# How to Tell If an Image Is AI Generated: 9 Checks

> Start with provenance, not pixels: check for Content Credentials, run a SynthID check in Gemini and reverse-search the picture to find where it first appeared. Then look for the classic slips in hands, text, reflections and background objects. No single check is proof, so treat a clean result as unknown rather than real.

- URL: https://fake-ai.com/images-and-video/how-to-tell-if-image-is-ai/
- Language: en
- Updated: 2026-10-19
- Publisher: Fake AI (OnlineBuilders LLC, https://onlinebuilders.biz)

For a 2022 experiment published in PNAS, Sophie Nightingale and Hany Farid showed people a mix of real faces and faces made by a generative model. Average accuracy was 48.2 percent. That is slightly worse than guessing. Participants who got training and feedback climbed to 59 percent, and the synthetic faces were rated 7.7 percent more trustworthy than the real ones ([Nightingale & Farid, PNAS 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8872790/)).

So the honest starting point is this: your eyes are a weak instrument. They still catch a lot of lazy fakes, but the checks that hold up best don't look at the picture at all. They look at where it came from.

## Why staring at pixels is the wrong first step

Most "spot the AI image" lists start with fingers. That made sense in 2022, when models routinely produced six-fingered hands and melted text. Current generators fix most of those slips, and the ones that remain show up mainly in busy scenes, small background details and anything that has to be consistent across the frame.

Visual checks also fail in the other direction. Real photos that were heavily compressed, upscaled or shot on a phone with aggressive night mode can look smeared and "synthetic" too. People flag those as AI all the time, which is how genuine news photos get dismissed.

The order that works better: provenance first, context second, pixels last.

**Checking a suspicious image in the right order**

1. **Content Credentials** Upload the file to a C2PA verify tool. A signed manifest names the camera or app that made it.
2. **Watermark check** Ask Gemini whether the image was made with Google AI. This reads the invisible **SynthID** mark.
3. **Reverse search** Google Lens, Bing Visual Search and TinEye show older copies and the first known upload.
4. **Context** Who posted it first, when, and does anyone else have the same scene from another angle?
5. **Pixels** Only now: hands, text, reflections, shadows, repeated patterns, background people.

## Check 1 and 2: provenance signals that can actually prove something

**Content Credentials.** The C2PA standard lets cameras and phones, as well as editing apps, attach a signed record to the file, a kind of nutrition label listing the device, the edits and whether a generative tool was involved. Since the Pixel 10, Google signs every photo taken with the Pixel Camera app, and Google Photos attaches credentials to images edited with its AI tools. OpenAI and Adobe sit on the C2PA steering committee, as do Microsoft and Meta. How the signatures work, and why they so often go missing, is covered in our [guide to Content Credentials](https://fake-ai.com/images-and-video/content-credentials-c2pa/).

**SynthID.** Images and video made or edited with Google's models carry an invisible watermark. In the Gemini app you can upload a file up to 100 MB and ask whether it was created with Google AI. A hit means all or part of the image came from Google's tools. A miss only means it didn't come from Google: Google itself notes that it "could have been created by other AI systems" ([Gemini Help](https://support.google.com/gemini/answer/16722517)).

The pattern for both: a positive result is strong evidence, a negative result is close to worthless. Most images on social media arrive stripped of metadata, and most generators outside the big platforms add no watermark at all.

## Check 3 and 4: reverse search and context

Reverse image search answers the question that matters most in practice: has this picture existed before today? An "exclusive" disaster photo that turns up in a 2019 stock library is not AI, but it is still fake news. Any picture with zero earlier copies, posted first by an account created last month, deserves suspicion even if it looks perfect. Our walkthrough on [finding the original source of an image](https://fake-ai.com/images-and-video/find-original-image-source/) covers the search engines and their differences.

Context checks are boring and effective. Is there a second photo of the same event from another person? Do the weather, the shop signs and the vehicle plates fit the claimed place and date? Does the account post anything else that looks like it was taken by the same person?

## Check 5 to 9: the visual tells that still work

When provenance comes up empty, look closely. Zoom to 200 percent and go region by region rather than judging the whole picture at once.

**Visual details worth zooming into**

- [ ] **Hands and feet**: count fingers, check where fingers meet objects
- [ ] **Teeth and eyes**: uneven teeth, pupils of different shapes, mismatched catchlights
- [ ] **Text**: signs, labels, licence plates and book spines that dissolve into pseudo-letters
- [ ] **Repeated structures**: fences, window rows, keyboard keys, tiles
- [ ] **Jewellery and glasses**: earrings that don't match, frames that merge into skin
- [ ] **Background people**: faces that blur into each other, limbs that don't belong to anyone
- [ ] **Light**: shadows pointing different ways, reflections that show a different scene
- [ ] **Texture**: skin with a waxy, airbrushed sheen across the whole face

The FBI lists the same weak spots in its December 2024 warning about AI-assisted fraud: "distorted hands or feet, unrealistic teeth or eyes" ([FBI IC3](https://www.ic3.gov/PSA/2024/PSA241203)). Those details matter most for profile photos used in romance and investment scams, which are usually portraits on a plain background where the background tells you little.

One flaw doesn't convict an image. Real photos have odd reflections, and phone cameras invent texture in low light. Two or three independent anomalies in different regions are a stronger signal.

## What AI image detectors can and cannot do

Upload sites that promise "99% accurate AI detection" usually mean accuracy on their own test set, with images from the generators they trained on. Accuracy drops on new models, on screenshots and on images that have been resized or recompressed.

The best public benchmark for video fakes makes the point clearly. Meta ran a Deepfake Detection Challenge in 2020, and the top model reached 82.56 percent average precision on the public data and 65.18 percent on a hidden test set it had never seen. The German federal security agency BSI cites that second number as evidence that automated detectors generalise poorly ([BSI](https://www.bsi.bund.de/EN/Themen/Unternehmen-und-Organisationen/Informationen-und-Empfehlungen/Kuenstliche-Intelligenz/Deepfakes/deepfakes_node.html)).

**What each method can tell you**

| | Can prove AI | Can prove real | Works on screenshots |
|---|---|---|---|
| Content Credentials | yes | partly | no |
| SynthID check | yes | no | partly |
| Reverse image search | partly | partly | yes |
| AI detector score | no | no | partly |
| Visual inspection | partly | no | yes |

Detectors are useful as one more opinion, especially if two different tools agree. They are not useful as a verdict, and they should never be the reason you accuse a person of faking something.

## A realistic routine for everyday images

For a meme in a group chat, none of this is worth ten minutes. For an image you are about to forward to 300 people, publish, or use as a reason to send money, it is.

Five minutes is enough for a sensible routine: run the file through a verify tool and Gemini, do one reverse search in two engines, check who posted it first, then zoom into hands, text and background. If nothing conclusive turns up, write "unverified" next to it or don't share it. That is a perfectly good outcome.

If you work in a newsroom or a company that publishes images at scale, the cheaper fix is upstream: sign your own images and ask suppliers for signed files. Teams that want provenance built into their content pipeline can get help from studios such as [OnlineBuilders](https://onlinebuilders.biz/solutions/ai-systems), which build AI systems with logging and review included.

Video needs a different set of checks because motion gives fakes away in other places. Continue with our guide on [how to spot a deepfake video](https://fake-ai.com/images-and-video/how-to-spot-deepfake-video/), or browse all [image and video guides](https://fake-ai.com/images-and-video/).

## FAQ

### Is there a free tool that tells me for sure whether an image is AI?

No. Free detectors return a probability, and they are trained on specific generators, so a new model can slip past them. Provenance checks like Content Credentials or SynthID are more trustworthy, but they only give a positive answer when the maker added the signal.

### Can AI still not draw hands?

Newer image models draw hands far better than the 2022 generation did, so six fingers are now the exception. Hands are still worth a look, together with jewellery, teeth and anything that has to repeat in a regular pattern.

### Does a missing EXIF block mean the photo is fake?

No. Social networks and messengers strip metadata from almost every upload. Missing EXIF tells you the file has been re-saved somewhere, nothing more.

### What should I do if I can't decide?

Don't share it, and say why if someone asks. An image you cannot place is not evidence of anything, and waiting a day often brings out the original source or a fact-check.

Photo: Heptagon, Public domain (https://commons.wikimedia.org/wiki/File:Magnifying_glass2.jpg)