How to tell if an image is AI-generated
No single test tells you whether a picture is real. You get a reliable answer by putting several small checks together. This guide takes you through them in the order that saves the most time.
1. Pause before you share
Fake images spread because they make people feel something strongly and quickly. If a picture makes you angry, scared or delighted, that is a good moment to stop and check.
Ask yourself three questions. Who posted this? When was it taken? Why am I seeing it now?
2. Find where it came from
This is the most useful step. An image with a named photographer, a news agency or a clear first post is much easier to trust than one that appears from nowhere.
Run a reverse image search. It shows other places the same picture, or similar ones, have appeared online.
- TinEye can sort results by oldest first, which helps you find the earliest copy.
- Google Lens and Bing Visual Search cover more of the web and often find near-matches.
Then check what you find. Is the earliest copy older than the story it is attached to? Does the original caption say something different? Did it first appear on an account that posts AI art?
A real photo with a false caption is one of the most common tricks. The image passes every AI test, but the story is untrue. RELTY calls this misleading context.
3. Look for Content Credentials
Content Credentials are a tamper-evident record attached to a file. They follow the open C2PA standard. They can show which camera or app made the image and how it was edited afterwards.
Upload the file to Content Credentials Verify to read them. Some cameras and phones add credentials to real photos. Some AI image tools add them to the pictures they generate, and the record will say so.
Most images online have no credentials. Many websites and messaging apps strip this information when a file is uploaded, and screenshots lose it too.
Missing Content Credentials do not prove an image is fake. They only mean this check cannot help.
Some AI companies also hide invisible watermarks in their images. The best known is Google's SynthID, which OpenAI now uses too. These need a special checking tool. Google's SynthID Detector looks for watermarks from Google and partners including OpenAI. Finding no watermark proves nothing, and a detector can sometimes be wrong when it finds one.
4. Look closely at the image
View the largest version you can find, then zoom in. AI images often go wrong in the details.
- Hands and teeth. Extra or merged fingers, odd joints, teeth that blur together.
- Text. Signs, labels and number plates with letters that do not form real words.
- Edges and objects. Earrings that do not match, glasses that melt into skin, hair that fades into the background.
- Light and shadow. Shadows that fall in different directions, or reflections that do not match the scene.
- Backgrounds. Faces in a crowd that turn into smudges, buildings with impossible lines, repeated patterns.
- Texture. Skin that looks airbrushed and glossy, with no pores or blemishes.
Treat these as clues, not proof. The newest image generators make far fewer of these mistakes. Real photos can look strange too, because of phone processing, compression, filters or motion blur.
Also check the scene against the real world. Does the weather match reports for that day? Are the shop signs in the right language? Are there other photos of the same event from different angles?
5. Use AI detectors with care
AI detection tools look for statistical patterns left by image generators. Services such as Hive and Sightengine return a score showing how likely they think it is that an image was AI-generated.
Detectors are useful, but they make mistakes in both directions.
- They work best on the original file. Screenshots, crops and heavily compressed copies make them less reliable.
- A new image generator may not be recognised until the detector is updated.
- Different detectors often disagree.
If you use more than one, look at each result separately. Averaging two scores hides the disagreement, and the disagreement is useful information. When detectors disagree, the honest answer is that the result is unclear.
6. Search for existing fact-checks
If an image is going viral, a fact-checker may already have looked at it. Search Google Fact Check Explorer with a few words describing the picture or the claim. In the UK, Full Fact publishes its checks in full.
Fact-checkers who follow the IFCN code of principles show their sources, so you can follow their reasoning yourself.
7. Decide what you can say
Put the evidence together. RELTY uses four verdicts, and they work for your own checks too.
- Verified. Reliable credentials, a trustworthy original source or independent reports support the image.
- AI-generated or altered. Credentials, the creator or strong technical evidence show it was made or changed.
- Misleading context. The image is genuine but is being used to tell a false story.
- Unverified. There is not enough evidence either way.
"Unverified" is a fair and common answer. It is better to say you do not know than to guess. If you cannot verify an image, do not share it as fact.
In the news
Recent reports from RELTY's AI news that show why this check matters.
- 11 November 2026 · Maldita.es
Image of a Madrid bus wrecked by strike pickets was made with AI (in Spanish)
Maldita.es found a SynthID watermark and illegible lettering on the bus. A second picture shared as the untouched "original" carried the watermark too. A supposed original needs checking as much as the fake.
- 5 November 2026 · Full Fact
Fake images of Donald Trump being carried after the US midterms
Full Fact found a SynthID watermark, and matched one picture to the first frame of an older AI video. A wider version showed several photographers, yet no other photos of the event existed. Missing coverage of a big moment is a clue in itself.
- 29 October 2026 · Newtral
AI-generated train photo spread after the Urda crash (in Spanish)
Newtral says a photo of a derailed train in a ravine does not show the accident. It carries a SynthID mark, showing it was created or edited with OpenAI's tools. Fake images appear quickly after real disasters.
- 29 October 2026 · Maldita.es
Netanyahu image naming Spain was altered with AI (in Spanish)
OpenAI's verification tool found a SynthID watermark, and the oldest version Maldita could find did not show the word. Tracing the source and checking for a watermark pointed the same way.
- 21 October 2026 · Full Fact
Google updates its AI image detector after Full Fact investigation
Google's tool labelled five images shared by politicians as made or edited with AI, though the people who posted them said they were not. Detectors can be wrong in both directions.
Sources
The pages this guide relies on, checked on 25 October 2026.
- C2PA technical standard. Coalition for Content Provenance and Authenticity.
- Content Credentials Verify. Content Authenticity Initiative.
- SynthID. Google DeepMind. Checked again on 1 November 2026.
- Content provenance. OpenAI developer documentation. Checked on 1 November 2026.
- Google updates AI detector after Full Fact investigation finds it flagging apparently real images. Full Fact, October 2026. Checked on 1 November 2026.
- Search with an image on Google. Google Search Help.
- TinEye reverse image search. TinEye.
- Fact Check Explorer. Google.
What are Content Credentials and how do you check them?
Why AI detectors disagree, and how to read their results
How to tell if text was written by AI
How to reverse image search a photo
Browse the RELTY directory of tools
Links to other organisations are for reference only. RELTY has no partnership with, or endorsement from, any of them.
