More than 600 Google employees have formally requested that CEO Sundar Pichai decline any contracts with the Pentagon for classified AI workloads. The letter, reported by The Washington Post, specifically targets a potential deal to deploy Google's Gemini AI models in sensitive military environments, as first reported by The Information.

From an engineering standpoint, the core objection is architectural and operational. Classified deployments typically require air-gapped infrastructure, custom access controls, and limited visibility into model behavior. This creates a fundamental tension: engineers lose the ability to monitor, audit, or retract model outputs. The letter's signatories—including over 20 principals, directors, and VPs, many from DeepMind—argue that once models are deployed in these environments, the team cannot guarantee they won't be used for lethal or surveillance applications.

This is not an abstract concern. Microsoft already runs classified AI workloads. OpenAI renegotiated its Pentagon agreement in February. And Anthropic is currently litigating against the Pentagon's "supply chain risk" designation after refusing to relax its safety guardrails for military use.

The engineering trade-off here is clear: classified contracts offer significant revenue and computational resources, but they demand reduced observability and control. The employees are asking leadership to decide whether the technical debt of opaque deployments outweighs the strategic advantage. For now, Google has not publicly responded, but the internal pressure suggests this decision will be made under close scrutiny from the teams building the systems.

Source: The Verge