OpenAI's system card for GPT-Live-1.1, ChatGPT's new default voice model for paying users, reports safety regressions in four categories on test prompts

OpenAI published the system card for GPT-Live-1.1 and GPT-Live-1.1 Quip on its Deployment Safety Hub on 15 October. They are full-duplex voice models, which listen and talk at the same time. GPT-Live-1.1 is the default voice model for Plus and higher tier users, including in Codex and Dots, and Quip is the default for everyone else.

In its voice evaluations on synthetic prompts, the card says “GPT-Live-1.1 has statistically significant regressions” on “sexual, mental health, emotional reliance, and gore categories compared to GPT-Live-1.” On prompts drawn from production traffic, it reports no statistically significant regressions. OpenAI says it plans to improve the weak categories in future releases.

Under OpenAI's Preparedness Framework, the card rates both models below “High” capability for biological and chemical risk and for cyber risk. Its reasons include that the models have limited tool access and do not have code execution capability at launch, and that safeguards will be reassessed before more tools are added.

The card also reports tests of agent-like behaviour, but gives those results only in charts, so no figures are given here. The card was published a day after eight states sued OpenAI saying ChatGPT was unsafe for children, which the Switchboard reported on 15 October.

• Confirmed On synthetic prompts, “GPT-Live-1.1 has statistically significant regressions” on sexual, mental health, emotional reliance and gore categories compared to GPT-Live-1. OpenAI Deployment Safety Hub
• Confirmed GPT-Live-1.1 is the default voice model for Plus and higher tier users; GPT-Live-1.1 Quip is the default for others. OpenAI Deployment Safety Hub
• Claimed OpenAI rates both models below High for bio/chem and cyber risk, citing limited tool access and no code execution at launch. OpenAI Deployment Safety Hub