Air Force AI Beats Human Planners by 90%, Pentagon Blacklists Anthropic, and Starlink Fails NATO's Forest Test — Military AI Just Got Real

Air Force AI Just Beat Human Planners — And It Wasn't Close

The U.S. Air Force ran its third DASH experiment — Decision Advantage Sprint for Human-Machine Teaming — this past fall. Results were published this week.

According to Breaking Defense, AI tools from roughly half a dozen companies were pitted against military professionals from the U.S., Canada, and the UK. The task: solve real battle management problems under time pressure. Planning airstrikes. Rerouting aircraft after base damage. Protecting a drifting Navy vessel.

The AI won.

Col. John Ohlund, director of the Air Force's Advanced Battle Management System Cross-Functional Team, told Breaking Defense that machine-generated recommendations were "up to 90 percent faster than traditional methods," with the best AI solutions showing 97 percent viability and tactical validity.

Human planners? They averaged about 19 minutes per course of action — with only 48 percent of those options deemed viable and tactically valid. The AI didn't just work faster. It made fewer mistakes.

Ohlund also noted: "Our team didn't observe hallucinations during the experiment." This matters given AI's known tendency to confidently generate wrong answers.

Caveats exist. Ohlund acknowledged the scenarios were deliberately designed to push humans outside their comfort zones — multi-domain problems crammed into one hour. That's a stress test, not a standard workday. Still, the results are hard to dismiss.

The Pentagon vs. Anthropic — Government Muscle vs. Private Ethics

While the Air Force was running those experiments, a separate conflict was playing out between the Pentagon and AI company Anthropic.

The Brennan Center for Justice reported this week that Anthropic asked the military to promise two things: first, that it would NOT use Claude — Anthropic's AI model — in fully autonomous weapons that identify and fire on targets without human input; second, that it would not use Claude to spy on Americans by analyzing commercially purchased location data and financial records.

The Pentagon refused. Then it blacklisted Anthropic entirely, designating the company a "supply chain risk." Anthropic has now challenged that designation, arguing it violates due process and the First Amendment.

The Brennan Center also reported that the Pentagon is already using Claude's competitor systems — including something called the Maven Smart System — to generate hundreds of targeting recommendations in Iran. The system pulls from satellites, data brokers, military drones, and social media. The Defense Department has allocated at least $75 billion to AI-driven programs since 2016, according to a Brennan Center report, with the actual total likely far higher due to classified programs.

A company said "don't use our AI to autonomously kill people or spy on Americans" and the government's response was to cut them off and label them a security risk. Whether Anthropic's conditions were reasonable or naive, the Pentagon's decision to punish a private company for setting ethical limits on its own product deserves scrutiny.

Back to Latvia: Starlink Fails in the Trees, and Soldiers Know It

Meanwhile, NATO's Crystal Arrow exercise in Latvia — which ran May 5-15 — confirmed a problem identified in previous reporting: Starlink dependency is a tactical liability.

Breaking Defense reported from the field that Latvian National Guard soldiers operating the Natrix UGV — a domestically-made ground robot tested in Ukraine — encountered severe communication degradation under forest canopy. Latvia's forests cover 50 percent of the country's territory. This is not a rare edge case. It's the primary terrain.

A Latvian National Guard soldier told Breaking Defense directly: "With recent developments we've seen it can be beneficial but also subject to disappearing suddenly." This came from a soldier on the Eastern Flank — 180 kilometers from the Russian border — expressing doubts about the primary communications link on his robot.

SpaceX did not respond to Breaking Defense's request for comment.

Canadian soldiers in a NATO reconnaissance unit reported the same problem with the American-made Raven-B aerial drone. Corporal Elana Clement told reporters: "How high and dense the tree line is messes with our equipment and signal."

The problem now extends beyond drones. Both aerial and ground robots are experiencing simultaneous communication failures across multiple allied nations, in the exact terrain where NATO expects to fight Russia.

What's Missing From Most Coverage

Most outlets covering the DASH results frame this as an exciting AI milestone. The Air Force is now openly comparing AI to humans in lethal decision-making, based on a single controlled experiment with stressed operators facing unfamiliar scenarios. The implications of that benchmark deserve more examination.

The Pentagon-Anthropic story is being covered mostly as a business dispute. The U.S. government is establishing a precedent that it can punish private companies for declining to remove safety limits on their own products. The implications extend far beyond defense contracting.

And the Latvia communications failures are being treated as a technical footnote. NATO is building an Eastern Flank deterrence strategy around technology that demonstrably stops working in half the country's terrain.

The Accountability Question

AI is being woven into targeting, battle planning, and intelligence analysis at scale — $75 billion and climbing — with limited public oversight and a government that demonstrated it will blacklist anyone who pushes back.

The technology works, sometimes impressively. It also fails in the woods. Washington has not clearly answered who is accountable when the algorithm is wrong and people die because of it.