For as long as humans have faced decisions they could not verify, they have built ways to act anyway. The Azande of South Sudan consulted a poison oracle; the Greeks climbed to Delphi. Anthropologists like E.E. Evans-Pritchard showed these were not the superstitions we picture. The oracle was not a truth-machine. It was a sanctioned way to decide under uncertainty and to place responsibility somewhere a person could live with — the oracle chose, so no single man carried the full weight of the choice. And everyone knew what it was. An oracle never claimed to have checked the facts.
Artificial intelligence is the newest technology for acting under uncertainty. But it is the first that presents itself as verification — a checker, a truth-machine, a producer of knowledge. That single difference is the whole story.
Consider where it lands hardest first — the place I know best. I spent years in and around investigative journalism, and the newsroom is where I watch this arrive ahead of everywhere else.
An investigative reporter today can receive two thousand pages from a records request, with four days to file and a week already full of interviews. Reading it all by hand, in time, is impossible — and she has no junior to split it with and no data desk to query it. So she feeds the documents to an AI workflow. It surfaces a pattern: three complaints about one facility, each followed by an inspection marked "no violation," and a memo, buried near page 1,400, flagging the site as high-risk months before someone died. This is not a toy example at scale. The Pandora Papers ran to 11.9 million records, only 4% of them structured — an investigation no team could have done by hand.
The AI did not write the story. It pointed. She read the memo herself, cross-checked the dates against the complaint logs, and walked into her interview able to ask the one question that mattered. AI was necessary. Whether to use it is already a settled question; she is using it.
The harder question is what she says about it.
Here journalism reveals a trap the rest of us will soon feel. Audiences say they want AI use disclosed. Yet research with newsrooms has found that when an outlet discloses AI use, trust tends to fall, not rise. The profession's own ethic points one way regardless: as one investigative data editor recently argued, journalists cannot demand transparency from the powerful while concealing their own tools, and the real test is whether we can work with AI without outsourcing our skepticism. Disclose — it is not something to be ashamed of. And yet the reporter who follows that ethic pays in the one currency she cannot afford to spend: the reader's trust.
Why should telling the truth cost her? Because the public's image of journalism — a lone human, unaided, digging the truth out of the dark — has not caught up to its reality, which is now human-and-machine. When she discloses, she is not punished for doing something wrong. She is punished for breaking a picture that is already out of date.
Underneath that is the deeper thing AI does to all of us. What people have always wanted is not certainty but verification — to know how we know, and who stands behind it. AI's danger is that it sells the feeling of verification while quietly detaching it from the act. It returns a fluent, confident answer that feels checked and is not. Where checking is possible, that tempts us to skip it. Where checking is impossible — will this venture succeed, is this the right call — it manufactures a certainty that cannot exist. The oracle never pretended. This one does.
Journalism feels this first because, in journalism, the credibility of the human is the product. A bank can use AI invisibly as long as the numbers come out right; a reporter cannot, because "how we know it" is the very thing the reader is buying. So the verification question turns existential in the newsroom before it becomes obvious anywhere else. That is what makes journalism a canary, not an exception.
And the signal is already spreading. Consulting — an industry that sold hours of expert output — is quietly repricing itself. BCG now structures around three-quarters of its largest AI projects around outcomes rather than billed time, because once AI makes the output cheap, what a client is actually paying for is someone willing to be accountable for the result. McKinsey reports a similar move. Strip it down and it is the same shift the newsroom is making: as the output commoditizes, the scarce and paid-for thing becomes verification, and a human willing to stand behind it.
The work ahead is not to use less AI. It is to keep one line visible — between an answer that feels checked and one that has been; between a decision-technology that knows it is one and a machine that claims to be the truth. The reporter still owns the story. The consultant still owns the outcome. The oracle, at least, was honest about what it was. Our task is to stay honest about what this one is.