Two of today's stories sit on either side of the same question, and it is the one the industry keeps trying to skip: not what AI can do, but how far we should trust it — and who gets to decide.

Can you trust what it tells you?

Start with the part most people actually touch. We looked at an NPR–NewsGuard test that fed six leading chatbots — ChatGPT, Gemini, Copilot, Meta AI, Grok and Claude — thirty questions built from real Russian, Chinese and Iranian falsehoods, and set them against the big search engines. The encouraging half: the chatbots debunked the lies about three-quarters of the time, and did it better than a Google or Bing summary. The catch is the other quarter. A person scoring three in four would be doing well; a machine that millions treat as an oracle, handing back a wrong answer one time in four with total fluency, is a different kind of problem. And the route matters — one model repeated a false claim for several paragraphs before hedging, which is not much of a correction if nobody scrolls that far.

The honest read is not "the machines have learned to spot lies." It is that a chatbot is a decent first filter and a poor last word — and the confident wrong quarter is exactly the part that travels.

Who gets trusted with it?

The second story runs the other way: not whether we can trust the AI, but which AI an institution will let through the door. The Pentagon's GenAI.mil platform added ChatGPT and Grok this week, putting them in front of three million defence personnel alongside Google's Gemini. One frontier lab is absent — Anthropic, whose Claude was purged after it asked for contractual limits the Pentagon would not accept: no autonomous lethal targeting, no mass surveillance of Americans. A federal judge ruled the designation behind that exclusion unconstitutional only last week. The platform onboarded the rivals regardless.

The thread

Put the two together and a shape appears. In the first, the danger is trusting an AI too readily. In the second, the fight is over the terms on which a powerful institution adopts one at all — and the lab that pushed hardest on the limits is, for now, the one left outside.

The capability question is increasingly settled: these systems are good enough that three million people will use them for real work and hundreds of millions will ask them what is true. What is not settled is everything both stories circle — whether you can trust the answer, and who sets the rules for where the technology goes next. On a quiet news day, that is the pairing worth holding onto: the frontier is shifting from "can it?" to "should we, and on whose terms?" — and the second question is far harder to benchmark.