How to tell if an image is AI-generated
No single test proves an image is AI-generated, but there is a sensible order to check things in. Here is how to trace an image to its source, read the evidence inside the file, use a visual detector without over-trusting it, and weigh what you find.
AI images used to give themselves away with six-fingered hands and melted lettering. Today’s generators, from Midjourney and FLUX to OpenAI’s GPT Image and Google’s Gemini, make those mistakes far less often, and the copies that circulate online have usually been compressed, cropped and re-shared until any data that came with them is gone. Spotting a generated picture by eye is no longer a dependable skill on its own.
What still works is a layered approach. Each check below catches things the others miss, and they run roughly from the strongest kind of evidence to the weakest.
1. Where did the image first appear?
Before you analyze a single pixel, find out where the picture came from. A reverse image search with Google Lens, TinEye or Bing Visual Search shows where else it has been posted, often including earlier and larger copies. Then ask the questions a picture editor would:
- Who posted it first, and do they usually share their own photos, or AI art?
- Did anyone else capture the same scene — other photographers, other angles, news coverage of the event?
- Does the caption claim a time and place you can test against weather, landmarks or known events?
- Has the creator said how it was made? Plenty of AI artists label their work openly.
Source checking is slow, but it is the one test a generator can’t beat by making better pixels. It also tends to turn up a higher-quality copy, which makes every later check more reliable.
2. What does the file say about itself?
Many AI tools write a record into the image file. OpenAI (ChatGPT and DALL·E), Adobe Firefly, Microsoft, Google and others attach C2PA Content Credentials to many outputs: a signed manifest naming the signer, the tool, the actions taken and a digital source type. Other signals are unsigned but still telling:
| Signal | What it looks like | How far to trust it |
|---|---|---|
| Signed Content Credentials | A signer such as OpenAI or Adobe, with the source type trainedAlgorithmicMedia (generated) or compositeWithTrainedAlgorithmicMedia (edited with AI) | High: the maker states it under its own signature |
| IPTC or XMP AI label | The same source-type values, without a signature | Good, but anyone can add or edit it |
| AI tool in the software tags | A generator or AI editor named in EXIF Software or XMP CreatorTool | Good, with the same caveat |
| Embedded prompt | Stable Diffusion, ComfyUI or InvokeAI text in a PNG, holding the prompt and settings | Very telling: cameras don’t write prompts |
| Camera-signed manifest | Content Credentials signed at capture by a camera such as a Leica, Nikon or Pixel | Strong evidence of a real photo |
| Camera EXIF | Make, model, lens and exposure settings | A weak hint, since it is easily copied or faked |
Drop the image into Toolvy’s AI image detector and these show up in the File evidence panel, read in your browser without uploading the file. To see every field, not only the AI-related ones, use the metadata viewer.
3. What does a visual AI check say?
When the file is silent, the pixels are what is left. AI image classifiers are trained on large sets of real and generated images to pick up traces people can’t see. Toolvy’s visual check uses a fine-tuned version of Community Forensics, a detector from a CVPR 2025 paper, and runs it on your device. It scores five regions of the image, the center and four corners, and reports the lowest, so it calls an image AI only when every region looks generated.
Treat its answer as a measured opinion. The model’s developer reports, on its own development data, 90–96% balanced accuracy with about 3–5% of real photos wrongly flagged, and a catch rate of roughly 66–93% on web-compressed images from recent generators. Those figures are not an independent benchmark, and they mean two practical things: some AI images will pass, and a few real photos will be flagged.
The verdict uses plain labels. Likely AI-generated means a high score in every region; Possibly AI-generated a weaker lean; Inconclusive no clear signal; No signs of AI found nothing detected, which is not the same as proof the photo is real. Expect the weakest results on small images (under about 300 pixels), heavily compressed copies, screenshots and photos of screens, real photos with heavy filters, and pictures where AI changed only a small area.
Check an image with the AI image detector4. What should you look for with your own eyes?
Visual tells are less reliable than they were, and real photos have oddities of their own, so treat these as reasons to look harder rather than as verdicts:
- Text and symbols. Signs, labels, logos and book spines that read almost right, or not at all.
- Hands, teeth and ears. Far better than they were, but still worth zooming in on: extra joints, merged fingers, mismatched earrings.
- Light and reflections. Shadows falling in different directions, or reflections in glasses and windows that don’t match the scene.
- Background logic. Railings that dissolve, windows that don’t line up, people in a crowd who blend into one another.
- Texture. Skin, fabric or fur with a uniform, airbrushed smoothness, or detail that repeats like a pattern.
Zoom to 100% or more, and spend as long on the background as on the subject.
5. Is there an invisible watermark?
Some AI providers hide a watermark in the pixels themselves. Google’s SynthID, used in images from its AI models including Gemini, is the best known. Because it isn’t metadata, it can outlast the re-saving that strips Content Credentials, but only Google can read it: Toolvy can’t, and Google offers its own SynthID Detector. A watermark check covers only the provider that made it, so a negative result says nothing about other generators.
How do you put the evidence together?
Rank what you found. A signed AI manifest is close to conclusive about what its signer declared; an unsigned label or embedded prompt is strong; a high visual score is a solid lead; visual tells are suggestive. Confidence comes from independent checks agreeing: a stripped file, a high visual score and an account that posts nothing but AI art make a fairly clear picture. A single weak signal does not.
Frequently asked questions
Can you tell if an image is AI from its metadata?
Sometimes, and nearly conclusively when signed Content Credentials declare it. But many images online have lost their metadata by the time you see them, so finding no AI data is the normal case and proves nothing.
Can a real photo be flagged as AI-generated?
Yes. Heavy filters, smoothing and repeated re-compression can make a genuine photo look generated to a classifier. If the photo is yours, the original file from your camera or phone is the best evidence you have.
Does a screenshot keep Content Credentials?
No. A screenshot is a new image made from whatever was on your screen, so the original file’s metadata, including any manifest, is left behind. That is one more reason to hunt down the original before you check.
Can you check whether text was written by AI, too?
Yes, with the same caution. Toolvy’s AI text detector scores an essay, PDF or Word file passage by passage on your device. AI-text detectors have blind spots of their own, since short, formal, AI-polished and non-native English writing can be flagged, so treat a result as a lead, never as proof.