Jump to a section
The short answer

Don't rely on your eyes: in a Which? test, 70% of people couldn't correctly identify every real and fake video they saw. Instead, ask who posted it, whether it's too good to be true, whether it's rushing you, and whether you can find it anywhere else. For images and video, free tools can sometimes check for AI watermarks, but a clean result doesn't prove something is real.

The honest answer: our eyes aren't enough

In May 2026, Which? showed 514 UK adults a mix of real and deepfake videos. 70% didn't correctly identify all the ones they saw. Which? concluded that 'more often than not, people cannot reliably spot AI fakery'.

The UK regulator Ofcom says the same thing from the other direction: 'users should not be left to identify deepfakes on their own'.

So if you've ever been fooled, you're in good company. The fix isn't sharper eyes. It's better questions.

Four questions that work better

These work whether something was made with AI, edited the old-fashioned way or is simply taken out of context.

  1. Who posted it? Which? says checking the profile behind a post is important. A famous face in a video posted by an account you've never heard of is a warning sign.
  2. Is it too good to be true? As Which? puts it, 'Anything that sounds too good to be true is often a red flag.' The same goes for anything perfectly shocking or perfectly heartwarming.
  3. Is it pushing me to act? If it wants a click, a payment or a share right now, slow down. That's how scams work.
  4. Can I find it anywhere else? If something big really happened, trusted news sources will have it. For images, a reverse image search with Google Lens or TinEye may turn up the original, Which? suggests.

AI chatbots make things up. Here's how to check what they tell you

The visual clues, as a bonus

Visual clues are still worth knowing, as long as you treat them as extra evidence rather than proof. Which? suggests looking for:

  • Distorted hands, shifting logos, unnaturally sharp eyebrows or overly smooth skin
  • Someone who struggles to look straight at the camera
  • Shadows, or reflections in the eyes, that don't match the scene
  • Expressions that don't fit, like smiling through serious news
  • Details that change or vanish between frames, such as background objects, or hair or eye colour

AI and old family photos: what it can do, and what to watch out for

Ask Gemini to check for a watermark

Google adds an invisible watermark, called SynthID, to content made with its own AI tools. You can ask Gemini to look for it. Upload the image, video or audio at gemini.google.com and ask: 'Is this AI-generated?'

If Gemini finds a SynthID watermark, Google says all or part of the content was made or edited with Google's AI. If it finds none, that does not prove the content is real or rule out Google's or another provider's AI. Editing and limited detection can make results unclear.

Upload the file, then ask
Is this AI-generated? Check for a SynthID watermark and any Content Credentials, and tell me what you can and can't tell from them.

Content Credentials: the 'nutrition label' for images

Content Credentials are an industry standard for recording where an image or video came from and how it was edited. They're developed by a group called the Coalition for Content Provenance and Authenticity, or C2PA. Google says Gemini can read them too.

They're useful when they're there. But Ofcom notes that watermarks can often be removed after basic edits, so a missing label tells you very little. In Ofcom's research, 85% of adults wanted platforms to label AI-generated content, but only 34% had ever seen such a label.

What each check can and can't tell you

No single check is definitive. Used together, they're good enough for everyday life.

Ways to check whether something is AI-made
CheckWhat it can tell youWhat it can't
The four questionsWhether to trust it enough to act on itExactly how it was made
Visual cluesSometimes, that something's been generated or editedThat it's real
Reverse image searchWhere else it's appeared, and possibly the originalAnything, if it's brand new
Gemini and SynthIDWhether Google's AI made or edited itWhether another company's AI did
Content CredentialsIts recorded history, when presentMuch at all, when missing

When you're still not sure

Then don't share it, don't pay and don't act on it until you've checked with a source you trust. Waiting costs nothing. If it's a video of someone famous selling an investment or a deal, our scams guide explains what's going on.

AI scams explained: what's changed, and the habits that still beat them

AI voice scams: a simple plan to protect your family

Make and edit pictures with ChatGPT and Gemini

What about AI-written text?

Text has no equivalent of a reliable visual tell. ChatGPT itself is not a dependable authorship detector: OpenAI says it may make up an answer when asked whether it wrote an essay or whether text was AI-generated.

A detector score can be a reason to investigate, but it is not proof. Look for the original source, working notes, drafts and whether the factual claims hold up. Do not accuse a child, student or writer on the strength of a percentage alone.

For a suspicious message, the practical question is what it wants you to do. Verify a payment or urgent request through a known contact even if the wording looks human.

Can teachers tell if you used ChatGPT? Do AI detectors actually work?

AI scams explained: what's changed, and the habits that still beat them

Four questions before you believe or share

  • Who posted it? Check the account, not just the face in the video
  • Is it too good, too shocking or too perfect to be true?
  • Is it pushing me to act, buy, pay or share right now?
  • Can I find it from a source I trust, or an older original version?

Sources

These sources support the guide. Tool settings, prices and policies change, so dated source boxes show when those details were checked.

AI helps with research and drafting. Dan checks the sources and is responsible for what is published.

How we check