AI-Generated Images: 6 Simple Checks That Still Work

Spotting AI-generated images used to be easy. Count the fingers, look at the teeth, read the garbled text on a shop sign. Those tells are largely gone. Current models render hands correctly and write legible text, so advice written in 2023 will now lead you to the wrong answer with confidence.

Written September 2026. Detection is an arms race, so treat any single check as evidence rather than proof.

Start with provenance, not the pixels

The most reliable check is not visual at all. Before you study the image, work out where it came from, because a picture with a traceable history is far easier to trust than one you are trying to reverse-engineer.

1. Find the earliest copy. Run a reverse image search and sort by date. A genuine news photograph usually appears on a wire service or a publication first. A synthetic one often appears first on social media with no earlier trace.
2. Look for the same scene from another angle. Real events get photographed by more than one person. If an image is the only view of a dramatic moment, be suspicious.
3. Check who published it and whether they say where it came from. “Courtesy of” and a named photographer is a claim someone is accountable for.
4. Check content credentials. Many cameras and editing tools now attach C2PA provenance data recording how an image was made and edited. Where present it is strong evidence. Absence proves nothing, because the data is easily stripped.

Our guide on how to check an AI answer applies the same principle to text: verification beats intuition.

Visual checks that still catch AI-generated images

When you have no provenance to work with, these are the artefacts that survive in current models. None is conclusive alone. Two or three together is a strong signal.

• Backgrounds fall apart under zoom. Generators spend their effort on the subject. Zoom into the crowd, the far side of the street, the shelves behind the person. That is where you find melted faces, doors that lead nowhere and objects that merge.
• Text at small sizes. Headlines render well now. Small print, number plates, price labels and background signage still tend to collapse into plausible-looking nonsense.
• Physical continuity of thin objects. Follow a strap, a chain, a railing, a cable or a necklace from one end to the other. Thin things that pass behind something and come out misaligned are a common failure.
• Repeated texture. Bricks, foliage, fabric weave and crowds often contain tiles that repeat identically. Real texture almost never repeats exactly.
• Lighting that does not agree. Check every shadow points away from the same source, and that reflections in glass, water and eyes match the scene. Multiple contradictory light directions are hard for a model to avoid.
• Skin and surfaces that are too clean. Real photographs carry sensor noise, dust, minor blur and skin irregularity. An image where every surface is flawlessly smooth is often synthetic or heavily processed.

Checks that no longer work

Save yourself time by dropping these. Counting fingers is unreliable now. Looking for garbled large text is unreliable. Judging by whether it “looks too good” is unreliable, because ordinary phone photos are heavily computationally processed and real professional photography has always looked polished.

Detector tools deserve their own warning. Automated AI-image detectors produce both false positives and false negatives at rates high enough that a single result should never be treated as an answer. They are one input, not a verdict.

Video needs a different approach

Synthetic video has caught up quickly, and the useful tells have moved. Watch the edges of a moving subject against a busy background, where warping shows. Watch for objects that change size or detail between cuts. Listen for audio that is unusually clean, with no room reverb and no incidental background sound, which is a common giveaway on generated speech.

For anything consequential, apply the provenance test first. A clip claiming to show a public figure saying something inflammatory should exist in more than one place, from a source willing to put its name to it.

A workable routine

For an ordinary social media image, thirty seconds is proportionate: reverse image search, and a zoom into the background. For something that would change your mind, or that you are about to share, spend five minutes and work through provenance properly. The cost of being wrong in public is higher than the cost of checking.

The honest summary for 2026 is that you cannot reliably identify AI-generated images by looking at them alone, and anyone who says otherwise is selling something. What you can do is establish where an image came from, and that skill will keep working no matter how good the generators get. If you are assessing a product listing or a review rather than a news image, our guide to spotting fake reviews covers the equivalent checks for text.

The situations where it matters most

You do not need to audit every picture you scroll past. The checks are worth spending time on in a small number of situations where being fooled has a real cost, and it is worth knowing which those are.

• Anything you are about to share. Sharing transfers your credibility to the image. That is the moment to spend five minutes.
• Marketplace and rental listings. Generated interiors and products are now common in scam listings. If a property looks better than the price suggests, check whether the rooms are geometrically consistent with each other.
• Profile photos on dating and networking sites. A reverse image search that returns nothing at all is itself informative: real people usually leave some trace.
• Breaking news. The first hours of any major event now reliably produce synthetic images. Wait for a wire service rather than trusting the first dramatic photograph.
• Anything asking for money. Charity appeals and crowdfunding campaigns built on generated imagery are an established pattern.

A small toolkit

Three free tools cover most of what you need. A reverse image search engine, of which Google Images and TinEye behave differently enough that trying both is worthwhile. An EXIF viewer, which shows the camera data attached to a file, though social platforms strip this on upload so its absence means little. And a content credentials inspector for checking C2PA data where it exists.

Note what is missing from that list: an AI detector. They are not reliable enough to be a fourth tool, and treating one as authoritative is how confident people end up wrong in public.

Common questions

Can AI image detectors be trusted? Not on their own. They produce both false positives and false negatives often enough that a single result is evidence at best. Use one as one input alongside provenance checks, never as the decision.

Does counting fingers still work? No. Current models render hands correctly most of the time. Advice built around anatomical errors is now several generations out of date and will give you false confidence.

What are content credentials? C2PA content credentials are provenance data attached by many cameras and editing tools, recording how an image was created and modified. Where present they are strong evidence. They can be stripped, so their absence tells you nothing.

What is the single best check? Reverse image search to find the earliest copy. Provenance beats pixel inspection, and it will keep working as generators improve.

Sources and further reading

Where the figures and rules above come from, so you can check them:

• Content provenance standard used by cameras and editing tools: C2PA
• Guidance on verifying images and video: Reuters Institute