The Pentagon turns commercial AI into workplace infrastructure

ChatGPT Mil and Grok for Government now share GenAI.mil with Gemini, a platform designed for more than three million people. The move is more than a technology purchase: it institutionalises competition between models and places their usage limits at the heart of public policy.

By Rubén Campoy3 September 2026Updated 11 September 20265 min read

Work area in an operations centre with laptops and screens displaying mapping data — Exclusive editorial image · BOLDERROR

On 31 August, the US Department of Defense expanded its GenAI.mil platform with two adapted versions of commercial products: OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government. Both join the government version of Gemini with which the system began operating in December 2025. The development matters because of its scale: the platform already reports more than 1.7 million unique users and is designed to reach a workforce of more than three million military personnel, civilians and contractors.

The tools are accredited to work in an Impact Level 5 environment, including controlled unclassified information. ChatGPT Mil launches with chat, files, projects and custom GPTs; the official focus is on document-related tasks, planning, internal policy, logistics and administration. Grok is presented with reasoning modes, persistent workspaces and reusable playbooks for market research, procurement, supply chains and collaboration. This is not a chatbot placed on an intranet, but an organisation-wide productivity layer capable of shaping working habits.

Three models, one institution and limits that are not interchangeable

The multi-model design has an obvious strength: it reduces dependence on a single supplier and allows responses to be compared. However, three model-generated opinions do not amount to three independent sources. The systems may share training data, biases and error patterns. Genuine diversity also requires authorised search engines, internal repositories with traceability and a policy requiring the evidence used for each decision to be retained.

The notable absence is Claude. Anthropic was excluded from certain military contracts after refusing to accept uses involving domestic surveillance and autonomous weapons. On 27 August, a federal judge ruled that the company’s designation as a supply chain risk was unlawful. A few days later, the Commerce Secretary said that Anthropic was back on the right side of the administration. The political tone has softened, but Claude is still not among the options on GenAI.mil, and the government retains the option to appeal.

THE KEY

That is the crux of the debate. On a civilian platform, the terms of use are a contract between supplier and customer. In defence, a restriction can become a dispute over who defines what constitutes a legitimate use of a strategic technology. The supplier provides the model and its safeguards; the state claims the authority to decide how it uses it. Neither position alone resolves problems such as assigning responsibility, human review or handling an incorrect response within a chain of command.

Security accreditation protects the environment and the data, but it does not make every model output correct. Role-based access controls, immutable logging of queries and actions, tool limits, separation between information and operational orders, and human review when a recommendation affects people or critical systems will still be necessary. The familiar interface may be the very risk: the more ordinary the tool seems, the easier it becomes to forget that it is still producing probabilities, not certainties.

The Pentagon turns commercial AI into workplace infrastructure

Three models, one institution and limits that are not interchangeable

Tags

Related

OpenAI turns Navier-Stokes into a stress test for agent-driven science

AlphaGenome Atlas delivers nine billion variants on a petabyte-scale scientific platform

Images 2.5 Moves Visual Generation from Single Renders to an Editing Workflow

China normalises cloud-native without abandoning the data centre

AWS and Azure turn private interconnection into a service, not a project

Anthropic Settlement Enters Awkward Phase: Deciding Who Gets Paid