Anthropic’s autonomous weapons stance could prove out of step with modern war

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here.

Anthropic’s stance on autonomous weapons may not survive the future

Much of the AI world is watching closely as Anthropic tangles with the Pentagon over how the government can use the Claude models. Anthropic has a $200 million contract with the Pentagon, but the contract says the military can’t use the AI company’s models as the brains for autonomous weapons or for mass surveillance of Americans. Defense Secretary Pete Hegseth insists, after the fact, that the military should be able to use the Anthropic models for “all lawful purposes.” 

Hegseth summoned Anthropic CEO Dario Amodei to the Pentagon for a Tuesday morning meeting, in which he reportedly gave Anthropic until 5:01 p.m. Friday to comply with the Pentagon’s demand. If Anthropic fails to do so, Hegseth threatened to invoke the Defense Production Act to compel the AI company to supply its models with no guardrails. Hegseth also said the government would declare Anthropic models to be a “supply chain risk,” meaning that all government suppliers would be directed to avoid or discontinue use of Anthropic models. 

Amodei said in an interview after the Hegseth meeting that his company has no intention of complying with Hegseth’s demands. (He’s got a strong case: After all, government officials agreed to the terms.) Amodei explained that the military relies on human judgement to avoid violating people’s constitutional rights. If AI is making the decisions, there will be no human being to object. 

Amodei is right, and his company’s willingness to stand up for its values is laudable. The trouble is, we’re rapidly heading for a future where autonomous systems become the norm in warfare. 

For years, the defense establishment talked about keeping the “human in the loop” in AI weapons systems. Often that human is a government lawyer who can make calls on rules-of-engagement issues on the battlefield. Today the Pentagon is talking more about fully autonomous weapons that can manage more of the “kill chain,” or the series of communications and decisions around the destruction of a target. Military leaders often say that whoever can use technology to shorten the kill-chain will win wars.

Things like electronic warfare (cyberwar), hypersonic missiles, and drone swarms are making war faster and response times shorter. This may eventually preclude the opportunity for human review and decision-making. Increasingly, the U.S. military may be forced to take humans out of the loop in order to stay competitive with its adversaries. 

So the result of Anthropic’s standoff with the Pentagon may be that a safety-conscious AI lab is forced out, and a generally less scrupulous company like xAI is chosen as the alternative.

Trump rips off Mark Kelly’s idea for powering new data centers

In his State of the Union address, Donald Trump spent a few minutes on the subject of new data centers for AI, which has over the past few months become a hot button issue for voters. While the tech industry says it needs hundreds of new data centers to support all the AI it’s building, a growing number of voters now understands that the power grid improvements needed to power the data centers may increase their energy bills. “I have negotiated the new Ratepayer Protection Pledge,” Trump crowed. “We’re telling the major tech companies that they have the obligation to provide for their own power needs.”

Politicos might recognize that message, as it closely echoes what Arizona Senator Mark Kelly, a Democrat, has been saying for months now.  Kelly’s “AI for America” plan would create an industry-financed “AI Horizon Fund” to pay for energy-grid upgrades and workforce reskilling. 

According to Kelly’s plan, Congress could require data center developers to buy or lease enough land to contain both their facilities and the renewable energy infrastructure to power and cool them. The data center operators could also be required to pay to connect the renewable sources to the local grid, should the power they generate go unutilized. 

Trump’s idea is more of a suggestion. As of now it’s non-binding, just words. And there was no mention of how the tech companies would generate their own power. Elon Musk’s xAI, for example, brought its own power to its massive Colossus data center in Memphis. Unfortunately, they were dirty methane-powered turbines, and the facility quickly became one of the area’s biggest polluters.

High numbers of young tech job seekers AI-cheated on skills tests

Cheating on technical hiring assessments went through the roof in 2025, with fraud attempts more than doubling, according to new research from CodeSignal, which runs a developer-skills evaluation platform used in hiring software engineers. The research found that 35% of proctored assessments showed signs of cheating or fraud last year, up from just 16% in 2024. The biggest culprits? Plagiarism, having someone else take the test for you, and sneaking in AI tools that aren’t allowed.

The jump was especially noticeable among entry-level candidates. Fraud rates for junior roles nearly tripled year over year—going from 15% to 40%—making early-career hiring a particularly vulnerable spot in the recruiting pipeline. In a press release accompanying the report, CodeSignal CEO and cofounder Tigran Sloyan partly blamed the normalization of AI tools, noting that 80% of Gen Z reportedly uses AI in daily life, which has made the line between acceptable help and outright cheating much blurrier. “Accessibility to AI also makes unauthorized assistance harder to detect and raises the stakes for maintaining fair and reliable skill evaluation,” he noted.

CodeSignal’s detection systems—which combine AI analysis, human review, and digital monitoring—identified a few common patterns across flagged assessments. About 35% of candidates frequently looked off-screen, suggesting they were consulting outside resources during the test. Another 23% showed unusually linear typing patterns, where complex solutions just appeared with barely any pauses or debugging. And 15% had answers that looked a lot like known solutions or leaked content. (It’s worth noting that these numbers reflect attempts that were actually caught, not cases where someone successfully slipped through.)

The data also surfaced some geographic and procedural gaps. Fraud attempt rates hit 48% in the Asia-Pacific region, compared to 27% in North America. Testing conditions made a big difference, too: Candidates in unproctored environments showed score jumps more than four times larger than those being actively monitored, which pretty clearly shows that proctoring works as a deterrent.

As for how CodeSignal catches all this: the company says it’s spent a decade building out its fraud-prevention infrastructure, which it’s now applied across millions of assessments. It uses a proprietary “Suspicion Score” and leak-resistant test design to flag things like plagiarism, proxy test-taking, unauthorized AI use, and identity fraud.

More AI coverage from Fast Company: 

• Harvard study shows AI stock trading rivals many picks made by fund managers
• He built a hit podcast about the Epstein files. It’s entirely AI-generated
• What if the SaaSpocalypse is a myth?
• This AI note-taking startup thinks it’s building the ‘steering wheel’ for chatbots

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.