How to Tell If an Image Is AI-Generated: 9 Signs + Free Tools

A photo lands in your feed and something feels off, but you can't say what. Two years ago you'd count fingers and move on. In 2026, the models render hands, teeth and even street signs well enough that the classic checklist mostly clears fakes.

To tell if an image is AI-generated, check three layers: physics (do lighting, shadows and reflections obey one light source?), texture (is fine detail suspiciously uniform?), and provenance (does the file carry C2PA metadata or match an earlier real photo?). Then confirm with a detector scan, because no single signal is proof on its own.

The old giveaways are disappearing

Every guide written between 2022 and 2024 taught the same tells: extra fingers, fused teeth, earrings that don't match, text that dissolves into alphabet soup. Those were real weaknesses of early diffusion models, and they're mostly gone. Midjourney, DALL-E, and Google's image models now handle hands and short text correctly far more often than not, and each release closes more gaps.

That didn't happen by accident. Hands and text were the most mocked failures of the early generators, so they're exactly what the labs fixed first — more targeted training data, dedicated text-rendering passes, aggressive tuning against the famous mistakes. The result is a strange inversion: the tells everyone memorized are the ones most reliably patched, while the failures nobody made memes about — shadow geometry, sensor-noise behavior, background logic — quietly persist. Guides that haven't updated since 2024 now actively mislead people into clearing fakes.

The research backs up how weak unaided eyes have become. In the ARIA study (Li et al., 2024), which built a benchmark of more than 140,000 real and AI images, participants correctly identified only 61.58% of AI-generated images — barely better than a coin flip — while clearing real photos 79.87% of the time. And those participants knew they were being tested. Casual scrolling is worse.

So the honest starting point is this: your eyes alone are no longer enough. But they're not useless either. The tells have moved from anatomy to physics, and physics is harder for generators to fake consistently.

How to tell if an image is AI-generated: the 9 signs

Diffusion models paint plausible pixels; they don't simulate a world. That difference leaks out in three places: light, surfaces, and scene logic.

Light and physics (signs 1–3)

1. Shadows that disagree. Pick two or three objects and trace their shadows back toward the light. In a real photo, every shadow points away from the same source(s). AI images regularly light the subject from the left while a lamppost throws its shadow toward the camera. Soft, directionless "studio glow" over an outdoor scene is the same failure in disguise.

2. Reflections that don't reconcile. Windows, mirrors, water, sunglasses, and eyes all have to re-render the scene from a different angle, and generators frequently fake it. Look for reflections showing objects that aren't there, skies that don't match, or two eyes reflecting different rooms.

3. Impossible focus. Real lenses produce one focal plane: things at the same distance share the same sharpness. AI images often keep a face and a distant building equally crisp, or blur one shoulder of a subject while the other stays sharp. If the depth of field has no consistent geometry, a camera probably wasn't involved. (Product photography is the honest exception — focus stacking keeps everything sharp on purpose — so weight this sign most heavily on candid, single-shot claims.)

Surface and texture (signs 4–6)

4. Uniform micro-detail. Skin is the classic case: real faces have pores, fine hairs, blotches and asymmetries that vary across regions. AI skin tends to be evenly, flatteringly detailed everywhere — a kind of airbrushed consistency. The same applies to wood grain, brick and grass.

5. Synthetic grain. Real sensor noise concentrates in shadows and smooths out in highlights. AI "grain" is typically identical across bright and dark regions, because it was hallucinated rather than captured. Zoom to 200% on a dark corner and a bright patch and compare.

6. Repetition where nature varies. Generators love patterns, so they repeat them: identical leaves, cloned windows on a building, the same face twice in a crowd, hair strands that loop back into themselves. Real scenes are messier.

Scene logic (signs 7–9)

7. Background melt. Attention concentrates on the subject, so coherence decays with distance from it. Chairs merge into tables, bicycles lose wheels, background pedestrians share limbs. Scan the edges of the frame, not the center.

8. Text that fails on the second look. Headline text is usually clean now, but secondary text — a shop sign down the street, a book spine, a jersey number — still tends toward almost-letters. Real-world typography is everywhere, and generators can't sustain all of it.

9. Suspicious perfection. Composition centered just so, color grading straight out of a cinema trailer, every surface clean, every face symmetrical. Individually meaningless, but a strong prior. Our breakdown of what gives Midjourney images away digs into this aesthetic signature in detail.

The compressed version, for when you're checking on the move:

Check the metadata before you trust your eyes

While you were squinting at shadows, the file itself may have been carrying an answer. C2PA Content Credentials are cryptographically signed provenance records embedded in the file — a tamper-evident label saying what made the image and what's been done to it since. OpenAI has embedded them in DALL-E images since February 2024, and Adobe Firefly, Microsoft's image tools and Google's latest image models do the same. Paste a file into the free Content Credentials verify tool and you'll see whatever manifest survives.

The catch is that word, survives. Major social platforms strip metadata on upload, screenshots carry none, and Midjourney never embeds credentials in the first place. That creates an asymmetry worth memorizing: a C2PA manifest naming a generator is close to proof; an empty result is no evidence at all. Our full guide to C2PA Content Credentials explains what these manifests contain and where they break.

Two more file-level notes to keep straight. Ordinary EXIF data — camera model, exposure settings, GPS — is weak evidence in both directions: it's trivially editable, so a fake can wear a Canon's fingerprint, and plenty of authentic photos shed it during processing. And Google plays a different game entirely: its Gemini-generated images carry SynthID, an invisible watermark woven into the pixels rather than attached as metadata. It survives handling far better than C2PA, but only Google's own tools can read it, so it won't help your independent check.

Pair the metadata check with a reverse image search — if the picture existed in 2019, no 2026 generator made it. The reverse-search verification workflow walks through that step by step.

Let a detector see what you can't

Visual inspection and provenance still leave a gap: a Midjourney image, stripped of metadata, with no lighting mistakes. That's where statistical detection earns its place. AI image detectors analyze pixel-level patterns — noise distributions, frequency artifacts, upsampling fingerprints — that generation pipelines leave behind and human eyes can't register.

They work, with caveats you should know before trusting any score. In the ARIA benchmark, most open-source detectors landed below 70% accuracy on AI images, and even commercial tools fell apart on certain generation modes. Compression is the other enemy: when Bellingcat tested a leading detector in September 2023, it correctly flagged ten uncompressed Midjourney images, then missed seven of the ten after social-media-level compression. We've published an honest look at AI image detector accuracy if you want the full picture, including where our own tool struggles.

The AI Detector 360 image checker approaches this with layered evidence rather than a lone percentage: a probability score with an explicit confidence level, likely-generator attribution, and provenance signals (C2PA, EXIF, generation parameters) pulled from the file in the same scan. When the input is a heavily compressed repost, the report says the confidence is low instead of bluffing.

Two habits make any detector more useful. Feed it the best copy you can find — a first-generation file carries far more statistical signal than a thumbnail that's been reshared four times. And ask what a mistake would cost before acting: clearing a meme wrongly costs nothing, while flagging a photojournalist's work wrongly costs a great deal, so the same 80% score justifies different responses in different contexts.

Put it all together: a 60-second workflow

For everyday verification, speed matters more than perfection. Here's the compressed version of the nine signs plus tooling:

1. 10 seconds — light and shadows. One light source? Reflections plausible?
2. 10 seconds — texture and background. Uniform grain? Melting objects at the edges?
3. 15 seconds — provenance. Content Credentials check, plus a reverse image search for the earliest copy.
4. 15 seconds — detector scan. Upload to AI Detector 360's image detector and read score plus confidence.
5. 10 seconds — decide. Two or more independent signals pointing the same way is a strong result. One ambiguous signal is a shrug — keep the skepticism.

Here's how that plays out on a real case. Say a "photo" of storm damage is circulating during a hurricane. The light looks plausible (10 seconds, no verdict). The debris field repeats the same crushed car three times (flag one). Reverse search finds no earlier copy, which fits both "just taken" and "just generated" (neutral). No Content Credentials (neutral — that's the internet's default). The detector returns 91% AI at high confidence (flag two). Two independent flags, zero counter-evidence: don't share it, and say why. Total time, about ninety seconds.

The same logic extends to moving images, where frame-level artifacts pile up fast; our AI video detector applies frame-by-frame analysis for exactly that reason, and the tells for video get their own treatment in our deepfake guides.

No method here is infallible, including ours — anyone promising certainty about AI images in 2026 is selling something. But stacked together, these checks catch the overwhelming majority of fakes in about a minute, which is more than most newsrooms managed in 2023.