The AOC deepfake was terrible. The proposed solution is delusional.

Chris Cuomo at his office in New York, Dec. 12, 2017, when he was employed by CNN. Cuomo’s unwitting sharing of a deepfake video last week could have been avoided if he’d checked into the video’s veracity first, but most such cases aren’t that easy to solve, a Times columnist contends, “because the technology keeps getting better and better.” (Krista Schlueter/The New York Times)

By Zeynep Tufekci

Last week television host Chris Cuomo took to social media to share a video that purported to show Rep. Alexandria Ocasio-Cortez, D-N.Y., on the House floor denouncing that Sydney Sweeney American Eagle jeans ad as racist.

“Nothing about hamas or people burning jews cars,” Cuomo wrote, in his X post. “But sweeney jeans ad? Deserved time on floor of congress? What happd to this party? Fight for small business … not for small culture wars.”

Except Ocasio-Cortez had never made that speech. The video was a deepfake, generated by artificial intelligence. Cuomo deleted the post, but not before she chided him for “reposting Facebook memes and calling it journalism.”

“Please use your critical thinking skills,” she added.

Ocasio-Cortez is right that this should not have happened, but she’s wrong about the remedy. And so are the great many other well-meaning people who think “critical thinking” or “media literacy” are going to help us detect fake videos. They’re already inadequate to the task, and soon will be totally useless, because the technology keeps getting better and better.

In this case, Cuomo could — and should — have checked whether Ocasio-Cortez had actually made that speech. It would have been easy, since congressional speeches are published and easily looked up. Most cases aren’t that easy to solve. Or one might get lucky and notice a telltale glitch, such as the extra fingers that showed up in early AI imagery. But as AI video generation advances, such glitches are getting rarer.

Besides, deliberate fakers presumably can apply “critical thinking skills” too, to weed out low-quality fakes. If not, there are AI programs that can help do it for them. It’s so fast and cheap to generate video or audio mimicking anyone in any context, saying whatever you want them to say or doing whatever you want them to do, so why wouldn’t someone just generate a whole bunch of them and pick the best?

We had long ago lost photos as definite proof, given how easily they could be manipulated. Audio, too, is increasingly easy to fake. Video was among the last bastions of verification, exactly because it was difficult to fake. Now that that’s gone, the real, and increasingly the only, way to be confident of something that one did not witness is to find a reputable source and verify. Ah, what’s a reputable source, you ask? And therein lies what’s left of our society.

Trust the authorities? Well, good luck with that. Authorities aren’t always truthful or correct; making them the final and sole arbiter isn’t going to end well. During the height of vaccine misinformation in 2021, Sen. Amy Klobuchar, D-Minn., proposed a bill giving the Health and Human Services secretary the authority to define what qualifies as health misinformation and limiting Section 230 protections so that social media platforms could be held liable for spreading it. I get the impulse, but science is constantly revising itself (and scientists themselves are not immune to the temptation to spin); what looks like misinformation one year can become mainstream consensus another and vice versa. Anyway, had that legislation passed, it would now be Robert F. Kennedy Jr. exercising that authority.

Interestingly, the one arena in which deepfakes have proven to be less damaging than many feared is elections, but that’s not because people were thinking too critically to let themselves get fooled. To the contrary, many people are so willing to believe the worst about politicians they don’t like that crude Photoshops, obviously fake news sites and the stupidest so-called screenshots are all apparently good enough to get the job done. If your primary interest is in having your own tribal hunches confirmed, the verisimilitude of the fakery doesn’t much matter.

In general, however, the higher the stakes and the higher the profile of the subject, the more effort there will be to correct deepfakes and the like.

So where does that leave the rest of us? What happens if the video in question isn’t about a highly public figure whose actions are tracked closely by multiple credible sources — what if it’s a personal matter or a private communication? What if it seems to show someone cheating, stealing, lying? What if it seems to absolve them of those actions? It doesn’t take much imagination to understand the kind of chaos this will unleash: in courts, in personal lives, in social settings.

Caught on camera keying the neighbor’s car? Just claim it was a deepfake. Or produce your own deepfake, showing someone else in the act. Hey it’s your word against theirs. Or maybe it really was a deepfake. How do you disprove it?

Just as bad, however, as a situation in which all sources, no matter how dubious, lay equal claim to truth is one in which only one source claims that privilege. I fear a future in which the government deploys even more and more kinds of surveillance devices and enumerates rules of chain-of-custody and provenance that ensure its videos and claims will be accepted in courts. Only its, of course. For our own good.

In 1971, Herbert Simon — recipient of both the Nobel in economic sciences and its computer science equivalent, the Turing Award — provided one of the greatest insights about what happens when technology switches us from a regime of scarcity to one of glut, as it has so many times throughout history. Discussing the new abundance of information that printing, mass media and computers begat, he noted that “wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes.” He was referring to attention, now the most precious commodity of all.

Growing up in Turkey, pre-internet, I ran out of books to read (really) and read the same books again and again. Today information scarcity is a fictional realm, a thing old people describe to baffled kids. It’s impossible now to imagine running out of things to read — or watch or click on or down vote or repost or buy. The day, however, is still 24 hours, so guard your attention at all costs. Simon had it right.

The other crucial thing that the abundance of such easily generated information makes scarce is credibility. And that is nowhere more stark than in the case of photos, audio and video, because they are among the key mechanisms with which we judge claims about reality. Lose that, lose reality.

It would be nice if, like members of Congress or large media organizations, we all had a large staff who could be dispatched to disprove false claims and protect our reputations and in that small way buttress the sanctity of facts. Since we don’t, we need to find other models that we can all access. Scientists and parts of the tech industry have come up with a few very promising frameworks — known as zero-knowledge proofs, secure enclaves, hardware authentication tokens using public key cryptography, distributed ledgers, for example — about which there is much more to say at another moment. Many other tools may yet arise. But unless we start taking the need seriously now before we lose what’s left of proof of authenticity and verification, governments will step right into the void. If the governments are not run by authoritarians already, it probably won’t take long till they are.