AGORA

Media literacy

How can you tell if a photo or video is AI-generated?

By the Agora desk · Published · How we count player splits

Often you can't tell from the picture alone, and AI detection tools can't reliably tell either. The dependable checks are about the image's history: reverse-search it to find where it first appeared, look for other photos or video of the same moment, and check for Content Credentials or a known AI watermark. Odd hands, garbled text and waxy skin are clues worth noticing, not proof.

In short

  • Visual tells are fading as AI tools improve; treat them as clues, not verdicts.
  • Reverse image search to find the earliest, original version.
  • Real events usually have many witnesses; a lone image with no others is a warning sign.
  • Content Credentials can confirm where an image came from, but their absence proves nothing.
  • AI detectors give false alarms and miss fakes, so never rely on one alone.

Practise on today's story · 22 November 2026

Lobby Market Winner: Fully release the Epstein files with only victims names redacted

Should the DOJ release all Epstein files with only victims' names redacted?

▶ Play today's story

Read the briefing first

Start with the story, not the pixels

Before zooming in on fingers, ask whether the scene is plausible. Would this person really be there, doing this? Who posted it first? Has any reputable news outlet, or the person in the picture, shared it? Full Fact's guide to AI images makes the point that real events are usually captured by several people from different angles; one dramatic image that exists nowhere else is a red flag.

Be honest about your own reaction, too. People are far more likely to believe an image that fits what they already think. That works in both directions: researchers also warn that the mere possibility of AI fakes lets people dismiss genuine photos and videos as "AI" when they're inconvenient.

Trace the image with a reverse image search

A reverse image search looks for copies of a picture across the web. It's the single most useful check, because it can show you where an image first appeared and whether it has been around for years.

  • Use Google Lens, TinEye, Bing Visual Search or Yandex. Upload the image or paste its address.
  • Crop to the most distinctive part (a face, a sign, a building) if the full image finds nothing.
  • Look for the earliest copy. If it traces back to an AI art account, or to a different event years ago, you have your answer.
  • For video, take screenshots of a few key frames and search those.
  • If the creator is findable, check their page: many AI artists say how an image was made.

Check for Content Credentials and watermarks

Content Credentials are a kind of "nutrition label" for media, based on an open standard from the Coalition for Content Provenance and Authenticity (C2PA), whose steering committee includes Adobe, the BBC, Google, Microsoft, OpenAI and Sony. Some cameras, editing apps and AI generators attach a signed record of where a file came from and how it was edited. You can upload a file to a Content Credentials verify tool to read it.

There's a catch: social platforms commonly strip this kind of metadata when files are uploaded, so most images you see online won't carry any. "Durable" credentials try to fix that with an invisible watermark that lets the record be looked up again. Google adds its own invisible watermark, SynthID, to content made with its AI tools, and offers a public SynthID Detector to check for it.

So read these signals in one direction only. Valid credentials or a detected watermark tell you something real about the file's origin. No credentials and no watermark tell you almost nothing: plenty of real photos have none, and plenty of AI tools don't add them.

Look for the classic glitches (but don't over-trust them)

Today's AI still makes mistakes, and several together are a stronger signal than one. But each is fading as the tools improve, and real photos can look odd too.

  • Hands, teeth and accessories: extra or merged fingers, too many teeth, earrings that don't match, chains that melt into skin.
  • Text: garbled lettering on signs, posters, packaging and badges. (Text can be added afterwards, so clean text doesn't rule AI out.)
  • An "aesthetic sheen": skin and surfaces that look too smooth, glossy or airbrushed.
  • Backgrounds: objects that blend into each other, warped lines, repeated faces in a crowd.
  • In video: lips that don't close properly on "m", "b" and "p" sounds, glitches at the edge of the face when the person turns or a hand passes in front, and expressions that don't match how the person usually talks.
  • In audio: odd phrasing or pronunciation, robotic pauses, and background noise that doesn't change naturally with the speech.

Why AI detectors can't settle it

Upload-and-check AI detectors are tempting, but they're unreliable. When researchers from the human rights group WITNESS tested free, public detectors in 2024 and reported the results for the Reuters Institute, simply cropping or resizing an AI image was enough to make one tool call it real, screenshots and compression threw results off, and AI edits to real photos, such as adding a tank, were judged authentic. A "no AI detected" result doesn't prove an image is genuine.

False alarms happen too. In October 2026 Full Fact found that Google's public SynthID Detector had labelled five photos posted by politicians as made or edited with Google AI, though the people who posted them said they were genuine and Full Fact found no sign of manipulation. Google updated the tool so those images came back as "Unsure". Full Fact's conclusion applies to every detector: useful as one signal, not to be relied on alone.

A five-step routine

Put together, a sensible check takes a few minutes. If an image is still unverified at the end, the honest label is "unverified", not "fake" or "real", and the safe move is not to share it.

  • 1. Ask whether the scene is plausible and who posted it first.
  • 2. Reverse image search it and find the earliest copy.
  • 3. Look for other photos, video or reports of the same moment from reliable outlets.
  • 4. Check for Content Credentials or a watermark, remembering that absence means little.
  • 5. Only then look closely for glitches, and treat any detector result as one clue among several.

Common questions

Can AI detectors tell if an image is AI?

Not reliably. Tests have shown simple edits like cropping can fool them, and genuine photos have been wrongly flagged. Use them as one clue at most, alongside reverse image search and checking the source.

If an image has no Content Credentials, is it fake?

No. Most images online have no credentials, often because platforms strip the data on upload. Credentials are meaningful when present and valid; their absence tells you very little.

How can I tell if a video or voice clip is a deepfake?

Find the highest-quality version, search for the original source, and check whether reliable outlets report it. Watch for poor lip-sync and glitches around the face; listen for odd phrasing and unnatural pauses. Experts find audio especially hard to judge, so treat a surprising clip as unverified until you find a trustworthy source.

What if someone says a real photo is AI?

Apply the same checks. A real photo will usually have an earlier history, other witnesses and coverage from reliable outlets. Calling something "AI" is an easy way to dismiss inconvenient evidence, so it needs checking as much as the image does.

Sources and further reading

Put it into practice: latest stories

Every story

Written by the Agora desk from the sources above. See our editorial standards. More Learn articles.