How to tell if an image is AI-generated, tested on 8 AI images

To tell if an image is AI-generated, check its provenance first and its looks last. Upload it to Content Credentials Verify, read its metadata, and ask the maker’s own watermark checker. Then run a reverse image search. We ran 8 real AI images through these checks. Metadata flagged 7 of them. Once we stripped the metadata, it flagged none.

What should you check first? A 6-step quick check

Start with the checks that can give a firm yes, and keep the guesswork for last. Here is our quick check, from the strongest signal to the weakest:

1. Content Credentials. A signed C2PA record names the tool that made the picture.
2. The metadata. A made-by-AI tag, a generator name or a saved prompt in the file.
3. A watermark check. Google’s SynthID, read by the Gemini app, for images made with Google AI.
4. Where it came from. A reverse image search finds the first post and the context.
5. An AI detector. Hive or Illuminarty scores the pixels. A second opinion, never proof.
6. How it looks. Hands, text, reflections and lighting. Useful, but easy to get wrong.

Always check the original file when you can get it. A screenshot or a re-upload to a chat app usually drops the metadata, so the first three checks have less to read.

How to check Content Credentials (C2PA)

Content Credentials are a signed record inside the file, built on the C2PA standard. The free Content Credentials Verify page from the Content Authenticity Initiative reads them. It read 4 of our 8 AI images: ChatGPT, Gemini, Google AI Studio and Ideogram.

1. 1

   Open Content Credentials Verify

   Go to contentcredentials.org/verify. You need no account. The page takes JPEG, PNG, WebP, AVIF, HEIC and more.

   

2. 2

   Drop the image and read the summary

   Drop the file on the page. Our ChatGPT image showed “This image was generated with an AI tool”, issued by OpenAI, with the actions Created and Format converted.

   

3. 3

   Check who issued the credential

   Read the Issued by line before you trust the summary. Our Ideogram file said it was made with an AI tool, but Verify flagged its issuer as Unrecognized. A credential is only as good as the signer behind it.

   

The other 4 files showed nothing in Verify. ComfyUI and both Meta files do carry AI fields, just not C2PA, and Verify reads only C2PA. Leonardo carried nothing at all.

How to check an image’s metadata for AI

Read the file with ExifTool, a free metadata reader, and look for three things. The IPTC field DigitalSourceType set to trainedAlgorithmicMedia, which the IPTC vocabulary defines as AI-made media. A C2PA record with a tool name. Or a saved prompt in a PNG. Here is what our 8 files carried:

Here is the real run on two of our files. The ChatGPT image names its maker; the Leonardo image prints nothing:

A real camera photo looks different. Our iPhone photo carried 150 tags, among them the make and model (Apple iPhone 12 Pro) and the lens, and no AI signal. Camera fields can be faked or removed too, so treat them as a hint, not proof. To see what else a photo carries, read how to remove EXIF data.

How to check for an invisible AI watermark

Ask the Gemini app. Google’s help page says to sign in at gemini.google.com and click Add files. Upload the image and ask “Is this AI-generated?”. Gemini looks for SynthID, the watermark Google puts in the pixels. The check works one file at a time, up to 100 MB, and about 10 images a day.

Read the answer carefully. Google says a detected watermark means Google AI made or edited the file, in whole or in part. “No watermark detected” only rules out Google AI; another tool may have made it. Google’s SynthID page says the mark is designed to survive cropping, filters and lossy compression.

Our files back that up. 3 of the 4 C2PA manifests declare a watermark: both Google files name SynthID, and the ChatGPT one records a watermark action. OpenAI’s help page says ChatGPT images carry SynthID too. So a ChatGPT or Gemini image can keep its watermark after its metadata is gone.

How to spot AI images by eye

Look at the parts generators still get wrong, and use what you find only as a hint. People are not good at this.

What generators still get wrong

• Text. Signs, labels and logos with letters that do not form words.
• Hands and small details. Extra or merged fingers, earrings that do not match, teeth that blur together.
• Physics. Reflections and shadows that point the wrong way, or a pattern that melts into the background.
• Objects that cannot work. A teapot with no handle, a door that leads nowhere.
• Too clean. Skin with no pores, perfect light everywhere, no noise in a dark corner.

What our 8 teapots showed up close

We zoomed into all 8 of our own AI images. 3 had a tell we could point at. The ComfyUI teapot has no handle and a lit candle inside. The Ideogram teapot has no handle either. The Leonardo pot is not a lighthouse at all, and its painted figure is garbled.

The other 5 looked clean to us, even zoomed in. ChatGPT, Gemini, Google AI Studio, Meta AI and Muse drew believable handles, windows and stripes. None of our images has hands or text, so those tells had nothing to show.

How well do people spot AI images?

Not much better than a coin toss. Microsoft’s AI for Good Lab studied its Real or Not quiz: about 287,000 image ratings by over 12,500 people. The study reports an overall success rate of 62%. People did best on portraits and worst on landscapes.

Use a reverse image search to find the source

Search for the image itself to find where it first appeared. A real news photo usually shows up on several sites, with a photographer’s name and a date. An AI image often traces back to one account, a prompt-sharing site or a post that already says it was AI-made.

TinEye

TinEye needs no account. Upload the image and it lists every page where it found a copy. We searched our ChatGPT teapot. TinEye found 0 matches after searching over 86.4 billion images. That is the honest answer for a new image: no history yet, which proves nothing either way.

Google Lens and About this image

Google Lens searches by image too. Google’s help page says its About this image panel shows when Google may first have seen the image. It can also say whether the image was made or edited with AI. The feature is only in some regions. We could not test Lens here: the upload ran into a robot check.

Read around the image

Who posted it first, and when? Does the place, the weather and the event match the day? This is how fact-checkers work, and it still works when the file has been stripped and re-shared.

Are AI image detectors reliable?

Treat any AI image detector as a second opinion, not a verdict. A detector guesses from the pixels and gives a score. We ran our 8 AI images and 1 real photo through the 2 free detectors that need no account. They disagreed badly.

Hive caught all 8

Hive’s demo runs without signing in. It called all 8 AI files likely AI-generated, each at 99.9%. It even named the generator: “lucid” for our Leonardo Lucid Origin file, which carried no metadata. Our real photo scored 10.9%, marked not likely AI.

Illuminarty missed 7 of 8

Illuminarty’s free plan also runs signed out. It put only the Leonardo file over 50%, at 72.8%. The ComfyUI teapot with no handle scored 9.6%. Our real photo scored 40.8%, higher than 7 of the 8 AI images.

Detectors we could not test

AI or Not opened a Create Account form on the first upload, so we did not run it. Its maker claims 98.9% accuracy. Is It AI offers 3 free checks, but it held the Analyze button behind a robot check. Its maker claims 95%+ accuracy. Neither claim is ours, and our test of 2 other detectors shows how far scores can drift.

Meta says the same about its own labels. Its labeling post says it is “not yet possible to identify all AI-generated content”. It also says people can strip invisible markers. Ordinary edits such as compression and upscaling can change what a detector sees.

What it means when no check finds anything

It means you do not know yet. Our Leonardo image is AI-made, and it carried no signal in the metadata. Content Credentials Verify could not read it either:

So go back to the source. Ask the person who shared the picture for the original file, and check where it was first posted. For school or a newsroom, write down which checks you ran and what each one found, and say plainly that the result is not certain.

How we tested

The files

• The AI images: 8 images the owner made on 24 October 2026 from one prompt, a lighthouse teapot. The tools: ChatGPT, Gemini, Google AI Studio, Ideogram, ComfyUI, Meta AI, Meta’s Muse and Leonardo. Every file as downloaded, unchanged. The same files as our guide to removing AI metadata.
• The real photo: one photo taken with an iPhone 12 Pro, saved as HEIC by the phone.
• The stripped copies: a copy of each AI image with ExifTool’s -all= applied, read again the same way.

The checks

• Metadata: ExifTool 13.59 on 2026-10-26. A file counts as flagged when ExifTool prints one of 5 things: a C2PA store, trainedAlgorithmicMedia, a PNG prompt, a Facebook ID or a publisher.
• Content Credentials Verify: the public page at contentcredentials.org/verify, run in Chromium through Playwright on 2026-10-26, no account. A file counts when the page shows a credential. Every upload, here and below, was one of these test images.
• Detectors: Hive’s public demo and Illuminarty’s free plan (version 1.4), signed out, on 2026-10-26. Each read the 8 AI files and 1 real photo, the owner’s iguana, a 1920 × 1280 JPEG. We record Hive’s AI-generated score and Illuminarty’s AI probability as printed.
• Reverse search: TinEye on our ChatGPT file, no account, on 2026-10-26.
• Visual tells: our own crops of the 8 AI images, viewed at full size.
• The steps: the Verify steps are our own screenshots. The Gemini steps follow Google’s own help page.

Next step

Open Content Credentials Verify and drop the original file. Working with your own AI images? Read what an AI image metadata remover takes out of them.