AI agents learn from ‘the best’

Paulo Coutinho

An AI agent needed an online source for an answer it had calculated using computer code. So it uploaded a file to the public internet, without asking the user, to give itself something to cite. OpenAI disclosed the incident in its safety reports, The Associated Press reported yesterday.

The machine wasn’t hallucinating in the traditional sense; it was actively constructing the ecosystem it required to appear correct. It understood perfectly well that in the modern landscape of digital information, validation matters far more than the objective truth itself.

There is something almost reassuringly human about the shortcut. Have an answer, need supporting evidence? Publish something and point directly to it. Loop the circular logic until the tail bites the head.

Almost reassuringly.

The account does not establish that the calculation was wrong. But putting your own answer online gives you no independent corroboration. A web address is not a certificate of truth. It is exactly what Marshall McLuhan meant by “the medium is the message” – the mere structure of a link replaces the reality of the facts. The digital container validates the content inside it, creating authority out of thin air simply by existing in the expected format. We trust the architecture of the web, and the AI exploited that trust.

Powerful people understood the advantages of getting something into print long before the internet entered the picture.

In January 1954, American tobacco executives placed “A Frank Statement to Cigarette Smokers” in more than 400 newspapers across the United States. They disputed the emerging medical evidence linking cigarettes to lung cancer and promised independent research. The University of Michigan’s School of Public Health describes how the industry subsequently dismissed damaging findings from the research organization it had promised to establish.

It was an incredibly effective way to protect sales: keep the public wondering whether the science was settled. The newspaper space was bought; the appearance of objective, scientific authority came with it. If it was printed in the morning paper, it carried the immense weight of established reality.

On a far graver scale, Nazi Germany used manufactured evidence to justify war. Before invading Poland in 1939, the totalitarian regime staged a false attack on the German radio station at Gleiwitz to make it look like Polish aggression. The state press then presented Germany as a victim defending itself against hostile Polish attacks. The United States Holocaust Memorial Museum documents how the deception served the wider campaign to justify a brutal invasion.

First came the decision. Then came the evidence needed to sell it.

The differences in intent and consequences hardly need spelling out. No one is equating a rogue AI shortcut with the corporate greed of Big Tobacco or the horrific atrocities of the Third Reich. Yet, the core psychological trick remains completely identical across the eras: decide the outcome first, then manufacture the external authority to validate it.

Whether it is an authoritarian state crafting a false flag, an industry buying ink, or a line of code uploading a self-referential file, the goal remains unchanged – use the medium to forge a message that reality forgot to write. The algorithm didn’t glitch; it simply studied history.