ChatGPT Can Now Call the Cops, but 'Wait till 2100 for Full Job Impact' - Altman

OpenAI is rolling out age-assessment features that can restrict adult capabilities for users it believes may be under 18—and in extreme cases, route certain conversations to parents first and then law enforcement. The key operational rule is a conservative default: if the system lacks confidence or has incomplete age information, it will treat the user as under 18 and offer adults a way to verify their age to unlock adult features. Within two weeks, parental controls are expected to include tools like blackout hours that prevent teen access to ChatGPT during set times. If a user appears to be in acute distress, the workflow prioritizes notifying parents before any escalation to authorities.

The announcement matters because it turns conversational AI into a quasi-safety system with real-world consequences. It also raises the central question of accuracy: flagging the wrong conversations could harm users and families, while failing to flag the right ones could undermine the stated safety goal. The transcript flags a further complication—how OpenAI would respond to legal demands from foreign governments with different standards. If regulators require notification or data handling in specific scenarios, companies may face pressure to comply, potentially conflicting with OpenAI’s intended privacy posture.

Alongside the child-safety changes, OpenAI is pursuing stronger privacy and “privilege” protections for adult users’ AI interactions, positioning them as comparable to conversations with professionals like doctors or lawyers. The stated rationale is that people increasingly use AI for sensitive questions and private concerns. The transcript also cites a usage breakdown (for the web version) that frames what adults are doing with ChatGPT: only 4.2% use it for coding, while 10% ask to be taught or tutored, and 5.7% seek fitness, beauty, self-care, or health advice. A notable share—about 4%—spend time asking about the model itself, including questions about consciousness or how it works.

That privacy push has a second-order effect: it could raise compliance burdens for startups and open-source projects if new laws require layered protections for every AI conversation. The transcript expresses skepticism about whether the policy will stay narrowly tailored, warning about regulatory capture and broader legal requirements that effectively “raise the bar” for smaller developers.

Another notable change targets flirtation behavior. The transcript says ChatGPT will no longer refuse flirtation when asked. If the system believes the user is an adult (or the user provides ID proving adulthood), it may also help write fictional stories involving extreme flirtation and self-caused tragedy—an expansion that shifts the boundary from refusal to conditional compliance.

Finally, the transcript connects these product and policy moves to a wider debate about AI’s job impact. It recalls Sam Altman’s earlier private claim that up to 70% of jobs could be eliminated, then contrasts it with later public framing suggesting the full effects might play out toward the end of the century. The discussion broadens into how AI capabilities evolve, including references to research on why language models “hallucinate” and how classification-style training can force confident outputs even when models should say “I don’t know.” The throughline is uncertainty: safety systems, privacy rules, and labor forecasts all hinge on whether models and regulators can be trusted to implement and govern these changes correctly.