Reality hygiene

While I’ve found various AI cheerleaders’ propaganda about the capabilities/ontology of Large Language Models to be absurdly exaggerated, I’m a lot more worried about this kind of thing:

Is seeing still believing? Based on the evidence of the past week, it is hard to say.

Consider Exhibit A: Rauiri Robinson, an Irish filmmaker and visual effects artist in Los Angeles, posted two short A.I.-generated videos on X, a hyper-realistic action-movie sequence depicting Tom Cruise and Brad Pitt fighting on a rooftop while arguing about Jeffrey Epstein. The clips were created, Mr. Robinson explained, by feeding a two-sentence prompt into Seedance 2.0, an A.I. video-creation tool newly released by the Chinese company ByteDance.

Its convincing imitation of an actual film sparked horror and outrage in Hollywood. “I hate to say it,” Rhett Reese, a screenwriter whose credits include the “Deadpool” films, wrote on X. “It’s likely over for us.”

But consider Exhibit B: The announcement on Thursday morning by Tom Homan, Donald Trump’s border czar, that federal immigration agents would soon withdraw from Minnesota. Although Mr. Homan declared the operation a success, the decision seemed a tacit acknowledgment of the political damage inflicted by bystanders’ videos of two fatal shootings of Minneapolis residents by federal agents last month.

The videos immediately undercut the administration’s false and derogatory claims about the victims, drawing rebukes from even some Republican politicians and conservative commentators. “Escalating the rhetoric doesn’t help, and it actually loses credibility,” Ted Cruz, the Republican senator from Texas, said on his podcast in late January.

It is a paradoxical moment, in which documentary evidence is still able to land a few punches, even as new technologies threaten its credibility like never before.

“It feels deeply contradictory,” said Sam Gregory, the executive director of Witness, a human-rights organization focused on gathering video evidence.

Mr. Gregory’s organization has for years trained observers in recording video of human-rights abuses, and more recently has studied the challenges that A.I. poses to such efforts. The success of observers in documenting the tactics and behavior of federal agents in Minneapolis is “clearly an affirmation that we can still show what’s real with video,” he said.

The internet is overrun with AI slop at the moment, but people who know about this technology (not me) tell me that the ability to create AI videos that are indistinguishable from recordings of reality is advancing very quickly, and that very soon it will be basically impossible for even experts to tell the difference, let alone the general public.

This seems like a giant political problem, that is going to be extraordinarily difficult to regulate. One of many passages from 1984 that have been cited over and over again during Trump 2.0 seems particularly relevant here:

The party told you to reject the evidence of your eyes and ears. It was their final, most essential command. His heart sank as he thought of the enormous power arrayed against him, the ease with which any Party intellectual would overthrow him in debate, the subtle arguments which he would not be able to understand, much less answer. And yet he was in the right! They were wrong and he was right.

This reflects Winston Smith’s belief in a kind of straightforward empiricism, that survives until physical and psychological torture breaks him completely.

But what happens when the evidence of your eyes and ears is something that, in a world increasingly mediated by the virtual reality of the internet, is increasingly unreliable?

In this regard, Neil Postman and Jean Baudrillard, who authored their essential works in the pre-internet era, are the prophets of our time, even more so than the dour lower upper middle class old Etonian, whose wintry vision of the future has come to seem ever-more relevant to our present moment.