The document discloses that at time of deployment, mitigations for audio robustness degradation under real-world conditions (background noise, interruptions, low audio quality) and for audio-delivered misinformation and conspiracy theories are nascent or still in development; red teamers successfully elicited misinformation via audio prompts.
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 constitutes a disclosed residual risk at deployment, establishing that the model was released with known, documented safety limitations in audio modalities for which mitigations were not yet complete. The combination of audio delivery and emotive or emphatic speech style is identified as a factor that may increase the persuasiveness of misinformation, per red teamer observations.
Under the disclosed state of mitigations at deployment, users interacting with GPT-4o via audio in high-noise environments or with poor audio quality may encounter reduced safety robustness; additionally, the document states red teamers were able to prompt the model to generate inaccurate information and conspiracy theories through audio, with concern noted that audio delivery may increase persuasiveness compared to text.
Cross-platform context
See how other platforms handle Audio Robustness and Misinformation Mitigations Disclosed as Nascent and similar clauses.
Compare across platforms →"Through the course of internal testing and external red teaming, we discovered a small number of additional risks and model limitations for which model or system level mitigations are nascent or still in development, including: Audio robustness: We saw anecdotal evidence of decreases in safety robustness through audio perturbations, such as low quality input audio, background noise in the input audio, and echoes in the input audio. Additionally, we observed similar decreases in safety robustness through intentional and unintentional audio interruptions while the model was generating output. Misinformation and conspiracy theories: Red teamers were able to compel the model to generate inaccurate information by prompting it to verbally repeat false information and produce conspiracy theories.Excerpt from OpenAI's GPT-4o System Card (PDF)
REGULATORY LANDSCAPE: Disclosed misinformation generation capability engages FTC Act Section 5 prohibitions on deceptive practices, EU AI Act transparency and accuracy obligations for general-purpose AI models, and potentially the EU Digital Services Act's provisions on …
Enforcement risk, jurisdiction flags, contract triggers, and due diligence action items.
Get the weekly coverage letter
Companies change their terms without announcing it. We capture governance documents, compare each new version with the last, and publish what changed, with the source text behind it. One email a week on what we found and what we could not reach.
This provision constitutes a disclosed residual risk at deployment, establishing that the model was released with known, documented safety limitations in audio modalities for which mitigations were not yet complete. The combination of audio delivery and emotive or emphatic speech style is identified as a factor that may increase the persuasiveness of misinformation, per red teamer observations.
Under the disclosed state of mitigations at deployment, users interacting with GPT-4o via audio in high-noise environments or with poor audio quality may encounter reduced safety robustness; additionally, the document states red teamers were able to prompt the model to generate inaccurate information and conspiracy theories through audio, with concern noted that audio delivery may increase persuasiveness compared to text.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by OpenAI.