If you want to know how to spot AI-generated images in 2026, start with the honest answer: looking closely at the picture is no longer the reliable test. Counting fingers, hunting for garbled text and checking whether the shadows fall the right way still catch plenty of fakes, but they now catch the lazy ones only. The check that holds up is to stop examining the image and start examining where it came from: who posted it, whether it carries a Content Credential, whether it shows up anywhere else, and whether any trustworthy source is reporting the same thing.
That shift is the single most useful thing to teach a child, because it survives the technology improving. Visual tells go out of date every few months. "Who is telling me this, and where else does it appear?" does not.
Do the visual tells still work?
Some, some of the time. They are worth knowing, and our spot the fake guide lists them in full for images, video and cloned voices. The short version for images:
- Hands, teeth and ears, where extra or merged detail still creeps in
- Text and logos, which often come out subtly wrong
- Backgrounds, where objects melt into each other or patterns fail to line up
- Skin and light, when skin looks plastic or shadows disagree with each other
Use them as a prompt to look harder, never as proof. A clean image is not evidence of a real one, and a slightly odd image is not evidence of a fake. Real photographs contain compression artefacts, odd lighting and strange hands too.
The check that actually holds up: provenance
Provenance means the record of where a file came from. Two systems are worth knowing about, and neither requires any technical skill from you.
Content Credentials, the first of the two. The Coalition for Content Provenance and Authenticity (C2PA) is an open standard backed by a steering committee that includes Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Sony and TikTok. It attaches a tamper-evident record to a file describing how it was made and what was edited. C2PA describes it as being like a nutrition label for digital content. You can check one for free by dragging an image onto the Verify tool.
SynthID, the second. Google's watermarking system embeds an invisible marker directly into images, audio, text and video made by its own generative AI products, including the Gemini app. Google says it stands up to modifications such as cropping, filters and lossy compression, which is more than can be said for ordinary metadata. There is a SynthID Detector portal, though it is currently in early testing with journalists and media professionals via a waitlist rather than open to everyone. In the meantime you can ask Gemini directly whether an image carries a SynthID watermark.
Now the limitation, because it is a big one. Both systems only tell you something when the tool that made the image took part. Plenty of AI image generators attach nothing at all, and a credential can be stripped by re-saving or screenshotting a file. So a positive result is genuinely informative, and a blank result tells you almost nothing. "No credential found" does not mean "real".
What to do when the tools come back empty
This is where most images land, so this is the part to practise. Three questions, in order:
- Where did this first appear? An image that exists only on one unfamiliar account, with no trail behind it, deserves no benefit of the doubt.
- Is anyone credible reporting it? If a photograph would be genuinely newsworthy and no established news organisation is carrying it, that absence is the answer.
- Does a reverse image search find it elsewhere? Google Lens or a right-click search will often surface the original, an older version, or a fact-check.
Our Is it real? tool walks a child through this sequence in a few taps, which is deliberately easier than remembering a checklist under pressure.
How to teach this to a child without frightening them
Children are often better at the mechanics than we are and worse at the pause. The pause is what you are teaching.
- Make it a game, not a warning. Look at three images together and guess, then check. Being wrong is the useful part.
- Give them one sentence to hold onto: if it makes me feel something strongly, check it before I share it. Strong feeling is what fakes are engineered for.
- Teach the source habit rather than the tell-tale list, because the list expires and the habit does not.
- Agree that sending you an image to check is never telling tales and never gets them in trouble. That single agreement does more than any tool.
For younger children, our AI guide for parents sets out what to say at what age, and where the conversation is genuinely worth having early.
When the fake is of your child
This is the version that matters most, and it is happening in the UK now. The Internet Watch Foundation reported on 10 August 2026 that by the end of June 2026, 268 actioned Report Remove reports involved suspected faked imagery, already more than the 221 recorded across the whole of 2025. IWF Chief Executive Kerry Smith said the figures suggest reports involving suspected AI-generated or otherwise manipulated imagery "are becoming an increasingly prominent threat".
If a fake sexual image of your child, or of a child they know, appears anywhere: it is not their fault, it is illegal, and there is a route to getting it taken down. Report Remove, run by the IWF with Childline, exists exactly for this and is free to use. Our guide to AI nudify apps and deepfakes sets out the steps, including what to keep and what not to do first.
The short version
Check the source before you check the pixels. Use the Content Credentials Verify tool and SynthID when they help, and treat a blank result as no information rather than reassurance. Teach your child the pause, not the tell-tale list. And make it easy for them to bring you something they are unsure about, because that is the one check that never goes out of date.