No single clue will reliably tell you whether an image is AI-generated. What works is stacking several quick checks. Find where the image first appeared, read its metadata for Content Credentials or generator tags, zoom in on the spots where generators still slip, and run it through an AI image detector. When three or four of those point the same way, you have a reasonable answer. When they disagree, treat the image as unverified.
Eyeballing alone is weaker than most people think. In a 2022 study published in PNAS, participants picked out AI-generated faces only 48.2% of the time, about the same as flipping a coin (Nightingale and Farid). So use the checks below in order, fastest first.
The eight checks at a glance #
| Check | Time | What it catches | What it misses |
|---|---|---|---|
| 1. Trace the source | 2 min | Anonymous accounts, recycled images | A real account posting an AI image |
| 2. Reverse image search | 1 min | Older originals, stock photos, known fakes | Brand-new images |
| 3. Read the metadata | 1 min | Content Credentials, generator tags | Stripped files, screenshots |
| 4. Look for labels and watermarks | 30 sec | Platform “AI info” labels, visible marks | Cropped or unlabeled uploads |
| 5. Zoom in on details | 2 min | Garbled text, broken geometry | Clean output from newer models |
| 6. Test the context | 2 min | Wrong weather, signs, uniforms | Plausible fakes |
| 7. Run an AI detector | 10 sec | Statistical fingerprints of generators | Heavy compression, very new generators |
| 8. Weigh it together | 1 min | Contradictions between checks | Nothing here is proof on its own |
1. Where did the image come from? #
Start with the account, not the pixels. Tap through to whoever posted it and look at their history. An account created last month that posts nothing but dramatic photos, with no location details and no follow-up shots, deserves more suspicion than a local reporter who has posted from the same town for years.
Then try to find the earliest post. Search the caption text, sort by oldest where the platform allows it, and read replies asking for a source. Somebody has often done the work already. If the only origin you can find is an aggregator account with its logo in the corner, you haven’t found the source yet.
2. Does a reverse image search find an older original? #
Upload the image to Google Lens, TinEye or Bing Visual Search. You’re looking for one of three things: the same picture from years ago (a real photo with a new caption), a stock or press photo it was cut from, or a fact-check that already debunked it. Reverse search is excellent at the first two and useless for an image generated this morning. We compare the two approaches in reverse image search vs AI detector.
3. What does the metadata say? #
Some generators write their name into the file. Images from Adobe Firefly, OpenAI’s image tools and Google’s Nano Banana Pro can carry C2PA Content Credentials, a signed record of how the file was made. Others leave XMP or IPTC fields, such as a “digital source type” that marks the picture as algorithmically generated, or a software tag naming the tool.
The catch is that metadata disappears easily. Most social platforms strip it on upload, and a screenshot creates a new file with none of the original’s history. Metadata that says “AI” is strong evidence, while missing metadata tells you almost nothing. Our guide to checking image metadata for AI shows where to look on a phone and a computer.
4. Is there a visible label or watermark? #
Check the corners and the post itself. Images made or edited in the Gemini app carry a visible watermark as well as Google’s invisible SynthID mark, according to Google. Meta puts visible markers on images made with Meta AI. Instagram and Facebook show an “AI info” label when they detect industry-standard AI signals or the uploader discloses AI use (Meta), and TikTok auto-labels content that carries Content Credentials.
A missing label proves nothing. Watermarks get cropped out, and plenty of generators never add one.
5. Zoom in where generators slip #
Open the largest version you can find and look at the parts that are hard to fake:
- Text. Shop signs, license plates, book spines, T-shirt slogans. Letters that melt into shapes are a classic tell, though newer models spell much better.
- Hands and small objects. Fingers merging into a cup, earrings that don’t match, glasses with one arm missing.
- Geometry. Railings that change spacing, floor tiles that don’t line up, windows at impossible angles.
- Light. Shadows falling in two directions, reflections that show a different scene, a face lit from the front while the sun sits behind it.
- Backgrounds. Crowds of half-formed faces, repeated objects, a soft painterly blur where a camera would show detail.
Artifacts are evidence for AI. Their absence is not evidence against it, because the best current generators produce clean hands and readable text most of the time. We cover the specific flaws and why each one happens in AI image artifacts: what generators still get wrong.
6. Does the scene make sense? #
Check the image against facts you can look up. If a photo claims to show a flood in a named city yesterday, was it raining there? Do the street signs use the right language and design? Are the police uniforms, license plates and vehicles right for that country? Generators produce plausible scenes, not accurate ones, and a context check catches fakes that pass every pixel test.
7. Run an AI image detector #
Detectors look for the statistical fingerprints generators leave behind: patterns in texture, noise and color that people can’t see. They’re the fastest check on this list and the only one that works on a brand-new image with no metadata.
Expose AI handles this step on a phone. Share an image to it from your browser, Photos or a social app, and it runs two passes on the device: a metadata read for Content Credentials and generator tags, then a neural network trained on the visual fingerprints of AI image generators. You get Likely real, Uncertain or Likely AI-generated, a confidence meter, and a “How we decided” card that says which pass made the call. The image is never uploaded, and checks are free with ads.
No detector is right every time. Compression, screenshots, heavy edits and generators released last week all reduce accuracy, which is why a well-built detector says “Uncertain” instead of guessing. See how accurate AI image detectors are for what the numbers mean.
8. Weigh the evidence together #
| What you found | Reasonable conclusion |
|---|---|
| Metadata or a platform label says AI | Treat it as AI-generated |
| Reverse search finds an older original | Real photo, possibly with a false caption |
| Detector says Likely AI and you see artifacts | Very likely AI |
| Detector says Likely AI, but the source looks solid | Ask the source before sharing |
| Detector says Likely real, but the context is wrong | Possibly a real photo from somewhere else |
| Everything is inconclusive | Unverified; don’t pass it on as fact |
You’re aiming for a decision you can defend, not certainty. If you can’t verify an image, you can still decline to share it, or share it with “unverified” attached.
Frequently asked questions #
Can you tell if an image is AI just by looking? #
Not reliably. In the PNAS study, people performed near chance on AI faces, and training with feedback only lifted accuracy to about 59%. Visual flaws help when they’re present, but many AI images now have none, so combine looking with source and metadata checks.
Is there a 100% accurate AI image detector? #
No. Every detector makes mistakes, including false positives on real photos. Vendor accuracy figures come from their own test sets, and real-world images that have been compressed, cropped or edited are harder. Use a detector’s result as one strong signal among several.
Does taking a screenshot remove AI metadata? #
Yes. A screenshot is a new file, so Content Credentials or generator tags in the original don’t carry over. A detector’s visual model can still analyze a screenshot, but crop away the app interface first for a cleaner read.
Can an AI detector tell which generator made an image? #
Sometimes. If the metadata names the tool, you’ll know exactly. Some online detectors also estimate the likely source model from the pixels, but that guess is less reliable than the basic AI-or-real call.
Do detectors catch photos that were only edited with AI? #
Partly. Face swaps and large generative edits often show up, but a small AI touch-up inside a real photo can read as real, because most of the pixels still came from a camera.