To tell whether a profile photo is AI-generated, combine three checks: look closely at details AI still gets wrong (ears, jewellery, teeth, hair edges, background text), run a reverse image and face search to see whether the person exists anywhere else online, and check for content credentials that record how the image was made. No single check is reliable on its own, and AI-detector scores in particular should be treated as a hint, not a verdict.
This guide covers the visual tells for still photos and video, what reverse image search can and can't prove about AI images, why detectors disagree, how C2PA content credentials work, and what to do once you suspect a fake.
Start with what the image claims
An AI image is only a problem when it is presented as something it isn't. A clearly labelled AI illustration is fine. A synthetic face used as someone's dating profile photo, or a fake picture presented as evidence of an event, is not.
So begin by writing down the claim attached to the image:
- "This is me, taken last month." (A profile photo.)
- "This happened yesterday in this city." (A news or viral image.)
- "This is the item I'm selling." (A listing photo.)
The rest of the checks test that specific claim. You often don't need to prove an image is AI-generated; showing that the claim doesn't hold up is enough to decide what to do.
Visual tells in still photos
AI image generators have improved quickly, and many older tells (six fingers, melted faces) are now rare in good outputs. Look for clusters of smaller inconsistencies rather than one obvious error. Zoom in on the full-resolution image, not a thumbnail.
| Where to look | What can go wrong in AI images |
|---|---|
| Ears and earrings | Mismatched ears, earrings that differ left to right or merge into the lobe |
| Glasses | Frames that change shape between sides, lenses that don't distort what is behind them |
| Teeth | Too many or too uniform teeth, blurry edges where teeth meet gums |
| Hair | Strands that dissolve into the background, hair growing through clothing or skin |
| Hands and fingers | Wrong count, fused fingers, unnatural joints, rings that melt into skin |
| Skin | Uniformly smooth, airbrushed texture with no pores, or repeated texture patterns |
| Background | Warped lines, doorways that don't line up, objects that make no sense, blurred "nowhere" settings |
| Text | Garbled signs, shop names and logos that look like letters but don't spell anything |
| Lighting and shadows | Light on the face coming from a different direction than shadows in the scene |
| Symmetry | Collar, buttons, straps or patterns that don't match from one side to the other |
Two patterns are worth knowing:
- Studio-style portraits from older face generators often have the eyes in nearly the same position in every image, a soft, blurry background and odd artefacts where the hair meets the background. If several profile photos have the eyes in exactly the same spot, be suspicious.
- A "set" of profile photos that don't quite match each other. AI tools struggle to keep one invented person consistent across many images. Compare the photos side by side: nose shape, eyebrow arch, moles and ear shape should stay the same.
Treat visual clues as questions. Compression, beauty filters, portrait mode blur and ordinary editing can all create oddities in real photos. A convincing-looking image can still be synthetic.
Signs of a deepfake in video and live calls
Live face-swap and lip-sync tools are now used in some romance and job scams, so a video call is no longer a complete guarantee. Common tells, though weaker in newer tools:
- Edges of the face shimmer, blur or slip when the person turns their head, especially at the jawline and hairline.
- Turning sideways makes the face distort or snap back. Ask them to turn fully to the side and back.
- Objects in front of the face, like a hand, a cup or glasses being taken off, cause glitches.
- Lip movement lags slightly behind the audio or doesn't form the shapes of the words.
- Blinking and expression look flat, or the face stays oddly still while the head moves.
- Lighting on the face doesn't change when the person moves relative to a lamp or window.
- Excuses about the connection whenever you ask for something unscripted.
Simple, friendly requests work: "Can you wave?", "Show me the view out your window", "Turn round, is that your cat?" A real person can do these easily. Someone using a face filter often avoids them. None of this is definitive, so combine it with the photo and behaviour checks below.
What reverse image and face search can tell you
Search can't analyse how an image was made, but it can test whether the person or the photo has a history.
Reverse image search
A copy search on Google Lens, Bing or TinEye can find:
- The original source, if the image was taken from an AI art gallery, a "fake person" generator site, or a stock library that sells AI-generated people.
- An earlier real photo that was edited, for example a real photo with an AI-changed face or background.
- Older appearances with a different caption, which is common for viral images presented as new. See how to verify a photo's caption and context.
An empty result does not mean the image is AI-generated. Many real photos have no copies online. And a match does not make the caption true.
Face search
A face search looks for the same person in other photos. This helps with profile photos:
- If a face search finds the same person across years of posts, tagged photos and different settings, the face almost certainly belongs to a real person. Whether that person is the one messaging you is a separate question.
- If a face search finds nothing at all for someone who claims an active social life, that is consistent with an invented face, though it is also consistent with a private person.
Facehunch runs face searches for $4.99 per search, with every result shown with its source link and a similarity score. For how the two search methods differ, see face search vs reverse image search.
AI detectors: useful hint, unreliable verdict
AI-image detectors are classifiers trained to spot statistical traces of particular generators. They can be helpful as one signal, but they have real limits:
- They lag behind new models. A detector trained on last year's generators may miss images from this year's.
- They are thrown off by processing. Screenshots, resizing, compression and filters, all common on social media, can push a score either way.
- They produce false positives. Heavily edited or beautified real photos are sometimes flagged as AI.
- Different detectors disagree. Running the same image through several tools often gives conflicting answers.
- The score is the tool's estimate, not a probability that holds across all images.
How to use one sensibly:
- Use the highest-quality version of the image you can get, not a screenshot of a screenshot.
- Record the result as a tool estimate, with the tool name and date.
- Keep it separate from source evidence (pages, dates, captions) and visual observations.
- Never accuse a person based on a single automated score.
| Check | What it can show | What it can't show |
|---|---|---|
| Reverse image search | Earlier copies, original source, changed captions | Whether an unmatched image is synthetic |
| Face search | Whether the face appears elsewhere as a real person | Who is controlling the account |
| AI detector | Statistical traces that suggest generation | A reliable verdict, especially after compression |
| Content credentials | Recorded origin and edit history, if present | Anything, if credentials were stripped or never added |
| Visual inspection | Specific inconsistencies to question | Proof, since real photos have oddities too |
Content credentials and C2PA
The most reliable signal, when it exists, is provenance data attached to the file itself.
The C2PA (Coalition for Content Provenance and Authenticity) publishes an open standard for recording where an image came from and how it was edited, cryptographically signed so tampering can be detected. The consumer-facing name is Content Credentials. Some cameras, editing apps and AI image tools add them; Adobe Firefly and OpenAI's image tools, for example, attach credentials that say the image was AI-generated.
To check an image:
- Get the original file if you can, not a screenshot. Screenshots and many uploads strip metadata.
- Upload it to the Content Credentials Verify tool.
- Read what it reports: the tool or device that created it, any edits, and whether an AI generator was involved.
Know the limits:
- Missing credentials prove nothing. Most images, real and fake, have none, and most social platforms strip metadata when you upload.
- Credentials record history, not truth. A real photo with valid credentials can still carry a false caption.
- Other markers exist. Some AI providers use invisible watermarks (Google's SynthID is one) that only their own tools can read, and some platforms show their own labels, such as Meta's "AI info" label on Facebook and Instagram, when they detect standard markers.
Putting it together: a step-by-step check
For a suspicious profile photo:
- Save the full-resolution images, cropping screenshots to the photo itself.
- Inspect each one against the visual tells table, comparing photos with each other.
- Run a reverse image search on each photo, sorted by oldest where possible.
- Run a face search if copy search finds nothing.
- Check content credentials on any original files you have.
- Optionally run a detector, and log the result as an estimate.
- Ask for a short video call and watch for the live tells.
- Write a one-line finding: "No history for this face anywhere, photos inconsistent with each other, refused video call" is a much stronger basis than any single score.
If the evidence is mixed, call the image unverified rather than forcing a verdict.
What to do if you find a fake
- Dating or social profile using an AI face: stop sharing personal information, don't send money or codes, save screenshots and the profile URL, then report it in the app as a fake profile. The fake dating profile guide covers reporting and securing your accounts.
- Money already sent: contact your bank immediately and report to ReportFraud.ftc.gov and the FBI's IC3 in the US, or your national fraud service elsewhere.
- A fake image of you: collect evidence first (see documenting photo impersonation), then report it to the platform. For intimate images, real or AI-generated, StopNCII.org helps adults block them across participating platforms, and NCMEC's Take It Down helps with images taken when the person was under 18.
- A fake news or viral image: don't share it, and if you already did, correct or delete the post.
Search tools here are for checking someone who has contacted you or protecting yourself, not for identifying strangers. If you are worried a real person's photos are being used in a fake account, the catfish detection page shows how to run a quick photo check.
Written by Facehunch with AI assistance and checked against the cited sources. Our editorial policy.