US’s DOD Didn’t Expect the AI Industry to Actually Have a Spine

US’s DOD Didn’t Expect the AI Industry to Actually Have a Spine

Microsoft backed Anthropic in court after the Pentagon flagged it as a security risk. Now the entire AI industry is watching which party gets to set the rules.

The US Department of Defense designated Anthropic a supply-chain risk last week.

Microsoft had filed an amicus brief by Tuesday, urging a federal court to block it. And then, a judge in San Francisco was already considering Anthropic’s request for a temporary restraining order by Wednesday.

That escalated fast.

Anthropic’s 48-page complaint, filed Monday in federal court, argues the Pentagon’s move is unlawful and seeks to have the designation declared void.

The core dispute is about guardrails. The Trump administration wants Anthropic’s Claude deployed in military contexts without the safety constraints Anthropic insists on building into its systems.

Anthropic refused. The DOD responded by treating the company as a threat to the supply chain it relies on.

Microsoft’s intervention is the part worth watching closely. The company is not a neutral observer in this case. It integrates Anthropic’s products into solutions it sells directly to the US military, which means the DOD designation hits Microsoft’s own government contracts.

Its amicus brief makes this explicit: the Pentagon gave itself six months to phase out Anthropic, but gave contractors zero transition time. That is a real operational problem, and Microsoft named it as one.

What makes this moment significant is the breadth of the coalition forming behind Anthropic.

Thirty-seven researchers and engineers from OpenAI and Google filed their own amicus brief on Monday. These are companies that compete with Anthropic in the market. They still showed up.

The Pentagon framed this as a national security question. The industry is reframing it as a governance question, one about whether federal agencies can unilaterally punish AI companies for refusing to remove safety constraints from their systems.

We think that reframing is correct. And it may be the more consequential argument in the long run.

Meta Opens WhatsApp to Rival AI Chatbots as EU Pressure Mounts

Meta is backing down, at least temporarily, by making a calculated concession in Europe.

Meta is easing its grip on WhatsApp and letting its AI rival back on to the platform. But there are specific terms and conditions- its own terms.

It comes after the EU regulators forced the tech giant’s hand after a substantial incident.

Meta blocked third-party AI chatbot providers from the WhatsApp Business API on January 15, leaving merely its own AI assistant on the platform. After this, the competitors complained to regulators. From there? The EU took notice quickly.

The European Commission threatened interim measures last month, citing potential irreparable harm to rivals. Italy’s antitrust authority had already acted in a similar way back in December.

As of now, Meta has eased its grip- at least for the next 12 months.

Meta says it will support general-purpose AI chatbots via the WhatsApp Business API in Europe. The tech powerhouse’s framing is that this voluntary step removes any urgency for the Commission to act while the broader investigation continues.

That’s reasonable. However, it sidesteps “why” the situation exists in the first place. And whether this is actually meaningful access.

Meta is charging a fee for that access, and smaller AI companies aren’t happy about it. Marvin von Hagen, the CEO of The Interaction Company (one of the complainants), puts it plainly-

“The pricing Meta introduced makes it just as impossible to operate on WhatsApp as the outright ban did, effectively replacing one anti-competitive restriction with another.”

That’s a targeted critique. And honestly, an opinion like this is hard to dismiss.

Opening a door while pricing out anyone who’d walk through it isn’t really opening the door. The same dynamic played out in Italy, where Meta reopened access after a court order. And competitors say the result is that there is no solution either.

The policy changes now extend to Brazil as well, after a court reinstated an antitrust injunction that’s suspended.

So Meta is dealing with this on multiple fronts simultaneously. And the pattern? It’s less like voluntary compliance and more like minimum concessions, market by market, wherever regulators push hard enough.

ChatGPT 5.4 Is OpenAI’s First AI Model with Native Computer Use Capabilities

Just when you think AI’s next step would be better responses, there’s been a shift. The new era of tech is systems that actually do the work.

OpenAI has released GPT-5.4. The update points to a clear direction for the industry. AI systems are moving beyond answering questions. They are starting to execute tasks.

GPT-5.4 can interact with computers directly. It can read what appears on a screen. It can move a cursor. It can type commands. It can navigate software to finish a job. The model does not just suggest steps. It performs them.

This changes how AI fits into everyday work.

Until now, most AI tools have behaved like advisers. They produced ideas, code, or explanations. Humans still had to open applications and carry out the steps. GPT-5.4 begins to remove that gap.

That is why the industry keeps using the term AI agents.

An AI agent does not simply respond to prompts. It receives a goal. Then it plans the steps needed to reach it. It gathers information. It runs tools. It adjusts if something fails. The model becomes closer to a worker than a chatbot.

For companies building software, that shift matters.

Enterprise tools often require long workflows. A report might require data extraction, analysis, formatting, and presentation. Today, a human moves through each step. An agent can potentially run the entire chain.

That is the promise OpenAI is chasing.

The company also claims GPT-5.4 reduces hallucinations compared with earlier versions. That matters if the model will run real tasks. Automation without reliability creates new problems.

The broader takeaway is strategic.

The AI race is no longer just about building smarter models that give accurate outputs. This new phase focuses on building systems that act inside digital environments. Whoever solves that first will redefine how people interact with software.

GPT-5.4 does not complete that transition. But it pushes the industry much closer to it.

Grammarly’s Expert Reviews Feature Comes with a Scary Realization

AI tools are moving from correcting sentences to simulating expertise. That shift is starting to worry the people being simulated.

Grammarly built its reputation fixing grammar mistakes. Now it wants to replicate expertise.

The company recently introduced an “Expert Review” feature that analyzes a document and generates feedback, inspired by” well-known writers, academics, and journalists. The idea is simple: your draft gets reviewed through the lens of recognized authorities in a field.

The problem is that those experts were never involved.

Reports found the system generating comments that seem to come from real individuals without their permission. Some users even saw feedback attributed to editors from substantial publications like The Verge and The New York Times.

Its feature relies on publicly available work and does not claim endorsement from the named experts, says Grammarly. But the presentation is where things get uncomfortable.

In tools like Google Docs, the suggestions appear visually similar to comments from a real editor. That design choice blurs the line between AI-generated advice and human critique.

For technology leaders, the controversy highlights a deeper tension in generative AI.

Large language models learn patterns from public text. That includes the tone, logic, and rhetorical habits of individual writers. Turning those patterns into a product- especially one that attaches a real person’s name- moves the conversation from training data to identity.

And identity is harder to defend as “fair use.”

The feature also exposes a practical limitation of AI expertise. Writing style can be modeled. Editorial judgment is harder. A system trained on published articles may mimic how someone writes, but that does not mean it understands how they think.

That difference matters.

AI is rapidly becoming a collaborator in professional work, from code reviews to legal drafts. But the Grammarly episode shows how quickly assistance can slip into simulation.

And once software starts simulating people, the debate is no longer about productivity. It becomes about ownership- of voice, reputation, and expertise.