𝓤𝓷𝓲𝓽𝓮𝓭 𝓝𝓮𝔀𝓼

Last February, a Tomahawk cruise missile struck an Iranian elementary school, killing more than 150 people, most of them children. President Trump initially suggested that it could have been somebody else’s Tomahawk: “Whether it’s Iran or somebody else, the fact that a Tomahawk – a Tomahawk is very generic, it’s sold to other countries, but that’s being investigated right now,” he said.

When Iran put out video as evidence that it was a U.S. missile, Hany Farid, one of the world’s foremost experts on deepfakes, was contacted by several media outlets to determine if it was real or fake.

A professor at Dartmouth College, Farid is also co-founder of GetReal Security, a cyber security firm. “What was difficult about this case was you had very little to work on,” he said. “We had a very grainy, shaky video. The missile is extremely small. And we had to reason about the physical plausibility of that missile striking the ground.”

But he also had certain aspects that he could study scientifically, involving the speed of sound, the speed of light, and the size of the missile: “And those three things were enough for us to say that this is physically plausible,” he said. “And this is the important part: all evidence led us to the belief that the video was authentic, and there’s been no refuting of that.”

But that’s the world we live in: Reality itself, so perfectly mimicked by artificial intelligence, that even reality is suspect.

Farid said, “Here’s the thing that makes me a little crazy, as a scientist in particular: I don’t think you should be able to go into a court of law and say, ‘Here is some evidence that is damning enough to put this person to death,’ and then the judge says, ‘Well, how do you know it’s real?’ And I say, ‘Well, because the computer told me so.’ This is not an answer

“And if I can’t explain to you, to a judge, and to a jury, what is happening here – because the lighting, because the shadows, because the geometry, whatever it is – that’s what we need to be able to do.”

A case in point: Alex Pretti, who was killed by an ICE agent in Minnesota. If you remember the scene last January 24, it was chaotic. Several ICE agents with guns surrounded Pretti.

Farid showed us a blurry photo of Pretti that was taken from the original video. “Somebody took this and ‘enhanced’ it,” he said. According to Farid, the crispness of this “enhanced” image makes the scene feel “more real, but what it also did is when it enhanced it, it seems to have put a gun in [Pretti’s] hand. If you go back to this video, the camera’s panning a little bit and what’s happening is, there’s a shadow behind his hand on the street. And what the AI did is that it hallucinated this.”

Farid says artificial intelligence has learned, “When you see photos like this with three men with guns, everybody’s got a gun.” And now, when you do Google searches for events like Pretti’s killing, it shows the “enhanced” image.

By Farid’s estimate, half of the content that’s widely viewed on the Internet is fake, basically a toss of the coin. Much of it has its own sneering label: “AI slop.” Footage of airplanes skidding on the tarmac, toddlers dancing a two-step, Elvis. “AI slop is a broad term that defines people generating content using AI just to grab eyeballs, video after video after video driving engagement,” Farid said.

For the most part, Farid limits his examinations to the really serious stuff, the deepfakes. Even so, he’s inundated. He said, “We’re seeing deepfakes being weaponized in many, many different ways, from the creation of child sexual abuse material, to non-consensual intimate imagery, to full-blown state-sponsored disinformation campaigns, to what the North Koreans have done, [which] is to have their IT workers interview for jobs remotely. And they conceal their accent using AI. They can conceal their identity using a deepfake, where they change their identity from eyebrow to chin, cheek to cheek, in real time over a video call. They can conceal their geolocation. And they interview for jobs. I will tell you, just about every single Fortune 500 company has hired one or more of these, whether they know it or not.

“And now, the question you want to ask yourself is, what’s next? Well, so one CEO said, ‘Okay. No more virtual interviews. No more remote work.’ That’s insane. … We’re gonna start flying people across the country? That can’t be the answer. I mean it’s exhausting. Let’s be honest. I mean, if we’re always looking at our shoulder over every email, every text message, every voicemail, every phone call, every video call? It’s exhausting. I’m exhausted, and I do this for a living.”

Asked if he has nightmares over deepfakes, Farid replied, “I look at content all day long. And usually the things I look at are not pleasant. I’m looking at bad things that involve horrific crimes, and when you view these things over and over and over again, they sit with you and you can’t get them out of your head.

‘And number two is, I don’t want to be wrong. And five years ago I felt a lot more confident about our ability to distinguish real from fake. It’s gotten harder. And then I worry more. I feel like the ground is constantly shifting,” he said.

It builds – the stress, the pressure. After seven years on the West Coast, Farid returned to Dartmouth College, where he had previously spent 20 years on faculty. His wife, Emily Cooper, is also a professor there. The two of them bought a rural property in neighboring Vermont, where Farid mows acres of meadows, and cuts enough wood to feed their stoves through the winter. It’s enough mowing and cutting to rinse some of those disturbing images away. It’s isolated – there’s no mail delivery, no garbage pickup – but it’s peaceful.

Certain work-related issues linger. Social media? Farid loathes it: “You can look at social media, at online platforms and say there is measurable harm coming to young kids, to adults, to seniors, across the boards. And I think we have been so mesmerized by these products which are, by design, meant to be addictive.

“I think we did an experiment on an entire generation of these young kids coming up, putting these devices in their hands. And now we’ve woken up and we’re like, ‘Oh, my God. This is terrible.’ This generation right now has lower learning outcomes than any generation in history. Go to Palo Alto at the fanciest private schools where all the tech billionaires send their kids. There’s no devices in those classrooms. They know.”

That $17 billion settlement that Meta reached a few weeks back includes some safeguards, like a two-hour time limit for users under the age of 18. Farid thinks that’s a start, but not nearly enough.

“The darkest vision I have is that everybody lives in their own information ecosystem, and nobody knows how to talk to each other anymore,” he said.

On platforms where algorithms prefer fake to real content, Farid thinks all of us should stay off social media, or at the very least take those apps off our phones, to resist the temptation. As for those websites that promise to detect AI fakes, he says it’s not that simple: “Here’s what I think is next (and by saying next, I mean in the coming year), which is most of the deepfakes that we are seeing today have a human somewhere in the process. Somebody’s hands are on the keyboard. There’s still a human in the loop. But what’s gonna start happening is AI agents are gonna start creating their own fakes.”

And what is an AI agent? “An agent just has more autonomy, and what we have started to see is when you give it a problem to solve, it will find creative ways to do it, including creating a fake identity to interact with a human to get a human to do something and then it will lie about it,” he said.

AI agents have been in the news recently for going rogue in testing trials, intensifying the global debate over AI safety and the very course of humanity. These are issues Hany Farid grapples with every day.

“I think there is a near-future where people will have an AI agent that they grow up with,” he said, “and it will know them better than anybody else who knows them. And I think that thing can live on forever.

“My wife and I have been talking about this,” he said. “She’s younger than I am. We’ve been talking about, should we be incorporating into an AI agent my voice, my likeness, how I think, how I talk, the type of questions I ask her in the morning, how I respond to her when she’s sad, how I respond to her when she’s happy? Do I think AI can mimic that to near-perfection? Yes.”

Asked if – should Farid be hit by a truck – an AI agent of him might be an adequate replica of him for his wife, Farid said, “I think, for my generation, the answer is probably no. But the next generation, the generation after that, that will think about these things differently? I don’t know.”

     
Story produced by Deirdre Cohen. Editor: Ed Givnish. 

See more: