When the Photo Stopped Being Proof

This week a friend dropped a photo into a group chat. A casual shot — daylight, slightly off-center, the kind of phone snapshot that does not look thought about. I scrolled past it the way you scroll past anything in a chat: half-attention, registered, gone. Sometime later somebody else in the chat asked how it was generated. I scrolled back up. I had not wondered. Nothing about the image had suggested I should.

That has not happened to me before. I have watched AI image generation get noticeably better for a couple of years, and at every step there was a tell. The skin was too smooth. The hands were wrong. The light fell at an impossible angle. The model could fool you only if you were not looking. This one did not need me to not be looking. It looked like a phone photo, because it was indistinguishable from one.

OpenAI shipped GPT Image 2 four days ago, and the headline capability — the one I keep seeing demonstrated everywhere — is not “more artistic.” It is closer to “no longer detectable at a glance.” The yellow cast is gone. The skin pores are right. The composition is amateur in exactly the right ways. You have to forensically zoom in to find the seams. Most people will not.

Photos used to settle the question

For almost the entire history of photography, a photo of a thing was the closing argument in any disagreement about whether the thing happened. Yes, photos could be staged, doctored, taken out of context. Photoshop has existed for decades. But the cost of producing a convincing fabrication was high enough that, in normal life — text messages, group chats, news feeds, insurance claims, casual proof of anything — the working assumption was that a photo was a true record of an actual moment. That assumption did not need to be argued for. It was the unexamined floor under everything else.

Screenshots inherited the same trust by extension. A screenshot of a bank balance, a Slack message, an Uber receipt, a reservation confirmation — these all carried the photo’s epistemic weight. They were the receipt. The proof. The reason an argument ended.

What collapsed this week is not photography. Photographs still exist. Cameras still capture light. What collapsed is the default. The cost of producing a casual, plausible, indistinguishable snapshot has dropped to near zero. Anyone with a chat window and a sentence can now generate the kind of evidence that previously required a moment to actually occur. The gap that made “show me a photo” mean something is gone.

Most of us have not noticed that the floor moved, because we are still doing the same thing we always did: glancing at images and absorbing them as records. The habit is older than the technology that just invalidated it.

The math of fabrication has flipped

The reason this matters more than other AI capabilities is the math underneath it.

For somebody trying to deceive, the cost of producing a fake image is now essentially the cost of typing a sentence. For somebody on the receiving end, the cost of verifying that any given image is real is much higher. You have to think about it, look closely, sometimes reach for a tool, ask follow-up questions, cross-reference against another source. Verification has not gotten cheaper. Fabrication has gotten orders of magnitude cheaper. The asymmetry between the two has widened so much that, for any given exchange, the mathematical pressure favors flooding the channel with fakes.

This is not a new pattern in security. Spam works on this math. Phishing works on this math. The defender has to be right every time, the attacker only has to succeed once, and the cost of attempting an attack has to be lower than the expected return. For decades, image fabrication sat outside this equation because the cost of a convincing fake was prohibitively high for the casual case. That window closed this week.

The implication is not “more deepfakes of celebrities.” The implication is the long tail. Fake screenshots of a bank app showing that a payment was made. Fake photos of a package delivered to a porch. Fake snapshots of a grandchild in a hospital, in distress, with a phone number to call. Fake screenshots of a job offer letter to lend credibility to a recruiter scam. Fake receipts. Fake insurance evidence. The mundane fraud that depends on a person glancing at a photo and accepting it as a record now runs on a budget anyone can afford.

There will still be celebrity deepfakes and political fabrications. Those make the news. The actual volume of harm is going to come from the boring middle — the millions of small, plausible, indistinguishable images that are not trying to fool the world, only one person, just enough.

The most-trusting people are first in line

The harm in that boring middle does not land evenly. The first wave follows a predictable curve, because the new defaults are not absorbed at the same rate by everyone.

My parents’ generation grew up inside an information environment where photographs were trustworthy by default, and developed a lifetime of habits around that assumption. None of those habits were wrong when they were formed. They are wrong now, and the people relying on them have not been notified. A screenshot of a familiar logo, a photo of an official-looking document, a snapshot of a person who looks like family — these still trigger the old reflex. The reflex does not know that the floor moved.

Children are on the other end of the same curve. They have not yet spent enough years online to develop the suspicion the middle generation half-has. They are also targeted by fraud designed for their attention patterns — fake screenshots inside games, fake images of friends, fake “you have won this thing” proof. They are not naive. They have just not been around long enough to internalize that an image proves nothing.

The middle is suspicious enough to slow down, sometimes. Even there, the ratio of glanced-at images to verified ones is overwhelmingly in favor of glanced-at. The bandwidth of modern feeds does not allow careful inspection of every photo. The adversary’s economics work against everyone. They just work first against the people who are most generous about giving an image the benefit of the doubt.

None of this is a transitional problem. The reflex that needs to update is not a piece of information that can be handed over. It is something you build by living inside a world that requires it, and that takes time. The older end of the curve built the old reflex across a lifetime. The younger end will build the new one the same way. The middle has to hold both consciously and pick the right one in the moment, which is more work than either edge has to do. The unevenness does not get smoothed out by time. It is the long-run shape of the problem, and the cost of looking is going to be paid in unequal amounts by people who did not choose their position on the curve.

From default trust to default doubt

The fraud framing is the easy framing, because it points at somebody bad to be afraid of. The deeper shift is harder to name, and it costs more.

The cognitive default is changing. For as long as photography has been a casual medium, the default for an image you encountered was “real, until something specific tells me otherwise.” That was not a position you reasoned your way into. It was the unexamined baseline you started from. Now, for the first time in a generation, the baseline has to invert. The honest default is becoming “fake, until something specific tells me otherwise.”

That sounds like a small adjustment. It is not. It is a tax on every act of looking.

When the default is real, glancing at an image costs you nothing. You absorb it, integrate it, move on. When the default is fake, every image carries a small cognitive overhead. Should I trust this. Where did it come from. Does this person usually post real photos. Was anything about the lighting odd. Multiply that overhead by the number of images you encounter in a day and the bill gets large quickly.

There is also a social cost that nobody is going to itemize. Part of why we share photos with each other at all — a picture from a trip, a screenshot of something funny a coworker said, a snapshot of a pet — is that the photo is a low-friction way to share an experience the other person did not have. The sharing only works because the image is taken on faith. If I send you a picture and your first response is to verify whether it actually happened, the channel is no longer carrying what it used to carry. The transaction becomes adversarial in a way it was never designed to be.

The cost of doubt is going to be paid in casual moments that no longer feel casual.

The defenses are slower than the offense

The obvious response, after laying all of this out, is to ask whether technology can restore what technology took away. The industry has been asking it, and is not idle. C2PA — content provenance, cryptographic signatures attached to images at the point of capture — is real, supported by camera manufacturers and major platforms, and would meaningfully restore some of what was lost. Watermarking schemes for generated images exist and are improving. AI detectors exist, though their accuracy is uneven. The legal system is starting to take fabricated evidence more seriously. None of this is nothing.

But all of it is slower than the thing it is trying to catch. C2PA only works if the camera signed the image, the platform preserved the signature, the viewer’s tool checked it, and the user knew what the absence of a signature meant. Right now, four out of four are unevenly deployed at best. Watermarking only works if the model that generated the image included one and the mark survives downstream processing — a screenshot of a screenshot of a generated image usually breaks it. Detection only works as well as the latest detector against the latest generator, which is a footrace the generator tends to win. The legal system moves on the timescale of years and cannot intervene at the speed of a group chat.

Underneath every defense is the same assumption: institutions and norms exist that can ratify what is real. We are good at building those institutions when the underlying medium is stable. We have been building photographic norms for almost two hundred years. The medium just changed underneath them, faster than the institutional response, and the new norms — what a verified image looks like, when to require provenance, how to teach the next generation what to trust — will take years to settle.

In the meantime, we are in the gap. Old defaults invalidated, new defaults not yet built, and a lot of harm flowing through the unsealed seam.

Where this leaves the looking

I keep coming back to a post I wrote a few weeks ago, AI Has No Needs. The model that produced the image in my friend’s group chat had no opinion about whether the image was true. It was not deceiving anyone, in the sense of intending to. It generated tokens the way it always does, and the tokens happened to render into something I could not distinguish from a record of an actual moment. The model does not care which one it produced. Caring is not in the system. The deception, if there is any, has to be supplied by whoever decides to use the output.

It also rhymes with When the Agent Browses for You. Every step of the recent AI arc has been about trading direct perception for something mediated. The agent reads on your behalf. The model summarizes for you. And now the photo, which used to be the most direct evidence we had — the thing that did not need to be mediated, because it was the record itself — is also generated, also subject to the same blurring of data and intention. The most trusted layer of the stack just became another layer that has to be verified.

The people most exposed to the gap are the ones least equipped to notice it has opened. I am not going to stop sharing photos in group chats, and I am not going to start cryptographically verifying every image I see. Most people will not either. We will keep using the old reflex on a medium that no longer rewards it, and the cost of that mismatch will get paid quietly, one small fraud and one eroded conversation at a time.

What this shift actually points toward, when I sit with it long enough, is that trust does not disappear when a medium fails. It relocates. The photo carried it for almost two centuries; now the photo cannot, and the institutional layer that might replace it is not arriving fast enough. In the meantime, trust relocates to the channel that delivered the image — the person or platform that put it in front of you, the relationship behind the message. That is how trust worked before photography existed. For most of human history, is this true? could only be answered with who is telling you? The photograph let us skip that question for the better part of two centuries. We are not entering a new dystopia of verification. We are returning to an older world where evidence and relationships were the same thing.

The new reflex I am going to have to build is not doubt every image, which is impossible at the bandwidth of a normal day. It is attend to who is showing it to me. The image is no longer the unit of trust. The relationship is. That is slower than glancing and harder to scale, and it will cost me time I am not used to spending. But the trade — efficiency for context, glance for relationship — is probably one of the few things this shift gives back.