Fake tariff news that swung stocks $2.4 trillion shows AI trading agents must confirm claims before acting

A source check on the April 7 alert would have come back clean. According to Fast Company, a system receiving the Reuters report could identify Reuters as the publisher and CNBC as the source it cited. Neither signal could establish whether the claim about the White House was true ^[13].

"Provenance isn't truth," said Nathaniel Bradley, CEO of Datavault AI ^[12]. On April 7 the claim moved from a post on X to CNBC's air to a Reuters report citing CNBC ^[2][3]. Each step added a recognisable name. CNBC repeated the claim without confirming it, and Hassett had not said what the reports claimed ^[3][5].

Agents make the job bigger. Before acting, an agent may have to judge whether a report is credible, what it means, whether it matters to its task, and whether there is enough evidence to act ^[11]. The April episode predates much of the push to give agents that kind of authority ^[8]. It shows a false claim moving through social accounts and newsrooms. It does not show what an agent would have done with it ^[3].

Citrini is the opposite case: the source was right and the label was on the page. "What follows is a scenario, not a prediction," the authors wrote ^[15]. Reuters later reported that it was one of a number of gloomy AI forecasts making the rounds while software and financial shares were under pressure ^[16]. Firecrawl cofounder Eric Ciarla put the judgment after his product's part of the job. "The reading, the weighting, and the conclusion happen in the model and the prompt," he told Fast Company. "Our part is giving them the best possible version of the page to work from." ^[18]

"Trusted sources matter, but trusting the source isn't the same as trusting the decision," said Kevin Frechette, CEO of Fairmarkit ^[19]. A team can turn that into a two-by-two for every input its agent reads. One axis is whether the claim has been confirmed by the person or body it is about. An author's scenario label counts as unconfirmed. The other axis is whether the agent's authority on that input is to recommend or to execute. Confirmed claims can go to execution. Unconfirmed claims go to a person as a recommendation, with the citation chain attached. The quadrant to close is unconfirmed-and-execute, where the April 7 alert sat.

I'd close it and accept the cost. The same rule holds the agent back on true reports until confirmation arrives. On April 7, the rally and the reversal both happened between 10:08 and 10:18 a.m. ^[4][7].