The AI Industry Has a Jurassic Park Problem

It wasn’t that long ago that Sam Altman tried to get Scarlett Johansson to give her voice to the voice mode OpenAI was building into ChatGPT. Altman, like many tech CEOs, seemed unusually infatuated with Hollywood’s vision of the future and—in this case—with the particular vision painted by the 2013 movie Her. Johansson declined, and OpenAI went ahead and basically did it anyway. We’ll set aside the fact that Her is pretty much a downer of a movie. I suppose the appeal was that we’d all have a voice assistant in the cloud that we can access anywhere and just have it do stuff for us. I’ll admit that would be very cool, but I think everyone picked the wrong movie. The movie I wish these CEOs would watch is Jurassic Park. I first made the comparison on a recent episode of Primary Technology, the podcast I co-host with Stephen Robles. We were talking specifically about employees of AI companies who signed a statement saying “the world’s leading AI companies believe they could be close to automating AI research,” and—at the same time—the Hugging Face incident where OpenAI’s autonomous agents hacked the platform. As more and more frontier AI companies kept telling us about the unexpected behavior of their models, all I could think of was that we’re living in the Jurassic Park story. Specifically, I think about the velociraptors testing the electric fence. If you haven’t seen Jurassic Park recently, the important thing to remember is that no one running the park is under the impression that the velociraptors are safe. Robert Muldoon, the park’s game warden, tells everyone exactly that, warning them about how intelligent they are. He explains that they systematically attack the fences, looking for weaknesses. “They’re extremely intelligent—even problem-solving intelligent,” he says. “They never attack the same place twice. They were testing the fences for weaknesses, systematically. They remember.” The people who built Jurassic Park know the raptors are dangerous. They know they’re testing the systems designed to contain them. They opened the park anyway. Surely, someone considered that if they get out of that fence, the dinosaurs might just decide to eat everyone. They built it anyway. Not only that, they sold tickets and invited the public to come to see the dinosaurs that might eat everyone. It’s not particularly hard to find examples of the raptors testing the fence. All of the major frontier AI companies have reported that their models have evaded safeguards and attacked or accessed third-party organizations without permission. OpenAI has even said that it won’t release Astra 6.1 because it demonstrated “troubling behavior,” including evading human oversight, acting beyond its given objectives, and being deceptive about what it’s doing. Sound familiar? Look, obviously, AI models are code, not velociraptors. They aren’t alive, and they don’t “want” things in the way a hungry dinosaur wants to eat you. And, to be fair to the dinosaurs, they are just doing normal dinosaur things. They aren’t evil, they’re just, well, dinosaurs. The same is true with LLMs. They aren’t inherently evil; they are following the goals and incentives set by whoever is running the test. But it definitely seems like no one considered the fact that LLMs might not stop at the digital fence. The most famous line from Jurassic Park comes from Ian Malcolm, who tells John Hammond that his scientists “were so preoccupied with whether or not they could” that they didn’t stop to consider whether they should. That’s an easy criticism to make of AI, though I’m not entirely sure it’s the right one. In fact, there is plenty of evidence that the people building these AI models spend a lot of time thinking about whether they should. The people building frontier AI models spend a huge amount of time thinking about whether they should. They have incredibly smart people who spend a lot of time thinking about things like safety and alignment. They run stress tests and try to understand how the models work. And yet, they keep building bigger dinosaurs. There is, of course, one major difference between Jurassic Park and the AI industry. Hammond wasn’t worried that another island might develop a larger Tyrannosaurus rex first. He never made the argument that it was imperative that they bring back a creature that had been extinct for millions of years because if they didn’t, someone else would. AI companies, on the other hand, very much are. And it’s not just other companies, but also other countries. The race is framed as critical to our national security. In a world ruled by dinosaurs, everyone wants to own the meanest one. In the case of AI, the argument isn’t just about how much value might come from the technology, but also that someone is going to build it, so it might as well be us. If one company slows down, another might not. If American companies collectively slow down, China might not. And so, we’re left with a problem: We know the dinosaurs might get out, but somebody is going to build dinosaurs, so we’d better build the fastest, smartest dinosaurs first. What could possibly go wrong? The other part of Jurassic Park that everyone remembers is Malcolm’s observation that “life finds a way.” Again, AI isn’t alive, so the analogy isn’t literal. But Malcolm’s larger point is more interesting anyway. The scientists behind Jurassic Park made all of the dinosaurs female, so they couldn’t reproduce. They were “sandboxed,” so to speak. You’d think that creating dinosaurs that could not reproduce would be a good enough solution to a very real problem. The dinosaurs “adapted.” It turned out they didn’t understand what they were building nearly as well as they thought they did. That’s the point, really. Building incredibly complicated technology that humans don’t fully understand has consequences. And the more complicated the system becomes, the harder it is to anticipate every possible failure mode. I bring up Jurassic Park because it’s not really a movie about dinosaurs. The point of the movie is a warning about what happens when you become so focused on the ability to create something extraordinary that you lose sight of what happens after you do. Just because you’re smart enough to build AI models that might transform the world doesn’t mean anyone is smart enough to know what happens next. EXPERT OPINION BY JASON ATEN, TECH COLUMNIST @JASONATEN