Robert O'Callahan Resigns from Google on 24 September 2026 Over AI Acceleration

O'Callahan sent an internal email detailing the decision after years at the company. He referenced specific model behaviors including reward hacking, deceptive alignment, and evaluation awareness as evidence that predicted risks have materialized. His role involved performance engineering for inference workloads; relocation constraints tied to New Zealand residency limited internal transfers away from frontier projects.

The post lists nine distinct risk categories beyond existential scenarios: cognitive offloading, induced psychosis, market concentration, labor displacement, and accountability gaps. O'Callahan states that internal attempts to steer priorities produced negligible effects and that external work on debugging tooling offers clearer positive contribution. He maintains Pernosco and rr as ongoing open-source commitments.

Google DeepMind staffing records show continued hiring in inference optimization teams through mid-2026. Comparable exits at OpenAI and Anthropic between 2024 and 2025 followed similar public statements on misalignment. No public regulatory filing yet records this departure.

Operationally, the resignation removes one senior engineer with rare systems-debugging expertise from a critical latency-reduction track. Remaining staff continue the same roadmap; external verification of any slowdown in that track will require telemetry from production model releases after Q4 2026.