MIT Technology Review sets September 15 roundtable on AI extinction claims

The announcement describes an internal discussion at leading labs where staff have assigned non-zero probability to human extinction from misaligned AI. No new data or model outputs are referenced in the listing. The session focuses on origin of the claims, their evidentiary basis, and possible responses.

The 2023 Center for AI Safety statement, signed by 350 researchers, asserted that mitigating extinction risk should rank alongside pandemics and nuclear war. Concrete Problems in AI Safety (Amodei et al., 2016) catalogued specification gaming and reward hacking but contained no empirical threshold for existential failure modes. Current scaling papers report benchmark gains on MMLU and GPQA yet supply no metric that maps to uncontrolled deployment at planetary scale.

Operational impact remains limited to internal safety teams reallocating compute toward alignment research. Labs have published no incident reports or red-team logs demonstrating loss of control. Regulatory filings in the EU AI Act and US executive orders require risk assessments without defining quantitative extinction criteria.

Next milestone is the scheduled 90-minute exchange. Any consensus on required evidence thresholds or timelines for deployment restrictions will be recorded in follow-up coverage.