After a long day, there’s a moment when you’re scrolling through a food delivery app and the photos start to feel almost suspiciously good. The pad thai shimmers. No kitchen pass ever really creates the angle from which the noodles receive light. The cilantro is perfectly positioned in a green crown. Because that’s what the picture is for, you order it anyhow, and forty minutes later a container containing a dish that is both identical and slightly different arrives. Photography used to play a role in the space between those two things: the meal and the image. The question now is whether the picture was ever connected to a meal at all.
Forkable, a San Francisco catering platform, replaced real restaurant photos on its website with AI-generated ones towards the end of last year, and it doesn’t seem that the restaurants were informed. It wasn’t until the business sent an email after the fact that the owners learned about it. It’s a little tale, the kind that appears and vanishes within a week. However, it stays with me because it’s such a clear example of where things have strayed. The food was authentic. It was a real kitchen. The invention was the image, the one thing intended to accurately depict everything.
This science isn’t as comforting as you might think. Oxford researchers conducted two studies in which participants were asked to select AI-generated food images from a lineup. The participants performed remarkably well, particularly when it came to ultra-processed foods. This is the unsettling part, though. People frequently preferred the AI images when they weren’t disclosed. Even though they could tell something was artificial, they still preferred it. Reading that gives me the impression that our eyes aren’t the real issue. We have a taste for it. In private, we’ve come to favor the nonexistent version.
The world of food is merely the tip of the iceberg of something much more difficult. At the end of September, OpenAI released Sora 2, which used a sentence of text to create a photorealistic video. It quickly became the most downloaded app in the photo and video section of the iOS App Store. Sam Altman singing in a restroom, Pikachu storming Normandy’s beaches, and copyrighted cartoon characters straying into unlicensed scenes were all part of the early clips, which were a fever dream. For about a week, it was humorous. Subsequently, Zelda Williams, the daughter of Robin Williams, was forced to publicly request that strangers cease sending her AI-generated videos of her deceased father. The joke quickly curdled.
To be fair, OpenAI has built-in protections. Each Sora video has an invisible provenance metadata and a visible, moving watermark designed to identify it as artificial. The problem is that metadata is removed as soon as a clip is screen-recorded and reposted, and watermarks bounce around the corner of a frame and are cropped. The credentials are present on every video, but there is hardly any visible labeling on sites like Instagram and TikTok, according to a journalist covering the rollout. Technically, the honesty is there, but it’s almost undetectable. Building public trust around that is a challenging task.
It turns out that we are very bad at this. The goal of a study that was published in the Communications of the ACM last fall was to assess people’s ability to identify synthetic media in real-world scenarios, such as when you might actually come across it, half-distracted, thumbing through a feed, rather than in a lab with endless time and a hint that half the images are fake. About 51% of the time, people correctly distinguished AI-generated content from human-made content. A toss of the coin, roughly. Human faces made detection worse, and the single-image posts that dominate social media feeds made detection even worse. Additionally, the researchers found something that should make everyone stop: people constantly overestimate their ability to recognize fakes. The risky aspect is the confidence.
A technical solution, such as stronger watermarks, more verification layers, or better detectors, is alluring. A group at Berkeley’s management review challenged that intuition, claiming that this is a trust issue disguised as a technological one. The ability of artificial intelligence to convincingly create reality is demonstrated by a new OpenAI system, and no detector can completely address the intuitive belief that an image represents an actual event.
There’s a detail in the restaurant reporting that keeps coming to mind. In December of last year, DoorDash permanently banned a driver for fabricating proof of delivery using AI-generated photos. A driver stole a single picture of a single bag that was dropped off. That’s the current texture of it: small domestic fabrications, the little lies that lubricate daily transactions, rather than some massive deepfake conspiracy. It’s odd to have to write that DoorDash and Uber Eats now specifically request that users submit real photos rather than AI-generated ones. A request to be honest, please.
Over all of this is a helpful old notion. Decades ago, the French theorist Jean Baudrillard wrote about the simulacrum, which is the map that takes the place of the territory and the copy that no longer has an original. He would have recognized the AI burger right away. It’s difficult to ignore the fact that we’re not moving toward a world of blatant fakes that we can ignore. In the future, authenticity will become something you have to actively claim rather than something you just have, and the synthetic version will be more appealing than the real.
Whether any of this settles is still up in the air. Perhaps a generation learns to read watermarks the same way we learned to recognize Photoshop, and ten years from now, the fear will seem silly. Perhaps provenance requirements and disclosure laws develop into something that truly travels with a file. Or perhaps we simply make a downward adjustment, ordering the pad thai while surreptitiously presuming the photo was fake and the dinner will be okay anyhow, trusting less, and checking more. Right now, that last one seems to be the most accurate. Not quite a crisis. The act of simply looking at something and believing it is more of a gradual deterioration of something we once took for granted.
