How to spot manipulated images and videos, including AI fakes

TUREFF• •0

Editing photos is as old as photography, but tools that can generate a convincing image, voice or video from a short text prompt have made fakes cheaper and faster to produce. No single trick will catch them all, and detection tools are not reliable on their own. The most dependable approach is to combine a few checks on the content with a few checks on where it came from.

Common kinds of manipulation

Miscaptioned or recycled media. A real photo or video from a different time or place shared as if it shows a current event. This is still one of the most common forms and needs no editing at all.
Selective editing. A clip cut to remove context, or a photo cropped to leave something out.
Traditional editing. Objects or people added, removed or altered with editing software.
Synthetic media. Images, video or audio created or heavily altered by AI, including cloned voices and face swaps, often called deepfakes.
Slowed, sped-up or dubbed clips that change how someone appears to speak or behave.

Look at the content closely

AI tools improve quickly, so treat these as clues rather than proof. Their absence does not mean an image is real.

Hands, teeth, ears and jewellery have often been rendered oddly, with extra or merged parts.
Text in the scene, such as signs and labels, may be garbled or inconsistent.
Lighting and shadows that fall in different directions, or reflections that do not match.
Backgrounds that blur strangely, repeat or contain impossible structures.
In video, watch for blinking, lip movement out of step with speech, edges of the face flickering, and skin that looks too smooth.
In audio, listen for flat emotion, odd breathing or pauses, and unusual pronunciation of names.

Check where it came from

This is usually more productive than staring at pixels.

Find the earliest version. Use a reverse image search, or take screenshots of key frames from a video and search those. If the same picture appears years earlier, it is not from today.
Identify the original uploader. Is it a known person or organisation, a brand-new account, or an account that posts a lot of viral material?
Look for other angles. Real events in public places are usually filmed by many people. One dramatic clip with no other footage is a warning sign.
Check the details against the claim. Weather, signage, language, vehicles, clothing and landmarks can confirm or rule out a location and season.
See whether fact-checkers or established newsrooms have already examined it.

Provenance labels and watermarks

Some cameras, editing apps and AI generators now attach Content Credentials, a provenance standard developed by the industry group known as C2PA, which records how a file was made and edited. Some AI developers also embed invisible watermarks in generated media. These are useful when present, but they have limits:

Metadata is easily lost. Screenshots, re-encoding and many upload processes strip it.
A watermark only shows up for content from tools that add one. No watermark does not mean the content is genuine.
Platform labels are inconsistent, and not every service displays provenance information.

Be careful with detection tools

Online AI detectors can be helpful as one signal, but they produce both false alarms and misses, and their results can change between versions. Never treat a single detector score as a verdict, especially when a person's reputation is at stake.

Before you share

Pause if a clip provokes a strong reaction. Outrage and shock are what fakes are designed to produce.
Share with context if you do share, saying what is and is not confirmed.
Correct yourself publicly if something you passed on turns out to be false.

Have you ever been caught out by a convincing fake, or spotted one early? Tell us how in the comments.