When AI-Generated Images Are Hard to Tell From Real Ones, We're Paying the Price for “Likeness”

There's a post I keep thinking about. Someone on Reddit complained that he was looking for dinner on DoorDash / Grubhub, went through a dozen or so restaurants, and almost all the dish photos on the menus were AI-generated—and “it made me feel sick to my stomach, so I just closed the app.”

The top reply wasn't bashing AI; it made a very practical point: a phone camera is more than good enough for photographing food, and using AI images is clearly a bad business decision. Someone else piled on: if you compare the photos on a fast-food chain's official website with the actual food, you'll be disappointed just the same.

The problem isn't that AI images are “fake”; it's that they're fake in the same way every time—that over-sharpened, greasy, gleaming texture in which every fry is impossibly perfect. The more you see it, the more it instinctively makes you distrust it. More subtly, this texture is becoming a universal “filter,” spreading from menus to ads, to product pages, to every feed we scroll through each day.

People Can't Really Tell the Difference

Two weeks ago, a little game made it to the front page of Hacker News called “Can You Tell Which Images Are AI-Generated?” It got 112 points and 85 comments. The game is simple: you're shown a series of images, 10 seconds each, and you decide whether they're real or fake.

The results were a bit sobering. Some commenters said that on their phones they were “basically guessing”; others came up with a simple but effective strategy—ask yourself: What reason would a person have to take this photo?

The example he gave was clever: a photo of a nightstand with a set of keys and a receipt on it, cleanly composed, flawless, with no visible AI artifacts. But think about it—who would deliberately photograph that? It's exactly this kind of “no reason to exist” photo that gives away its origin. Another highly upvoted comment was even easier: “Did I take this photo myself? Yes → not AI; no → AI.”

Interestingly, several people mentioned that the 10-second time limit was what really made the game hard—“most people looking at an image on social media won't spend more than 10 seconds on it at all.” In other words, in a real feed, the time we spend judging whether an image is real may be even less than what the game gives us.

A few “giveaways” that the comment section generally agreed on: lighting so perfect there are no shadows, objects arranged too neatly with no traces of real life, details that blur when you zoom in, and the most common one—there's no reason this scene would ever have been photographed.

It's Already Everywhere

AI-generated images aren't just showing up on menus.

There's a much-discussed phenomenon overseas: grandparents excitedly use AI to generate “photos of their grandchildren” and send them to the family group chat, while young parents are quite put off—because those photos use the child's face but place them in scenes that never happened. The disagreement between the two sides isn't about technology; it's about “who gets to decide a child's image.” It sounds like a family squabble, but it's actually another version of the same problem: an image carries not just a visual, but the default premise that “this really happened.”

Ads, e-commerce, second-hand homes, job postings… anywhere that needs “a good-looking image” is being filled with AI images. Efficiency is indeed high, but the cost is accumulating: when every image is flawless, people's default trust in images is being eroded bit by bit. Trust is expensive—it takes years to build and only a few images to wear away.

The Watermark Debate Is Fierce

Last month, Alibaba's Qwen Image 2.1 went viral overseas for being open source, having 7 billion parameters, and being able to run locally. The tech world cheered, but one sobering comment also surfaced in the discussion: “With that much capability, it still doesn't have a watermark—that's a little concerning.”

This is exactly the most awkward part of the current moment: the very features that make an open-source model useful—free, locally deployable, not passing through anyone's server—also make forgery easier. Watermarks and provenance aren't technical showmanship; they're the only somewhat reliable anchor in this “authenticity inflation.”

What We Should Care About

I don't think AI images are inherently a problem. They make creating images faster and cheaper, which is a good thing. Some people use them for e-commerce images, scientific illustrations, personal portraits—the efficiency gains are real.

What's truly worth caring about is default honesty. Whether an image was photographed or generated shouldn't require users to play a 10-second detective game to figure out. Whether the dish photo on a menu is the actual dish, whether the person in an ad is real, whether a child's photo has been altered by AI—the answers to these questions should be right next to the image, not hidden in the details for people to guess.

Technology has already perfected “likeness.” What needs to be added next is “truth.” And for people who make and use images, taking “Where did this image come from?” seriously earlier may matter more than learning a few more prompts.