AI Spending Doubts Meet Autonomous Weapons Without Rules

Uber burned through its entire annual AI budget before April was over. That's the kind of stat that should make any CFO reach for antacids — and it turns out Uber's own leadership is asking the uncomfortable question that follows: what exactly did we get for all that?

Uber president and COO Andrew Macdonald recently admitted the company can't draw a straight line between its surging AI tool usage and any measurable improvement in what users actually experience. The company is spending heavily on Claude Code, Anthropic's AI coding assistant, and token consumption is climbing fast. But more tokens does not apparently equal more useful features shipping to riders and drivers.

Macdonald put it plainly: the underlying metrics might be trending in a dramatic direction, but translating that into "we shipped 25 percent more useful consumer features" just isn't happening yet. He suggested clarity might come over the next few quarters or years. That's a long runway for a company already deep in the spend.

To be fair, Uber isn't exactly penny-pinching. The company dropped $3.4 billion on research and development in 2025 alone, a 9 percent jump from the year before. CEO Dara Khosrowshahi has framed the AI investment as a trade-off: spend more on AI tools, hire fewer humans. It's a calculation a lot of tech companies are quietly making right now.

But Macdonald's comments crack open something the industry has been reluctant to say out loud. The standard pitch for AI coding tools is that they supercharge developer productivity — you get the same output with less headcount, or more output with the same headcount. Uber is essentially saying that pitch is still unproven in their case, at least at the level of detail needed to justify the cost on a spreadsheet.

This matters beyond Uber. Across tech, companies have been signing massive AI contracts based on the promise of productivity gains that are genuinely difficult to measure. When a company as data-obsessed as Uber admits it can't connect spending to outcomes, it raises a fair question about how many other organizations are in the same boat but not saying so.

The replace-humans-with-tokens strategy also carries real risk. If the productivity gains don't materialize at the scale expected, companies will have reduced their human capacity without getting the AI-driven output they banked on. Macdonald's framing — that the trade becomes harder to justify without a direct line to shipped features — suggests Uber is at least being honest with itself about where the math currently stands.

For an industry that has spent two years insisting AI ROI is just around the corner, Uber's candor is either refreshing or alarming, depending on how much you've already committed to the bet.

In November 2017, a short film screened at a United Nations forum in Geneva changed the atmosphere in the room almost immediately. The video, called Slaughterbots, depicted a fictional AI-powered drone capable of selecting and killing human targets without any human pulling the trigger. The audience found it unsettling. The reason it was so unsettling is that the Pentagon was already building something like it.

That Geneva meeting happened just after the launch of Project Maven, a US Department of Defense program using AI to analyze drone surveillance footage. Google was one of its early partners. The attendees weren't watching science fiction — they were watching a slightly dramatized version of work already underway.

Nearly eight years later, fully autonomous lethal weapons haven't arrived. But the infrastructure, the intent, and the investment are all in place, and the rulebook governing any of it remains dangerously thin.

The most visible flashpoint right now involves Anthropic, the AI company behind Claude. Anthropic has staked out what it calls two hard limits for any military use of its technology: no domestic mass surveillance, and no weapons systems capable of identifying, tracking, and killing targets without a human making the final call. In a field where most AI contractors have avoided drawing any lines at all, Anthropic's position stands out.

But the broader context is easy to lose in the noise of contracts, lawsuits, and political maneuvering. The reality is that AI has been reshaping warfare for years already, in ways that don't require a fully autonomous killer robot to be deeply consequential. Targeting assistance, surveillance analysis, logistics optimization — these are live applications, not future concepts.

The governance gap is the real story here. The Convention on Certain Conventional Weapons, the primary international forum dealing with autonomous weapons, has been meeting twice a year in Geneva for years. Progress has been slow to the point of being nearly invisible. Countries that are actively developing military AI — including the United States — have little incentive to agree to restrictions that would limit their own programs.

The challenge isn't just technical or political. It's definitional. What counts as meaningful human control over a weapons system? If a human approves a target list but an AI executes strikes autonomously, is that a human in the loop or not? These questions don't have agreed answers, and the technology keeps moving while the debate continues.

What Anthropic's stand illustrates, whatever its ultimate outcome, is that private companies are now making decisions about the ethics of autonomous warfare that governments have failed to resolve. That's not a reassuring place to be. The most important rules governing AI on the battlefield shouldn't be written in a startup's terms of service.