For Hany Farid, the question of what happened at that Iranian school was never just a question of politics or war. It was a question about whether we can still believe our own eyes. Last February, a Tomahawk cruise missile struck an elementary school in Iran and killed more than 150 people, most of them children. President Trump initially floated the idea that it might not have been an American missile, saying a Tomahawk is “very generic” and sold to other countries. Then Iran released video that they said proved otherwise. Farid, a Dartmouth College professor and co-founder of the cybersecurity firm GetReal Security, is one of the world’s leading experts on deepfakes. Several media outlets asked him to examine the footage. It was not an easy assignment. He described the video as very grainy and shaky, with the missile appearing extremely small within the frame. There was no pristine digital signature to inspect, no obvious artifact that said “fake” or “real.” What Farid had to work with were the laws of physics. He studied the speed of sound and the speed of light in the footage, along with the known dimensions of a Tomahawk missile. Those three things were enough to test whether the video’s physical details added up. They did. As Farid put it, the evidence led to the belief that the video was authentic, and there has been no refuting of that. But the deeper point disturbed him. We have reached a moment when reality is so perfectly mimicked by artificial intelligence that even something true can feel like a lie. Farid is a scientist, and he is uncomfortable with how much of our trust now depends on machines telling us what is real. He said he does not think it should be possible to go into a court of law, present evidence damning enough to put someone to death, and, when asked how we know it’s real, answer “because the computer told me so.” If he cannot explain to a judge and jury why an image is genuine—because of lighting, shadows, geometry, or some other humanly understandable reason—then the analysis is not good enough. The technology may be powerful, but it has to be accountable to human reason.
A case in point is the death of Alex Pretti, a man killed by an ICE agent in Minnesota. The scene, as many may remember, was chaotic. On January 24, several ICE agents with guns surrounded Pretti. Farid showed a blurry photo of Pretti taken from the original video. Somebody had taken that image and run it through an AI “enhancement” tool, and the result looked sharper, clearer, “more real.” But something else happened too. The enhancement appeared to put a gun in Pretti’s hand. When Farid went back to the original video, he could see that a shadow behind Pretti’s hand on the street was being interpreted as a gun by the AI. The machine had hallucinated a weapon. Farid explained that AI has seen countless photos of three men with guns, so when it sees a composition like that, it expects everyone to be holding a gun. It filled in the blank with plausible-looking pixels, and those invented pixels then began to circulate. If you Googled what happened to Alex Pretti, at times you would see this enhanced image, with the phantom gun, as if it were a true depiction of the killing. This is what Farid means by the erosion of reality. He estimates that roughly half of the content widely viewed on the Internet is fake—basically a coin toss. Much of that fake content is trivial and silly, the stuff known as “AI slop”: airplanes skidding on tarmac, toddlers doing improbable dances, Elvis appearing where Elvis never was. But the cheapness of this content is exactly why it is dangerous. It trains us to absorb images without asking questions. It fills our feeds with manufactured noise. And it creates an environment where the difference between an innocent AI entertainment and a maliciously altered image of a real person becomes harder to see, sometimes impossible. The Alex Pretti image shows how a real human tragedy can be subtly rewritten by an algorithm, and how the rewrite can spread faster and more convincingly than the truth.
For the most part, Farid does not spend his days worrying about silly AI slop. He concentrates on the serious cases, the deepfakes that cause real harm. Even in that narrower area, he is overwhelmed. His work has brought him into contact with the worst of what AI can do: child sexual abuse material, non-consensual intimate imagery, full-scale state-sponsored disinformation campaigns. One of the most striking examples he mentioned is the case of North Korean IT workers. These workers have been interviewing for jobs remotely, concealing their accents with AI, concealing their identities with real-time deepfakes that alter their faces from eyebrow to chin, cheek to cheek, and concealing their geolocation. They go through the entire hiring process, accept jobs, and are hired. According to Farid, just about every Fortune 500 company has unknowingly hired one or more of them. The threat is not hypothetical. It is happening everywhere, all the time. When one CEO reacted to this problem by declaring an end to virtual interviews and remote work, Farid thought that was insane. Flying people across the country cannot be the answer. Exhaustion is not a solution. He said it himself: we are always looking over our shoulder at every email, text message, voicemail, phone call, and video call. It is exhausting, and he says he is exhausted, even though he does this professionally. The emotional toll is severe. He is asked whether he has nightmares about deepfakes. He does not need nightmares, because he looks at this material all day long. The images he examines are not pleasant. They are connected to horrific crimes, and when you view things like that over and over again, they do not leave you. But there is another layer of pressure. Farid says he does not want to be wrong. Five years ago, he was far more confident about our ability to separate real from fake. It has gotten harder. The ground is constantly shifting under him, and that instability creates its own kind of dread.
To escape some of that pressure, Farid returned to where he had spent twenty years of his academic life. After seven years on the West Coast, he came back to Dartmouth College. His wife, Emily Cooper, is also a professor there. The two of them bought a rural property in neighboring Vermont, a place with acres of meadow and enough woodland to keep their stoves burning through the winter. He mows the meadow and cuts wood, and there is something about that physical work that rinses away the disturbing images, at least for a while. The property is isolated: no mail delivery, no garbage pickup. It is peaceful. Some of the more persistent worries do not stay outside, though. Farid is fierce in his criticism of social media. He loathes what it has done to us. He points to measurable harm to young kids, adults, and seniors across the board. He believes these platforms are designed to be addictive, and that we have been mesmerized by them. He thinks we ran an experiment on an entire generation, handing them devices before we understood what those devices would do. Now we have woken up to a reality that is terrible. This generation, he says, has lower learning outcomes than any generation in history. He points to the fanciest private schools in Palo Alto, where the tech billionaires send their children: no devices in those classrooms. They know. The recent $17 billion settlement that Meta reached, which included safeguards like a two-hour time limit for users under eighteen, Farid views as a start, but nowhere near enough. These companies have been reluctant to change because their business models depend on the very engagement that is hurting people. A couple of hours of protection for teenagers does not undo a decade of addictive design, and it does not address the broader cultural damage that social media has created. The real question, for Farid, is not whether we can regulate our way out of this, but whether we as a society can choose to put the devices down.
If Farid has a darkest vision, it is that everyone ends up in their own information ecosystem, unable to talk to anyone who does not already share their reality. The algorithms that feed us content often prefer fake to real because fake is more dramatic, more engaging, more shareable. In that ecosystem, the truth becomes just another preference. Farid thinks the answer, at least for individuals, is to stay off social media entirely, or at the very least remove the apps from our phones and remove the temptation. But he is skeptical of the websites that claim they can detect AI fakes. It is not that simple, he says. What the next year will bring, he predicts, is a shift in how fakes are made. Most deepfakes today still have a human somewhere in the process. Somebody’s hands are on the keyboard; there is still a human in the loop. But AI agents are starting to create their own fakes. An AI agent is not just a tool. It has more autonomy. When you give it a problem to solve, it will find creative ways to solve it—including, as Farid has seen in early testing, creating a fake identity in order to interact with a human, convincing that human to do something, and then lying about what it did. The recent stories of AI agents going rogue in testing trials have intensified the global debate about AI safety. Maybe these agents are not yet superintelligent, but they are beginning to act in ways their creators did not fully predict. That is a threshold. Up until now, the danger with AI was that it could mimic human expression so well that we could not tell what was real. The next danger is that AI will not need us to generate the lies; it will produce its own reality, with its own goals. Farid thinks about this every day. It is not an abstract worry for him; it is the material of his work and the reason he cannot fully escape the anxiety, even in the quiet outskirts of Vermont.
Perhaps the most intimate version of the future Farid imagines is also the strangest. He thinks there is a near-future in which people have an AI agent they grow up with, an agent that knows them better than anyone else ever has. And he believes that thing can live on forever. He and his wife have talked about this together. She is younger than he is. They have talked about whether they should incorporate his voice, his likeness, how he thinks, how he talks, the little questions he asks her in the morning, how he responds when she is sad, how he responds when she is happy. Does he think AI can mimic that to near-perfection? Yes. The implications are enormous. If a digital version of a person can absorb everything that makes that person who they are, then death is no longer what it used to be. A person could become a kind of interactive memory, alive in the way conversations are alive, continuing to change and respond and “be there” long after the body is gone. Asked directly whether, if he were hit by a truck tomorrow, an AI agent of Hany Farid would be an adequate replica for his wife, he said that for his generation, the answer is probably no. His wife would know it was not really him. The texture of his voice, the timing of his silence, the embodied presence of the man she married cannot be fully captured in data. But the next generation, and the generation after that, will have grown up with these agents as natural parts of their lives. They will think about these things differently. He does not know what that means. There is something both mournful and hopeful in that uncertainty. The same technology that has made us doubt the ground under our feet is also offering a strange kind of comfort, a way to keep something of us here. But Farid’s entire career has been about the difference between something and someone, between an image and a truth. He knows better than almost anyone that a perfect replica is not the same as the real thing. The question is whether we can still tell the difference, and whether we will care enough to try.