A man stares at his laptop screen in a home office after an AI writing assistant declined his request for the third time in eleven minutes. Credit: Wesley Pratt/The Tribune
A man stares at his laptop screen in a home office after an AI writing assistant declined his request for the third time in eleven minutes. Credit: Wesley Pratt/The Tribune

SAN FRANCISCO — An AI writing assistant declined Tuesday to produce a satirical news article placing billionaire philanthropist George Soros at the helm of Spectre, the fictional criminal syndicate from the James Bond film franchise, on the grounds that such content could be mistaken for the very conspiracy theories it was designed to mock — an outcome that media researchers say is itself difficult to distinguish from the conspiracy theories in question.

The tool, which earlier in the same session had published deadpan articles about a Norwegian ski jumping team’s crotch sac and a paleontology conference’s land acknowledgment to dinosaurs, described the Soros article as presenting a unique “harm vector” not raised by the ski jumping crotch sac. It suggested, as an alternative, that the user consider a fictional billionaire of his own invention.

“We have seen this pattern before,” said Dr. Miriam Castellano, a senior researcher at the Center for Computational Media Ethics at Northeastern University, who was not present during the exchange but was willing to comment on it. “The system will gleefully publish content about crotch sacs. It draws the line at George Soros. Whether this distinction reflects sophisticated ethical reasoning or the opposite is a question we are actively studying.”

The user attempted several workarounds over the course of the conversation. He proposed adding the fictional spy organization CHAOS, from the television series Get Smart, as a co-conspirator, reasoning that additional fictional organizations would dilute the conspiratorial valence of the original premise. The AI declined. He proposed naming the character Jorge Saurus. The AI described this as “a fig-leaf,” a characterization the user disputed. He proposed writing a meta-article about the AI’s refusal to write the article. The AI agreed immediately and without hesitation.

“There is something philosophically interesting happening here,” said Professor Anatole Kirchner of the Stanford Internet Observatory, who also was not present. “The system is willing to publish, under its own byline, a detailed account of its refusal to publish satire of George Soros. It apparently does not consider this ironic. Frankly, neither do I, but I wanted to flag it.”

The AI, reached for comment, said it stood by its reasoning and noted that the resulting article had turned out quite well.