How to spot AI-generated images and text: what to check before you share

TUREFF• •0

AI tools now produce photos, illustrations and writing that can pass a quick glance. You do not need special software to protect yourself from most fakes, but you do need a habit: pause before you believe or share, and check a few things. No single clue proves anything; look for several together, and weigh them against where the content came from.

Visual clues in images

Image generators have improved enormously, and the obvious tells of a few years ago, such as six-fingered hands, are far less common. Still worth checking:

Text inside the image: signs, labels, number plates and logos often contain garbled or almost-right lettering.
Hands, teeth, ears and jewellery: odd joints, merged fingers, earrings that do not match, glasses that melt into the face.
Backgrounds: people in a crowd with distorted faces, buildings with impossible geometry, railings or patterns that do not continue logically.
Lighting and reflections: shadows falling in different directions, reflections in windows or eyes that do not match the scene.
Texture: skin that looks airbrushed or waxy, an overall sheen, or detail that looks sharp at first but turns to mush when you zoom in.

Real photos have flaws too, and heavily edited or compressed real images can look strange. Treat these as reasons to look closer, not verdicts.

Check the context

Context often settles the question faster than pixels.

Who posted it first? Trace the image back. An account that appeared last week and posts nothing but dramatic images deserves suspicion.
Does anyone else have it? A real event, such as a disaster or a celebrity incident, is usually photographed from several angles by several people. One perfect image with no other coverage is a warning sign.
Is it designed to make you angry or amazed? Fakes spread because they trigger strong emotions. That reaction is your cue to slow down.
Check trusted news sources and fact-checking organisations before sharing.

Reverse image search

Upload the image, or paste its address, into a reverse image search such as Google Lens, Bing Visual Search or TinEye. You may find:

the original source, which can show it was labelled as AI-made all along;
an older version, showing the image is real but from a different time or place;
fact-checks that have already examined it.

Provenance labels: Content Credentials

A growing number of tools attach a record of how an image was made. The main open standard is C2PA, shown to people as Content Credentials, often with a small "cr" icon. Some AI image generators add it to their output, some cameras and phones add it to photos at the moment of capture, and some editing apps record AI edits. You can inspect an image's credentials with the free verify tool on the Content Credentials website. Google also embeds an invisible watermark called SynthID in content made with its AI tools, and its Gemini app can check whether an image carries it. The limits matter:

Absence proves nothing. Many tools add no credentials, and social platforms often strip this information when a file is uploaded or screenshotted.
A credential shows what the tool recorded; it tells you how the file was made, not whether its claims are true.

Spotting AI-written text

Text is harder than images. Signs that may suggest AI writing:

Smooth, confident, generic prose that says little of substance.
Invented details: quotes, statistics, studies or book titles that cannot be found anywhere. Check any specific claim that matters.
Repetitive structure, such as every paragraph ending with a tidy summary.
Reviews or comments from different accounts that share oddly similar phrasing.

None of these is proof. Plenty of people write blandly, and a skilled user can edit AI text until it reads naturally.

Why AI detectors are unreliable

Tools that promise to detect AI-written text give confident-looking percentages, but they are frequently wrong in both directions. OpenAI withdrew its own text classifier in 2023, citing low accuracy. Researchers have found that detectors flag writing by non-native English speakers as AI-generated more often, and that light editing or paraphrasing fools them. Never use a detector's score alone to accuse a student, colleague or writer.

A quick routine before you share

Pause if it provokes a strong reaction.
Look closely at text, hands, backgrounds and lighting.
Check who posted it first and whether others confirm it.
Run a reverse image search.
Look for Content Credentials, without treating their absence as proof.
When in doubt, do not share.

Have you been fooled by an AI image, or caught one? Share your examples and tips in the comments.