How to Tell If a Photo Is AI Generated: 5 Checks (2026)

Quick answer

You can no longer tell if a photo is AI generated by eye, so use five layers. Check the context: a new account, every photo lit the same way, no group shots or tagged photos. Look for the tells that survive: teeth, earrings, hair edges, text, hands, reflections and background lines. Check provenance with Content Credentials, Google's About this image and the SynthID Detector, and treat detector websites as probabilities, not proof. Then ask for a short live video call: a generated still can't turn its head on request.

Key numbers

31%accuracy of people with typical face recognition skills at spotting AI-generated faces before training, rising to 51% after five minutes of instructionRoyal Society Open Science study reported by PetaPixel, December 2025
$893 millionlost across 22,364 complaints to the FBI in 2025 that involved artificial intelligenceFBI IC3 2025 Internet Crime Report
180 billion+images and videos carrying Google's invisible SynthID watermark since 2023, now checkable by anyone at synthid.comGoogle, October 2026

A profile photo looks great. Maybe a little too great: flawless skin, perfect light, a smile that could sell toothpaste. You're trying to work out how to tell if a photo is AI generated before you invest another week in the conversation, or agree to meet.

The honest starting point is that, as of October 2026, you usually can't tell from the face alone. In a 2025 study published in Royal Society Open Science, people with typical face recognition skills identified AI-generated faces only 31% of the time, worse than a coin toss. And generated faces are now a standard tool for fraud: the FBI logged $893 million in losses across 22,364 complaints that involved AI in 2025. So this guide teaches a layered method: context first, then the tells that still slip through, then provenance tools, then detectors, then the one test no image generator can pass.

Can you tell if a photo is AI generated just by looking?

Not reliably. Studies in 2025 and 2026 found that untrained people spot AI faces less than half the time, although a few minutes of training on specific tells helps a lot.

In the Royal Society Open Science study, 664 participants judged real faces against StyleGAN3 fakes. Untrained participants with typical face skills got 31% of the fakes right; even "super recognizers" managed only 41%. After five minutes of instruction on common errors, mainly unnatural hair patterns and the wrong number of teeth, typical participants rose to 51% and super recognizers to 64%. A 2026 study of 1,664 adults found higher average accuracy (85%) on portraits from ChatGPT-4o and Imagen 3, but accuracy fell with age, was lower on a phone than on a computer, and the ChatGPT-4o portraits were hardest to call.

The FBI's December 2024 alert on generative AI in fraud says criminals use AI-generated images for believable profile photos and for "proof" photos sent in private chats to convince victims they're speaking to a real person. Your eye is one weak layer. Stack it with four more.

Context clues: the account and the photo set, not the pixels.
Visual tells: the details generators still fumble.
Provenance: labels and watermarks attached when the image was made.
Detectors: probability scores, used carefully.
The live test: a short video call with an unscripted request.

What context clues suggest a profile photo is AI generated?

The strongest clue isn't in the pixels: it's a profile where every photo looks like it came from the same shoot and nobody else ever appears. Real people have messy photo histories. Generated people have a portfolio.

Same lighting, same lens, same distance in every photo. Soft even light, blurred background, subject centered.
No candid shots. Nothing mid-laugh, nothing badly framed, nothing taken by someone else at a party.
No group photos, and no photos where the face is at a sharp angle, partly hidden or in motion.
No tagged photos from other people and no comments from friends who clearly know them.
The account is new, or the posts were uploaded in a batch, with few followers. Our guide on how to tell if an Instagram account is fake shows where to find the join date and country.
The face doesn't age. Photos supposedly years apart show the same haircut, weight and skin.
The backstory is generic: a job title with no employer, a city with no neighborhood, interests that match yours exactly.

One older test has weakened. Asking for "a selfie holding up three fingers" used to catch stolen photo sets. Today's image models can keep a face consistent across new poses, so treat a passed photo request as a small plus, not proof.

How do I check the person behind the photo?

Check the person, not just the picture: reverse search two or three photos, then confirm that the name, job and city belong to one real person with a history. An AI-generated person has no past, and that absence shows.

A reverse image search gives you two very different results. If the photos were stolen from a real person, you'll usually find the original: a model's portfolio or a stock site. If the photos were generated, you'll usually find nothing, because the image never existed before it was made. So "no results" means either a private person or a synthetic one, and you need the next step to tell which. Our guide to how to reverse image search a photo has the steps for Google Lens, TinEye and Bing on phone and desktop.

The next step is the name. A Search A-Z dating match search takes the full name and country they gave you, plus any detail you know (city, employer, age range, phone or email), and tells you how sure we are that a real person matches. The report shows their current job and company, work history, education, public social and professional profiles, news and web mentions from a live web search, and any public scam or fraud warnings. You can add up to three photos from the profile and we run them through Google Lens to show where else those exact images appear; this is a search for copies of the picture, not face recognition, and the photos are deleted once the check has run. Dating searches show only the city, never a home address or relatives, and the person is never told you searched.

Read the result with AI in mind. A detailed story (a named employer, a degree, a city) attached to a person the report can't find at all is a strong signal, because real people with those details leave traces. If we can't find the right person, the report says so.

What visual tells still give away an AI generated image?

The tells that survive in 2026 are at the edges and in the details: teeth, jewelry, hair, text, hands, reflections and background geometry. Zoom in and count how many you find.

Tell	Where to look	How reliable
Teeth: too many, too few, merging into each other, or oddly even	Zoom on the mouth in any smiling photo	Medium. One of the two cues in the 2025 training study
Earrings and jewelry: a mismatched pair, a necklace that fades into skin, a clasp with no chain	Both ears, the neckline, watch and rings	Medium
Hair: strands that merge with the background or skin, hair that turns into fabric or a collar	Edges of the head, shoulders, where hair meets clothing	Medium to high. The other training cue
Text: letters on clothing, signs, badges or books that are almost words	Anything with writing in the frame	High when present
Hands: wrong finger count, fused fingers, joints bending the wrong way, held objects that don't connect	Hands and anything they hold	Medium. Models have improved; a clean hand proves little
Glasses and reflections: reflections that don't match the scene, frames that don't meet behind the ears, mismatched lenses	Lenses, frames, mirrors, windows	Medium
Background geometry: tiles, railings, window frames or door edges that bend or don't line up	Behind and beside the person	Medium
Shadows and light: a shadow under the nose that doesn't match the light on the wall	Under the chin and nose, versus the background	Medium. "Inaccurate shadows" is on the FBI's own list
Skin and symmetry: poreless, perfectly even skin with very sharp eyes; a face that's a touch too symmetrical	Cheeks versus eyes	Low on its own. Filters and retouching do this to real photos too

None of these is proof. Compression, filters and bad photography create some of them in genuine photos. Two or three tells in one photo set, plus the context clues above, is a pattern. One tell is a reason to keep checking.

How do I check a photo's provenance with Content Credentials, SynthID and About this image?

Provenance tools read labels attached to an image when it was created, so they can confirm "this came from an AI tool", but they can't confirm "this is a real photo", and they only work if the labels survived the journey to your screen.

Content Credentials (C2PA)

Go to contentcredentials.org/verify and upload the file or paste a public image URL. If the image carries Content Credentials, the Verify tool shows who produced it, which app or device was used, and, if a generative AI step was involved, which AI tool. If there's no manifest it shows "No Content Credential", the normal result for a screenshot or a photo that passed through most social apps, since credentials live in the file's metadata. A blank result is not a clean bill of health.

Google's "About this image"

In the Google app or on google.com, open an image in the image viewer, tap the info icon, then "About", then "Learn more about an image". Google's help page says it shows when Google first saw a similar version of the image, how other sites describe it, and, when metadata exists, a "How the image was made or edited" section that can indicate whether it was captured with a camera or generated with AI. Google warns that this metadata can be modified by editing tools, and the feature is only available in some regions.

Google's SynthID Detector

Since October 7, 2026, anyone can sign in at synthid.com and check an image, video or audio file for Google's invisible SynthID watermark, which Google says has been applied to more than 180 billion images and videos since 2023. The important limit: it only finds SynthID watermarks, which are added by Google's AI tools and by partners Google names, including OpenAI, NVIDIA, Kakao and Apple. An image from a tool that doesn't watermark comes back clean. A Notebookcheck test the day after launch found it caught a fresh Gemini image and, as expected, didn't flag an image from a generator outside the program.

Meta's "AI info" label

On Facebook, Instagram and Threads, content Meta detects as fully AI-generated carries an "AI info" label under the post; for content that was only edited with AI tools, the label sits in the post's menu, a change TechCrunch reported in September 2024. Meta says the label comes from industry-shared signals or the poster's own disclosure. No signal, no label, so a missing label means nothing.

Put together: a provenance hit ("made with Gemini", "Content Credentials: generative AI") is strong evidence. A provenance miss is weak evidence, because most profile photos, real or fake, arrive with their metadata stripped.

Do AI image detectors actually work?

Detector websites give you a probability, not a verdict, and independent testing shows they perform much worse on real-world images than their own accuracy figures suggest.

A 2025 benchmark called AIGIBench tested 11 advanced detectors across images from 23 generators plus real samples from social media and AI art platforms. The authors found "significant performance drops on real-world data", limited gains from the usual training tricks, and sensitivity to ordinary pre-processing such as the resizing and compression every social app applies. In plain terms, a detector trained on last year's generators and pristine files meets this year's generator and a compressed screenshot, and guesses.

Used carefully, they're one more layer:

Upload the largest, least-compressed copy you can get, not a screenshot of a screenshot.
Run two different detectors (Hive's AI detector, Sightengine and AI or Not are commonly used) and compare.
Treat anything short of a very high score as "unknown". Never treat "likely human" as clearance.
Weigh the score against the context clues and the tells. A detector result never overrides a video call.

Paid apps that promise certainty, or quote 99% accuracy measured on their own test set.

What test beats every AI photo check?

A short live video call with an unscripted request beats every photo check, because a generated still image can't turn its head on request, and the person behind an AI profile can't suddenly produce the face.

Ask in a friendly way: "Fancy a quick five-minute video chat before we meet?" On the call, ask them to turn to the side, pass a hand slowly in front of their face, and show you the room. Our guide to deepfake video call scams covers the live-call tests and what a face swap looks like when it breaks.

People who used AI photos usually don't take the call. Expect a broken camera, a bad connection or a promise for "next week". Repeated excuses after a week of eager chatting, plus any of the context clues above, is your answer. If you also see signs the whole identity is invented, our guide to how to tell if you're being catfished explains what to do next. And never send money to someone you've only met online, however real the photos look.

The bottom line

You can't reliably tell if a photo is AI generated by eye anymore, so don't try to. Read the account and the photo set for context, zoom in on teeth, hair, jewelry, text and backgrounds, run the provenance tools knowing a miss proves little, treat detectors as hints, and ask for a short live video call. A real person passes all five layers without noticing. A generated one fails the last.

Frequently asked questions

How can I tell if a dating profile picture is AI generated?

Look at the whole profile first: identical lighting in every photo, no group shots, no tagged photos, a new account and a generic backstory. Then zoom in on teeth, earrings, hair edges, text and background lines. Check provenance with contentcredentials.org/verify and Google's About this image, and ask for a short live video call. Refusal plus the context clues is your answer.

Is there a free tool to check if an image is AI generated?

Yes, three worth knowing. Google's SynthID Detector at synthid.com (open to everyone since October 2026) finds Google's and partners' invisible watermarks. The Content Credentials Verify tool at contentcredentials.org/verify reads C2PA labels if the file still has them. Google's About this image shows when the image first appeared and how it was made, where metadata exists. All three confirm AI when they find a label; none can prove a photo is real.

Why did a reverse image search find nothing for a suspicious photo?

Either the person is private and their photos aren't indexed, or the image was generated and never existed online before. Reverse search finds copies, and an AI image has none. Use the name instead: a Search A-Z dating match search shows whether a real person with that name, job and city exists, and the report says so if nobody matches.

Do AI generated photos have metadata?

Often very little. A camera photo usually carries EXIF data (camera model, date, settings); a generated image usually has none, and some generators attach Content Credentials saying which tool made it. But most platforms strip metadata on upload and screenshots never carry it, so missing metadata is common for real photos too. Treat it as a hint, not evidence.

Can AI image detectors be trusted?

Only as one input. A 2025 benchmark of 11 detectors found large performance drops on real-world images from social media and after ordinary compression. Use the best-quality copy you have, run two detectors, treat middling scores as unknown, and never let a 'likely human' result replace a live video call.

What should I do if I find out a match used AI generated photos?

Stop replying, don't explain what you found, and report and block the profile in the app. Don't send money, gift cards or crypto for any reason, and don't share more personal details. If you already sent money, contact your bank or payment app immediately and report it to the FTC at reportfraud.ftc.gov and the FBI at ic3.gov; in the UK use Report Fraud, and in the EU your national police.

Sources

Figures are the latest published by each source when this guide was updated. Reported fraud losses are a floor: most fraud is never reported.

How to tell if a photo is AI generated: a five-layer check for dating profiles and beyond