How to tell if an image is AI-generated

There is no single reliable tell, and anyone offering a five-second trick is selling something. We can say that with unusual confidence because checking is our day job: we run a detection benchmark that scores thousands of images through the compressions and screenshots real photos actually survive, and the numbers in this guide come from it. What exists instead of a trick is a ladder of checks, from free-and-fast to rigorous, each catching things the others miss. This guide walks the ladder in order and is explicit about where each rung fails.

Start with the checks that end investigations early

Reverse image search first

Search the image before you study it (Google Lens, TinEye, Bing). If it appeared years earlier in another context, you have answered a different and often more useful question: it is not what the caption claims, whether or not it is AI. A large share of viral "is this real" images fall at this rung, and it costs nothing.

Then the source

Who posted it first, and do they exist beyond this one image? A fresh account with one viral photo is its own signal. Look for corroboration: other angles, other witnesses, a place and time you can verify. An image that cannot be tied to anyone or anywhere deserves suspicion regardless of what a detector says.

What you can check by eye

Visual inspection is the weakest rung, and it gets weaker every year. The famous tells (six fingers, garbled text, melted jewelry) were real flaws of 2022–2023 generators, and current models have largely fixed them. Treat the absence of visual glitches as no information at all. When a tell is present, though, it is still a lead:

Text and signage

Look at anything with letters: shop signs, book spines, jerseys. Generators have improved, but long or incidental text in the background still comes out mangled more often than foreground text.

Geometry and light

Follow straight lines: railings, tiles, window frames. Check that they stay straight behind an occluding object and come out the other side in the same place. Then pick two objects and ask whether their shadows agree on where the light is. Reflections (in mirrors, water, glasses, eyes) should contain a plausible version of the scene, and inconsistent reflections remain one of the harder things for generators to get right.

Textures and repeats

Crowds, leaves, bricks, fabric patterns. Generators sometimes tile subtle near-copies of the same face or texture patch. At full zoom, look for the airbrushed sheen: skin without pores, hair strands that merge, teeth without individual edges.

If you find one of these, you have a lead, not a verdict. Real photos contain weird geometry, bad lighting, and compression mush all the time.

Generator-specific tells age fast

The folklore problem

"Midjourney images look cinematic," "this model can't do hands," and every similar rule of thumb has an expiry date, usually the next model release. The tells that circulate on social media describe last year's generators. Folk knowledge is not useless, but it tells you what images from a model tended to look like at one moment, under default settings a deliberate user can simply change.

What actually persists: watermarks and credentials

The durable generator-specific signals are the deliberate ones. Some large providers embed invisible watermarks in their output (Google's SynthID is the prominent example), and some tools attach cryptographically signed content credentials (C2PA). When such a mark is present and verifiable, it is strong evidence of origin. The catch is coverage: most images you will actually encounter carry neither, marks do not survive every transformation, and absence of a watermark is evidence of nothing at all.

What a detector measures instead

Statistical detection does not rely on any of this. It reads traces the generation process itself leaves in pixel statistics, which is why it works across generators rather than memorizing one model's look. It is also honest about the limit: detection is strongest on generator families represented in its calibration, and weaker on a model released last week. Our benchmark measures that difference per generator family instead of hiding it in one pooled number.

Where statistical detection fits

AI generators leave statistical traces in pixel data: patterns invisible to the eye that a trained model can score. A detector reads the image and returns a score; the crucial part is how that score becomes a verdict.

The number to demand from any detector is its false-positive rate: how often it calls a real photo AI. A detector that flags 27% of genuine messenger photos (which is what we measured from a public detector at its default threshold on our own benchmark) is worse than useless, because its alarms carry no weight. Our analyzer only issues verdicts at thresholds calibrated so that at most 5% (standard) or 1% (strict) of real photographs get flagged, and the report prints the cap it used.

That rigor has a price, and you should hear it from us: at the ≤5% false-positive setting, across our benchmark's conditions, including images that have been through WhatsApp and screenshots, the detector catches roughly 55–64% of AI images, not 99%. If a service quotes you a number like 99%, ask what it means. The answer is usually "on clean images from generators we trained on."

Two consequences follow, and they are printed on every one of our reports:

1. A flag is strong evidence. At the strict setting, a flagged image comes from a process that is wrong about real photos at most 1% of the time.
2. "Not flagged" is not "real." A calibrated detector deliberately lets borderline AI images pass rather than inflate false alarms. Absence of a flag is absence of evidence, never a certificate of authenticity.

What nothing can tell you

• No method proves an image is real. Detection can flag synthesis; it cannot certify authenticity.
• Brand-new generators are a moving target. A model released last week may not be represented in any detector's calibration yet, including ours.
• Small, cropped, and heavily edited images carry less signal. The less pixel data survives, the less any statistical method can say.
• Certainty is not on the menu. Anyone offering a definitive yes/no, human expert or software, is overclaiming.

A practical order of operations

1. Reverse-image-search it. If it is old or recontextualized, you are done.
2. Look for the visual tells above: leads, not verdicts.
3. Check the source: earliest posting, account history, corroborating images.
4. Run a calibrated detector and read the verdict with its false-positive cap and its limits, not as a yes/no oracle.
5. If the stakes are high (money, safety, reputation, evidence), do not act on any single check. Combine them, and treat "unclear" as an acceptable answer.