The New Shapes of Fake

For most of the AI image era, "fake" had a face. It was the smooth, plasticky portrait with one too many fingers, the magazine-cover lighting on a face that didn’t quite belong to a person, the supernaturally clean skin and oversaturated eyes. Detection was almost a parlor game: count the fingers, look at the teeth, check the hands.

Those days are gone. The 2026 generation of image models — GPT Image 1.5, Midjourney v7, FLUX.1 Kontext, Gemini 3 Pro Image, Sora’s image-only outputs — has solved the obvious tells. Hands look like hands. Text looks like text. Eyes are no longer made of glass. The category that used to be a coin flip with a few easy clues has bifurcated into a dozen distinct subcategories of fake, each with its own aesthetic, its own purpose, and its own detection problem.

This is a field guide. It is not exhaustive — the landscape is moving too fast for exhaustive — but it captures the major shapes the new fakes have taken. Some of these will be familiar. Several of them, until recently, didn’t exist as categories at all.

I. The Aesthetics of Real

The most important shift in AI image generation in the last twelve months has nothing to do with model architecture. It has to do with prompts. Specifically: the discovery that you can ask for imperfection.

Early AI images failed by being too perfect. Studio lighting. Magazine compositions. Pores so smooth they read as digital. Detection systems and human reviewers learned to flag the polish itself as a signal — if it looks like a stock photo of a person who has never existed, it probably is.

The new prompt vocabulary inverts that instinct. "Shot on a disposable camera, slightly overexposed, motion blur." "iPhone 6 photo, indoor low light, JPEG compression visible." "Grainy 35mm film, scanned, dust on the negative." "Security camera footage, 480p, fluorescent lighting, slight fish-eye." Every one of these is a request for the appearance of a real-world image artifact — and the models are now good enough at honoring those requests that the resulting outputs look meaningfully more authentic than their polished counterparts.

This is the meta-game of 2026. The fakes that win are the ones that look like they shouldn’t be worth faking. A perfect portrait raises the question; a blurry phone snapshot doesn’t. Forensic detectors that learned to flag overly clean images now have to learn the inverse: to recognize when imperfection itself has been generated.

II. CCTV, Dashcam, and the Cult of Low Quality

The category to watch in 2026 is fake surveillance footage — video and stills produced to mimic the visual language of security cameras, body cams, dashcams, and convenience-store recordings. The reason is straightforward: low-quality footage is socially trusted in a way high-quality images aren’t.

When a clear, well-lit, professionally framed image of a public figure doing something scandalous appears online, our default instinct is suspicion. When grainy 240p footage of the same scene appears, with a timestamp burned into the corner and a slight rolling-shutter wobble, the instinct flips. Bad quality reads as authenticity. Bad quality reads as accidental capture. Bad quality reads as evidence.

Sora and its peers have made this category cheap to produce. The same generators that struggle to render a perfect face will happily render a smudged, low-resolution face from across a parking lot, and that’s the face that ends up in the viral clip. We have already seen early instances: fake CCTV footage of staged crimes, fake dashcam recordings of celebrity confrontations, fake convenience-store videos used in coordinated harassment campaigns.

Detection here is brutally hard. The signals image forensics relies on — sharp edges, clean noise, predictable compression — are the very things that have been intentionally degraded. The future of detection in this category is less about pixel forensics and more about temporal coherence: subtle frame-to-frame inconsistencies that real cameras don’t produce, even when their output is bad. We are not there yet at scale.

III. The Document Frontier

In April 2026, screenshots circulated on X showing what appeared to be photographs of bank checks generated by a new image model, accompanied by mobile-deposit confirmations. Whether the deposits cleared is a separate question; the underlying check images were almost certainly synthetic. They were also nearly indistinguishable from real checks at thumbnail size.

Documents are the newest, fastest-growing fake category, and the existing detection infrastructure was almost completely unprepared for it. Forensic image detectors trained on portraits, landscapes, and stock photography have very little to say about a fake bank check, a fake court filing, a fake screenshot of a Twitter post, or a fake driver’s license — because the visual structure of a document is fundamentally different from the visual structure of a scene.

The implications cross every industry that relies on documentary evidence. Journalism that quotes "leaked memos" now needs to forensically verify the memo. Insurance claims supported by photographic evidence become arguable. Court filings that include screenshots of communications become contestable on the basis that the screenshot itself might be generated. Every place where a document used to be its own proof, it isn’t anymore.

On the detection side, document fakes leave different signals than scene fakes. The character spacing in a generated MICR line on a check is mechanically wrong in ways a real printed check never is. The kerning of fake court-document text drifts in ways no actual word processor would produce. Fake screenshots almost always have subtle errors in the chrome — wrong icon spacing, off-by-one font weights, slightly misplaced timestamps. A document-aware detector has to learn this entire second alphabet of forensic signals.

IV. The Art World Problem

For two years, AI-generated "art" was easy to spot because it was bad. The hands were wrong, the perspectives drifted, the color was uncanny. Galleries, auction houses, and online marketplaces had little to worry about because the generators were making things that didn’t look like the artists they were imitating.

In 2026, that changed. Midjourney v7 will produce a credible imitation of an oil painting in the style of any working contemporary artist; FLUX.1 will generate a watercolor that, scanned and printed at the right resolution, can pass casual inspection in a gallery. The generators have learned to produce the surface artifacts of physical media — visible brushstrokes, the way watercolor pools at the edges of shapes, the granularity of dry pigment on rough paper, the subtle warping of a canvas under tension. These are not pixel-perfect simulations. But they are good enough that the burden of proof has moved.

The category most exposed is mid-tier secondary-market art: estate sales, online auctions, the kind of "rediscovered work by a known artist" that occasionally appears at smaller houses. A skilled prompter with a good printer and access to a 19th-century-style canvas can manufacture a "lost" piece that requires expert authentication to debunk. The economics of authentication are not on the side of the buyer.

Within the contemporary scene, the fight is different. Living artists are seeing their style scraped, replicated, and sold under their name on dropship sites. Adobe’s Content Credentials and the C2PA standard offer one path forward — provenance metadata that proves a work was made with a real camera or a real digital tool. But adoption is voluntary, and the economic incentives for forgers to ignore the standard are exactly as strong as you’d expect.

V. Documentary and Historical Fakes

A genre that has emerged with surprising velocity is the fake archive image. Prompts like "1920s Berlin street photography, gelatin silver print, slight chemical staining" or "American Civil War battlefield, 1863, wet-plate collodion process" produce images that, viewed at the resolution typical of digitized archives, are difficult to distinguish from real period photographs.

These show up in three places. First, on Pinterest and TikTok, where "rediscovered historical photos" routinely go viral with no provenance whatsoever. Second, in propaganda — a fake "1940s photograph" of a political ancestor doing something incriminating is easier to plant than a fake recent photo because viewers don’t expect crisp resolution. Third, increasingly, in books, where unscrupulous publishers have used AI-generated "archival" imagery to fill out historical nonfiction at a fraction of the licensing cost of real material.

The detection challenge here is the genre’s defense in depth. Real archive images have surface artifacts — chemical degradation, film grain, scanner moiré, the warping of a digitized print — that look very similar to the artifacts a generator produces when prompted to imitate them. The forensic signals collapse into each other. The most reliable detection in this category is not pixel forensics; it’s reverse image search and provenance research.

VI. The Personal-Use Layer

Quietly, beneath all of the above, sits the largest volume of synthetic image traffic on the internet: personal use. Fake selfies for dating profiles. Fake "professional headshots" for LinkedIn. Fake vacation photos for Instagram. Fake "before and after" weight-loss photos. These are not used for fraud in any criminal sense. They are used for self-presentation in a culture where image is currency.

This category is uncomfortable to write about because the harm is diffuse. No single fake selfie damages anyone in particular. But the aggregate effect — an internet where the baseline assumption shifts from "this is probably real" to "this might be real" — corrodes a layer of social trust that we have not figured out how to replace. Online dating becomes a forensics exercise. Hiring screens for headshots become unreliable. The wedding photographer’s portfolio competes with prompts.

For detection, the personal-use category is a numbers game. There is nothing technically distinctive about a fake selfie compared to a fake portrait used in any other context. What there is, is volume. By some industry estimates, more than ten percent of profile photos uploaded to major dating platforms in late 2025 contained generative-AI manipulation, ranging from light retouching to fully synthetic faces. The platforms are mostly silent on this. Detection at scale is the only path to any kind of containment.

VII. What Still Trips the Models Up

It is worth being concrete about where the current generation still fails, because these are the signals detectors continue to rely on most heavily.

Reflections remain a weak point. A generated person standing next to a window will frequently have a mirror image that is subtly inconsistent — different posture, different lighting, missing details. Models reason about the scene as a flat composition, not as a 3D space, so the geometry of reflections degrades under scrutiny.

Shadows are inconsistent for the same reason. Multiple light sources should produce multiple shadows that agree on direction and intensity; generated images often produce shadows that contradict each other across the frame. This is most visible in indoor scenes with mixed lighting.

Crowd faces in the background are still telling. A primary face in a generated portrait will be rendered with the model’s full attention, but a crowd in the background — at a concert, a sporting event, a city street — frequently contains faces that, individually, look wrong: misaligned features, smeared skin, eyes that don’t resolve. The model has spent its budget on the foreground.

Hands holding objects remain unreliable, even though hands themselves have largely been solved. A hand resting empty looks correct; a hand gripping a cup, holding a phone, or interacting with a complex object often produces subtle physical impossibilities — fingers passing through the object, grips that wouldn’t support the weight, wrists at angles a body doesn’t bend.

And finally, text in the background — signs, labels, the spines of books on a shelf — is still a useful signal even though foreground text has been mostly fixed. The model concentrates its text-rendering capacity where it expects you to look. Look elsewhere.

These are the tells that detectors are most confidently exploiting in 2026. They are also the tells that the next generation of models will likely close, leaving us with the harder forensic signals — noise topology, frequency-domain inconsistencies, metadata absence — and with the categorical shift that this whole essay has been about: detection moving from a single problem to a dozen overlapping ones.

VIII. The Map, Going Forward

The unifying lesson, if there is one, is that the meaning of "AI image" has fragmented. In 2023 it was a category. In 2026 it is closer to a medium — like "video" or "audio" — within which there are dozens of distinct genres, each with its own conventions, its own use cases, and its own detection profile.

A detector that works on portraits will not work on documents. A system tuned to spot generated landscapes will be uncertain on fake CCTV. A model that flags the smooth-skin tells of 2024 will miss the grainy-nostalgia tells of 2026. The field is no longer "AI image detection." It is many fields.

For consumers, the most useful thing to internalize is that the question "is this image real?" is no longer answerable by visual inspection alone. The right question, increasingly, is "what kind of fake might this be, and what signals would distinguish that kind?" That is a research question, not a glance.

For us — the people building detection systems — it means continuous calibration against an expanding catalogue of subcategories. It means treating documents as different from scenes, scenes as different from video stills, video stills as different from archive photography. It means, eventually, that the verdict cards we show users will need to communicate not just a confidence score, but a category: "consistent with low-quality video generation," "consistent with stylized art generation," "consistent with document forgery."

The last thing to say is that this is not a doom forecast. Detection lags generation, but it does not give up. Every category of fake described here has been countered, partially, by tooling that didn’t exist a year ago. The pace will not slow. But neither will ours. The tradeoff between generation capability and verification capability is not a fixed loss — it is an ongoing negotiation, and one that is, for the first time, getting serious institutional attention.

In the meantime, the most valuable thing you can do as a reader is calibrate your skepticism by category. The fake portrait of a celebrity is the easy case. The fake CCTV clip, the fake quote-tweet screenshot, the fake "rediscovered" historical photo, the fake bank check — these are the cases where the old visual instincts no longer help, and where forensic tooling matters most.

The shapes of fake have changed. The map is being redrawn. This is one snapshot of where it stands.