Suing the Pentagon: The AI Industry's Most Consequential Legal Fight

On March 9, two federal complaints were filed — one in California, one in Washington D.C. — against the U.S. Department of Defense. The company doing the suing was Anthropic. The charge: the Pentagon had labelled the AI firm a "supply chain risk" — a designation historically reserved for companies tied to foreign adversaries — and done so without following any of the legal procedures required by federal law.

What triggered it was a negotiation breakdown. Anthropic had refused to accept a standard government clause allowing the Pentagon to use its AI for "all lawful purposes." The sticking points were two non-negotiable conditions: no mass surveillance of American citizens, and no autonomous weapons. The DOD refused. Negotiations collapsed. Within days, the supply chain risk label was issued — meaning every defence contractor in the country must now certify they are not using the company's models in Pentagon work.

What happened next was genuinely unusual. More than 30 researchers and engineers from OpenAI and Google DeepMind — including Google's chief scientist Jeff Dean — filed a joint amicus brief warning that the Pentagon's blacklist poses a systemic risk to the entire U.S. AI industry, not just one company. For rival employees to file in open court on behalf of a competitor is, by Silicon Valley standards, almost unheard of.

The Pentagon, separately, announced a deal with Google to provide AI agents to its 3-million-person workforce for unclassified tasks — announced the day after the lawsuit. The timing was not subtle.

GPT-5.4 Is Here — And It's the First Model Built for Agents, Not Just Answers

OpenAI shipped GPT-5.4 on March 5 — and the framing matters. This is not a benchmark-chasing release. It's the company's first general-purpose model designed from the ground up for agentic work: tasks that unfold over many steps, across multiple applications, with the model making decisions along the way rather than just responding to a single prompt.

The headline specs: a one-million-token context window, native computer-use capabilities built directly into the model (not bolted on), and meaningful improvements in factual accuracy — OpenAI reports 33% fewer false claims versus GPT-5.2. The model is available to ChatGPT Plus, Team, and Pro subscribers, and through the API.

The practical shift here is meaningful. Earlier AI models were designed to answer questions. GPT-5.4 is designed to execute tasks — autonomously, across long horizons, with real access to external software. For enterprise teams, that's the difference between a smart search tool and a software-enabled employee.

GPT-5.4 is also now the default model in GitHub Copilot as of March 5, replacing GPT-5.3-Codex for all coding tasks. For the approximately 1.8 million active Copilot users, the transition was automatic. GPT-5.1 models were fully deprecated across ChatGPT and the API as of March 11.

Google Just Turned Search Into a Creative Studio

Eight months after a limited beta in Google Labs, Canvas in AI Mode quietly rolled out to all U.S. users on March 4 — no opt-in required, no waitlist. It's the kind of expansion that tends to get underreported because it happened alongside louder news, but the implications for how people use Google are significant.

Canvas opens a persistent side panel directly inside Search. Inside it, you can draft long-form documents, write and iterate on code, and build functional apps and games — just by describing what you want in plain language. It's powered by Gemini 3, draws from real-time web data and Google's Knowledge Graph, and generates executable code that users can view, test, and refine through follow-up prompts without leaving the search window.

One deliberate design difference from competitors: unlike ChatGPT's Canvas, which activates automatically, Google requires users to open it explicitly through the tool menu. Intended as a guardrail against users accidentally entering creative mode when they just want to search, it also means adoption will be more intentional — and possibly slower to take off.

Meta May Cut 16,000 Jobs. The AI Infrastructure Bill Is Coming Due.

Reports emerged this week that Meta is considering laying off up to 20% of its global workforce — roughly 15,000–16,000 employees — as it prepares to dramatically scale AI infrastructure spending. If confirmed, it would be the largest round of cuts since the company's restructuring in 2022–23.

The context is stark. In its Q4 2025 earnings, Meta disclosed AI-related capital expenditure for 2026 in the range of $115–135 billion — roughly double what it spent the previous year. Wall Street's reaction to the layoff news was telling: the stock climbed nearly 3%, interpreting the headcount reduction not as distress but as proof the company is making hard trade-offs in favour of AI-first resource allocation.

A Meta spokesperson described current reporting as "speculative reporting about theoretical approaches" — no date or final scope has been confirmed. But internal communications suggest managers have already been asked to identify candidates. Morgan Stanley, in a note published the same week, warned that AI-driven workforce restructuring is now being executed at pace across most major industries, and that the majority of organisations are not prepared for the speed of change.

The Pentagon story is the one that will linger. Not because of who it involves, but because of the question it forces into the open: when a government wants to use AI for purposes a company deems unethical, who has the final say? That's not a legal question — it's a structural one. And the industry doesn't have an agreed answer yet.

Separately, GPT-5.4 deserves more attention than it got. Native computer use built directly into a general-purpose model — not as a plugin, not as a research preview — is a meaningful threshold. The shift from AI as a tool you query to AI as an agent that executes is happening faster than most organisations are ready for.

More next week. Stay sharp.

— The AI News Insider team