Level : Beginner
How to spot an AI-generated image: a 4-step method
By The Info Agent IA editorial team ·
To spot an AI-generated image, do not rely on your eyes alone: check the source first, then look for the marks some generators leave behind (Content Credentials, a SynthID watermark) and run a reverse image search. None of these methods is foolproof on its own, but together they usually give you a solid answer.
Why it has become so hard
Image generation has improved enormously. The errors that used to give a fake away, such as mangled hands or gibberish text, are becoming rare. Our guide to generating images with AI shows how easy it now is to produce a realistic photo in seconds.
A fake picture of a flood, a protest or a celebrity can spread across social media within hours. When it shows a real person, it is called a deepfake. The same logic applies to video, as explained in our guide to creating videos with AI.
The good news: several major companies now add origin marks that you can check yourself, for free.
The 4-step method
| Step | What you look for | Tool | What it proves |
|---|---|---|---|
| 1. Context | Who posted the image, when, with what source | Common sense | Nothing alone, but guides the next steps |
| 2. Content Credentials | An origin “label” attached to the file | Content Credentials site | The declared history, if present |
| 3. SynthID watermark | An invisible mark in the pixels | Gemini, Google | Creation or editing by Google AI or a partner |
| 4. Reverse search | Other appearances of the image online | Google Lens or similar | Likely origin and date |
The rule is simple: a positive result at step 2 or 3 is a strong clue. A negative result never proves that an image is real.
Step 1: check the source and context
Before using any tool, ask yourself three questions.
- Who is posting it? A well-known news outlet, an official body, or an anonymous account created last week?
- Is a source given? A photographer, an agency, a specific place and date?
- Does it fit what is known? A major event is almost always covered by several independent sources.
A dramatic, unsourced image designed to make you angry or afraid always deserves a check. Our article on how to fact-check an AI answer walks through the same habit for text.
Step 2: read the Content Credentials (C2PA)
Content Credentials are a kind of digital label attached to a file. They describe how an image was created and edited. According to the C2PA website (accessed October 22, 2026), the Coalition for Content Provenance and Authenticity develops this open standard, and its steering committee includes Adobe, Google, Meta, Microsoft, OpenAI, Sony and the BBC. The C2PA compares this information to a “nutrition label” for content.
In practice, the Content Credentials website presents a small “CR” pin that signals the presence of this information. It also offers a Verify tool: you drop an image in and, if it carries Content Credentials, you see the declared creation method and editing history.
Two important limits:
- This information is metadata. The C2PA FAQ acknowledges that it is typically embedded in the file but can be separated from it. A screenshot or a re-upload through a messaging app often strips it.
- Not every generator or camera adds it.
Step 3: look for a SynthID watermark
An AI watermark is an invisible mark embedded in the image itself rather than next to it. According to Google DeepMind’s SynthID page (accessed October 22, 2026), the watermark covers images, video, audio and text, and is designed to withstand cropping, filters and compression.
Google offers two ways to check it:
- In Gemini: according to the Gemini Help Center (accessed October 22, 2026), you go to gemini.google.com while signed in, add the file and ask “Was this created or edited by Google AI?”. The page states a limit of 100 MB per file and one file per check.
- With the SynthID Detector portal, described on the DeepMind page, which according to Google recognizes content from its own tools and from partners such as OpenAI, NVIDIA and Kakao. According to Google, this can therefore include images made with OpenAI tools such as ChatGPT.
Step 4: run a reverse image search
A reverse image search means giving a search engine an image to find where else it appears. According to Google Search Help (accessed October 22, 2026), on a computer you click “Search by image” on google.com, then upload the file or drag it into the search box. In Chrome, right-clicking an image also offers “Search with Google Lens”.
What you are looking for:
- The earliest appearance: if the picture was already online years before the event it supposedly shows, it has been repurposed.
- The original source: an artist’s account saying “made with AI,” a stock photo site, a news outlet.
- Other versions: larger, sharper, or still carrying their metadata.
Visual clues: useful, never conclusive
Looking closely still helps, as long as you do not treat it as proof. The most common signs:
- Hands, teeth, ears: extra fingers, shapes that melt together.
- Text: signs, labels or license plates with garbled letters.
- Backgrounds: blurry, distorted people, impossible buildings.
- Light and shadows: shadows falling in different directions, missing reflections.
- Too-perfect look: plastic-smooth skin, a composition that feels staged.
What about automatic AI image detectors?
Plenty of websites promise to tell you in one click whether an image is artificial. They work like AI text detectors: they estimate a probability from what the image looks like.
Their verdict can be one clue among others, never proof, especially when it comes to accusing someone.
Check yourself You receive a shocking photo of a fire in your town. Gemini finds no SynthID watermark. Can you share it?
Not yet. The absence of SynthID only shows the image did not come from Google’s tools. Use a reverse image search to find who posted it first, and check whether local news or the authorities mention the fire. Without a reliable source, do not share it.
What the rules say
In the European Union, the AI Act requires providers of generative AI to make their output identifiable as generated, and requires people who publish a deepfake to disclose it. According to our watermark glossary entry, checked on October 22, 2026, the deadline for the labeling obligation was set at December 2, 2026, after a delay. So many images circulating today carry no mark at all, and rules differ outside Europe. For questions about rights over images, see our explainer on AI and copyright.
What to do with a suspicious image
- Do not share it until you have found a reliable source.
- Report it to the platform if it is misleading or harms someone.
- Be wary of money requests that come with a photo, for example of an injured relative or a celebrity endorsing an investment. Our guide to AI scams covers these scenarios.
To brush up on the technical terms in this article, our AI glossary explains them in plain English.
Frequently asked questions
Is there a tool that can tell for sure whether an image is AI-generated?
No. The reliable tools only check for a mark left by some generators, such as a SynthID watermark or Content Credentials. If no mark is found, the image may still come from an AI. You need to combine several clues, starting with the source.
How do I check an image with Gemini?
Go to gemini.google.com while signed in, add the file and ask whether it was created or edited by Google AI. Gemini then looks for a SynthID watermark. A negative result only means the image did not come from Google's tools.
Are extra fingers still a giveaway?
Less and less, because recent generators make fewer of these mistakes. A visible flaw is still a warning sign, but a flawless image is not automatically real. Visual clues support a check; they never replace it.
Does a screenshot erase the proof of origin?
It can strip metadata such as Content Credentials, which are attached to the file. An invisible watermark embedded in the pixels, like SynthID, is designed to survive edits better. Always try to find the original version of the image.