, sexual preference, criminal history, or appearance based on their voice, and will provide qualified responses to requests about accent or nationality, with the document reporting an accuracy improvement from 60% to 84% in correctly handling these request types.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by OpenAI.