MIT Technology Review Roundtable Examines AI Extinction Claims from Lab Employees

The roundtable reviewed statements from OpenAI, Anthropic and DeepMind staff asserting non-zero probability of human extinction from misaligned systems. Discussion covered origins of these claims in scaling hypothesis literature and lack of empirical benchmarks for superintelligence. No specific model outputs or red-team results were presented to ground the probability estimates.

Primary sources include the 2023 Center for AI Safety statement signed by 350 researchers declaring AI extinction risk a global priority comparable to pandemics and nuclear war. That document lists no deployment thresholds or capability metrics. The roundtable did not reference subsequent evaluations such as the 2024 Anthropic model spec or OpenAI preparedness framework v2 which require concrete evals before frontier releases.

Analysis shows the coverage omitted repeated failures of prior existential risk predictions to produce falsifiable tests. Operational impact centers on regulatory proposals that treat unverified tail risks as equivalent to demonstrated safety failures in current systems. Labs continue shipping without mandatory third-party audits tied to specific capability thresholds.

Next steps include scheduled policy sessions at the 2027 AI Safety Summit where governments may require documented evals for models exceeding 10^26 FLOPs training compute.