To tell if a photo is AI, check provenance first: a watermark checker such as SynthID Detector or openai.com/verify, then the file's Content Credentials, then a reverse image search, and only then visual tells. A found watermark is strong evidence; nothing found proves little. Classifier-style AI photo detectors give a probability, and independent research shows their accuracy drops sharply on images from the real internet.
To tell if a photo is AI, check where it came from before you judge how it looks: run the original file through a watermark checker and a Content Credentials inspector, reverse-search it, then look for visual tells. A detected watermark is strong evidence that an AI tool made or edited the image, but a clean result only means that particular signal was not there, so no single AI photo detector gives you a yes or no.
Checked October 2026 against the SynthID Detector page and Google's 7 October 2026 announcement, Gemini's help page on verifying AI-generated media, Google's May 2026 provenance update, OpenAI's help page on provenance signals, the Content Credentials inspector, and the tools' own pricing pages linked in the table. Accuracy figures come only from the two named studies below.
This guide is for anyone looking at a suspicious image: a profile photo that seems too perfect, a viral picture, a picture of you that you never posed for. If you make AI images yourself, the same checks show you what your audience and platforms can see, and our guide to Instagram AI labels covers when you have to disclose.
Is this photo AI? The 5-step checklist
Work in this order. The early steps give hard evidence when they hit, and the later ones are judgement calls you only need when the early ones come back empty.
- 1Get the original file
Download the highest-quality version you can, not a screenshot. Screenshots and re-saves strip metadata and weaken watermarks.
- 2Check for an AI watermark
Upload it to synthid.com, and to openai.com/verify for OpenAI images. Expected: watermark detected, or not detected.
- 3Read its Content Credentials
Drop it into verify.contentauthenticity.org. Expected: the tool that made it and its edit history, or no credentials found.
- 4Reverse-search it
Use Google Lens or TinEye to find the oldest copy and who posted it first. Expected: a source, or nothing older than the post you saw.
- 5Look for visual tells
Zoom into text, hands, jewellery, shadows and the background. Expected: clues, never proof.
Step 1: check for an AI watermark with SynthID
Many AI image tools now hide a watermark in the pixels. SynthID, built by Google DeepMind, is the most widely adopted, and its checker opened to the public on 7 October 2026. The SynthID Detector page says it detects media made with consumer products from Google, NVIDIA, OpenAI and Kakao, and Google says Apple is coming.
- synthid.com: upload an image, video or audio file. The result is "watermark detected" or "watermark not detected".
- The Gemini app: upload the file and ask "Is this AI-generated?". Google's help page allows one file at a time up to 100 MB and about 10 image checks per rolling 24 hours (checked October 2026). It recognises Google AI watermarks only.
- openai.com/verify: checks images and audio for OpenAI's watermark or its signed Content Credentials. OpenAI says images made with ChatGPT, Codex or its API carry both.
Read the result carefully. The SynthID page says a "not detected" result does not rule out AI, because the image may come from a model without SynthID or may have been edited heavily, and that in very rare cases it can trigger on unmarked content. Gemini can also answer "unclear", for example when an edit was too small to carry a watermark.
Step 2: read the metadata and C2PA Content Credentials
Content Credentials are a signed record inside the file of which tool created it and how it was edited, built on the C2PA standard. OpenAI, Google and many camera and software makers write them, and not only for AI: Google says the Pixel 10 was the first phone to add them to photos from its built-in camera, and that Meta will label camera-captured media with Content Credentials on Instagram.
- Woven into the pixels
- Survives resizing, recolouring and compression in most cases
- Says only "made or edited by a supported AI tool"
- Checked at synthid.com, in Gemini or at openai.com/verify
- Stored as signed data attached to the file
- Lost to screenshots, format changes and many uploads
- Can show the tool, the date and each edit
- Checked at verify.contentauthenticity.org or in Gemini
Source: OpenAI Help Center and SynthID Detector, checked October 2026
The Content Credentials inspector accepts JPEG, PNG, WebP, HEIC, AVIF, TIFF, DNG and several video and audio formats, and warns that credentials are still rolling out, so most files will show nothing. Ordinary metadata (EXIF) is weaker evidence: a camera model in the file suggests a real camera, but anyone can edit it, and missing EXIF is normal for any image that passed through social media.
Step 3: reverse image search for the source
Provenance checks tell you how a file was made. A reverse search tells you where the picture came from, which often settles the question faster. Look for the oldest copy, the account or outlet that posted it first, and whether a news agency or photographer credits it.
- Google Lens and Circle to Search: Google said in May 2026 that you can ask "Is this made with AI?" in Lens, AI Mode and Circle to Search, with SynthID checks live and C2PA checks to follow.
- TinEye: its help centre says searches are free for non-commercial use. Sort results by oldest to find the first appearance.
Two patterns point to AI: the image appears nowhere before a single viral post, or near-identical versions turn up with small differences in the faces or background, which happens when someone generated several and posted the best.
Step 4: visual tells of an AI photo
Visual tells are where most guides start. They belong last, because modern image models have fixed the famous ones, and because people are not good at this. Use them to decide whether to dig further, not to decide the answer.
| Where to look | Common AI tell | Caveat |
|---|---|---|
| Text and signs | Letters that melt, repeat or spell nothing; logos close to a real brand but wrong | Real photos have readable text, or blur it evenly |
| Hands and teeth | Extra or fused fingers, nails at odd angles, too many teeth | Better models get these right, so a good hand proves nothing |
| Jewellery and glasses | Earrings that do not match, frames that merge into skin, one lens reflecting a different scene | Look at both sides of the face |
| Light and shadow | Shadows falling in two directions; a face lit from a side with no light source | Check shadow direction against the brightest area |
| Backgrounds | Repeating windows, people with smeared faces, lines that bend for no reason | Zoom into the parts the photographer did not care about |
| Skin and hair | Poreless, airbrushed skin; hair that dissolves into the background | Phone beauty filters cause the same look, so this alone is weak |
- People did best on portraits
- They struggled most with natural and urban landscapes
Source: Microsoft AI for Good Lab study, arXiv 2507.18640, read October 2026
That study is the honest baseline: a careful eye gets you only a little past a coin flip on average. If you are checking an AI-made portrait of a person rather than a landscape, your odds are better, but not good enough to accuse someone on a hunch.
AI photo detectors: what they can and cannot do
"AI photo detector" covers two different kinds of tool. Provenance checkers look for a signal a known tool put there on purpose. Classifiers guess from the pixels whether any AI made the image. The first kind is precise but narrow; the second covers everything but can be wrong in both directions.
Drop in AUC on deepfakes that circulated online in 2024, compared with earlier benchmarks. Commercial and fine-tuned models did better, but still below trained forensic analysts. Source: Deepfake-Eval-2024, arXiv 2503.02857, read October 2026
The lesson for anyone using a free detector: a model that scored well on a lab test set can fail on the compressed, cropped, re-shared images you actually find online. No vendor accuracy figure tells you how it will do on your image, which is why we quote none here.
AI photo detector tools compared
Every tool below exists and was checked on its own site on 16 October 2026. Prices are in US dollars and change; check the linked page before you pay.
| Tool | What it checks | Price and limits (checked October 2026) | The catch |
|---|---|---|---|
| SynthID Detector (synthid.com) | Invisible SynthID watermarks in images, video and audio from Google, OpenAI, NVIDIA and Kakao tools | Open to everyone since 7 October 2026; no price shown | Not a general detector: it cannot see images from tools that do not use SynthID |
| Gemini app | SynthID watermarks, plus C2PA Content Credentials (version 2.2 or later) | Free with a Google account; about 10 image checks per rolling 24 hours, files up to 100 MB | One file at a time; it recognises Google AI watermarks only |
| openai.com/verify | OpenAI's SynthID watermark or an OpenAI C2PA manifest in images and audio | OpenAI says its provenance API checks are free with usage limits | Built for OpenAI output only; images made before signals existed are missed |
| Content Credentials Inspect (verify.contentauthenticity.org) | C2PA Content Credentials: which tool made or edited the file, and its edit history | Web tool, no price listed; accepts JPEG, PNG, WebP, HEIC, AVIF, TIFF and more | Most files have no credentials yet, and screenshots strip them |
| Hive AI Detector | Classifier for AI images, video, audio and text, with a likely source model when it can tell | Free Chrome extension with no login; the API is billed by usage | A probability score, so false positives and misses both happen |
| AI or Not | Classifier for AI images and text; video, voice and deepfakes on paid plans | Free: 20 image checks. Pro $5 a month: 500. Business $100 a month: 6,600 | Same limits as any classifier |
| Sightengine | Classifier API for AI images and video, aimed at developers | Free: 2,000 operations a month (500 a day). Starter $29, Pro $99 a month | Needs an API call or their demo; no provenance data |
| TinEye | Reverse image search: where else the image appears and its oldest copy | Free for non-commercial searches | Says nothing about AI on its own; it finds the trail |
| Google Lens and Search | Reverse search, and since May 2026 SynthID checks when you ask "Is this made with AI?" | Free | C2PA checks in Search were still listed as coming |
Sources: SynthID Detector, Gemini Apps Help, OpenAI Help Center, Content Credentials, Hive, AI or Not pricing, Sightengine pricing, TinEye help and Google. A third-party extension called "HIVE AI Detector for Chrome" is not made by Hive; install the one Hive links.
Mistakes that lead to the wrong answer
- Checking a screenshot instead of the original file
- Reading "no watermark found" as "real"
- Trusting one classifier score without a second check
- Calling a beauty-filtered selfie AI because the skin looks smooth
- Ignoring a lightly AI-edited real photo, which is still edited
- Publicly accusing a person on visual tells alone
- Uploading a private or intimate image of someone to a detector you have not vetted
Detection also cuts the other way. A real photo lightly edited with an AI tool can carry a watermark, and a fully generated image passed through enough edits may carry none. OpenAI's help page says a watermark does not show how much of an image a model made, or who is responsible for it.
If the AI photo is of you
Save the evidence before anything else: the image, the link, the account and the date, plus a SynthID or Content Credentials result if you get one. Then report it on the platform. Whether the law helps depends on where you live; our deepfake laws map covers the EU, UK, Australia and US states, and the EU's labelling duty is explained in EU AI Act deepfake labeling.
For creators the rule runs the other way: never make an image of a real person without their consent. Our AI face swap guide covers consent and the law, and if you build a fictional AI persona, our Gemini AI photo prompts show how Gemini marks every image it makes. Building that persona openly, with a consistent face, disclosure and a content plan, is what the AI Influencers program teaches.
How to tell if a photo is AI: FAQ
How can I tell if a photo is AI-generated?
Check provenance before pixels. Upload the original file to a watermark checker such as synthid.com or openai.com/verify, inspect its Content Credentials at verify.contentauthenticity.org, reverse-search it to find the earliest copy, then look for visual tells such as broken text, odd hands and mismatched reflections. A found watermark is strong evidence; finding nothing proves little on its own.
Is there a free AI photo detector?
Yes. SynthID Detector at synthid.com opened to everyone on 7 October 2026 and checks for watermarks from Google, OpenAI, NVIDIA and Kakao tools. The Gemini app gives signed-in users about 10 image checks a day, and Hive's Chrome extension needs no login. AI or Not's free plan includes 20 image checks and Sightengine's free plan 2,000 operations a month. Checked October 2026.
How accurate are AI photo detectors?
It depends on the image, and no vendor score transfers to your photo. Researchers behind the Deepfake-Eval-2024 benchmark found open-source image detectors lost 45% of their AUC on deepfakes that circulated online in 2024, compared with older benchmarks. Commercial tools did better but still fell short of trained forensic analysts. Treat a classifier score as one clue, never as a verdict.
Can people spot AI photos by eye?
Not reliably. In a study from Microsoft's AI for Good Lab, over 12,500 people made about 287,000 judgements of real and AI images and got 62% right overall, a little better than chance. They did best on portraits and worst on natural and urban landscapes. Visual tells help, but newer models fix many of them, so use them alongside provenance checks.
What does it mean if no watermark is found?
Only that the checker did not find its own signal. Google says a SynthID "not detected" result does not rule out other AI systems, and OpenAI lists stripped metadata, compression, cropping and older or unsupported tools as reasons its checker can miss its own images. Screenshots and re-uploads remove Content Credentials too. Keep looking at source and context.
What should I do if someone made an AI photo of me?
Save the image, the link, the account name and the date first, then report it to the platform under its impersonation or synthetic media rules. Run it through SynthID Detector and save the result as evidence. Depending on where you live, deepfake, privacy or intimate-image laws may apply, and some platforms and police treat non-consensual intimate imagery as an urgent report.