AI
How to Spot an AI-Generated Image
People spot AI images correctly only 62% of the time, and the old tells are dead. Check the source, reverse-search, and read the provenance label.
Last updated: 2026-11-15 · 7 min read
The short answer: looking is no longer enough
A few years ago, spotting an AI-generated image was a parlour game. Count the fingers, look at the hands, find the warped text on the sign in the background, and you had your answer. That game is over. The image generators of 2026 have learned hands, text, reflections and lighting, and the pictures they produce pass a casual look without effort.
The numbers say so plainly. A study by Microsoft's AI for Good Lab asked more than 12,500 people to sort real photographs from AI-generated images, collecting around 287,000 judgements — and people were right only 62 percent of the time, barely above chance, on images made by 2024-era generators. The generators have improved since then. Human eyes have not.
That does not mean detection is hopeless. It means the method has to change. The honest answer for 2026 is that looking at the picture is now the weakest test, not the strongest. The checks that work are the ones that ignore the pixels and investigate the image instead: where it came from, whether it exists anywhere else, and whether it carries a provenance label. This article explains all three, and what to do in the common case where none of them settles the question.
Why the old tells are dead
Most guides to spotting AI images still list the same tells they listed in 2023: six fingers, melted hands, garbled text, mismatched earrings, impossible reflections. Some of these still appear, especially in images from free or outdated generators, and a picture with three arms is still a giveaway. But the mainstream generators fixed the famous flaws first, because they were the most embarrassing. Treating a clean pair of hands as proof that an image is real is now exactly backwards: the best generators produce clean hands every time.
The deeper problem is that real photographs have flaws too. A genuine photo can have a warped sign, a strange reflection in a window, a hand hidden behind a bag, or a face blurred by motion. Judging by flaws alone means flagging real photos as fake and passing fake images as real, often on the same day. The tells were never reliable evidence. They were shortcuts that worked while the generators were clumsy, and the generators stopped being clumsy.
There is one partial exception, and it is worth knowing. A study from the Australian National University in June 2026 found that teaching people broad qualities of AI-generated faces — the overly smooth skin, the too-perfect symmetry, the oddly empty eyes — nearly doubled their accuracy, where teaching them to hunt for specific flaws did not help much. The lesson is to look at the whole face, not the fingers. But even that is a probability, not a proof, and it works best on faces, not on landscapes, objects or scenes.
What actually works: check the source
The strongest check costs thirty seconds and needs no expertise. Ask where the image came from. A photograph published by a news agency, a government department or a photographer with a verifiable history is far more trustworthy than the same image posted by an anonymous account that appeared last week. Scroll back: does the account have a history, or was it created to post this one picture? Does the story the image tells match what reliable sources are reporting?
The second check is a reverse image search. Upload the image, or search by its address, and see where else it appears. A real photograph of a real event usually has a trail: the same image on a news site, the photographer's portfolio, or an earlier date. An AI image often has no trail at all, or a trail that leads to an AI gallery or a social media post with no source. A reverse search can also catch the oldest trick of all: a genuine photograph from years ago, presented as if it happened yesterday.
The third check is the context around the image. Does the caption make a claim that no reliable source supports? Is the image the only evidence for an extraordinary claim? Is it being shared by accounts that benefit from you believing it? None of these proves an image is fake, but together they tell you how much to trust it — which is the question that actually matters.
Provenance: the label that travels with the image
The most important development in 2026 is not a better detector but a label. The C2PA standard, known to users as Content Credentials, embeds a signed record inside the image file itself: which camera or software made it, when, and whether AI was involved. Cameras from major manufacturers and tools from Adobe and others now write these credentials, and you can read them in a browser or an app by clicking the small Content Credentials icon.
Google's SynthID works the same way from the other side: it embeds an invisible watermark into images made by Google's AI tools, identifying them as AI-generated even if the file is renamed, cropped or re-shared. The two systems are different, but the idea is the same: instead of guessing from the pixels, read the label the image carries with it.
The catch is coverage. A label only helps if the image has one, and most AI images shared on social media in 2026 still do not, because the platforms strip metadata when they recompress uploads. The law is pushing in the right direction: the European Union's AI Act bans non-consensual intimate images made by AI from 2 December 2026, and the United States' TAKE IT DOWN Act requires platforms to remove such images within 48 hours of a request. But provenance is a slow fix, and for now the absence of a label proves nothing. A missing label means you fall back on the source and the reverse search.
What to do when you still cannot tell
Often, after all three checks, you still will not know. That is the normal outcome in 2026, and the right response is to act on that uncertainty instead of resolving it by gut feeling. Do not share the image as fact. Do not use it to settle an argument. If the claim matters — a news event, a product, a person accused of something — wait for a reliable source to confirm it, and check whether one already has.
If the image is being used to harm someone, report it. Platforms have reporting paths for non-consensual intimate images and for deceptive media, and the TAKE IT DOWN Act gives victims in the United States a legal route to removal within 48 hours. If you created an AI image yourself, say so: a clear label protects your audience and costs you nothing. The age of spotting fakes by eye is ending. The age of checking the source, and of labelling honestly, is the one we are in now.
Related on Gargloo
Frequently asked questions
Can you still tell an AI image by the hands?
Not reliably. Early generators often drew extra or melted fingers, but the mainstream generators of 2026 have largely fixed hands, text and reflections. A picture with obvious hand errors is still likely AI-made, but clean hands prove nothing — the best generators produce them every time. Judge the source and the provenance, not the fingers.
How accurate are people at spotting AI images?
Barely above chance. A Microsoft AI for Good Lab study with more than 12,500 people and around 287,000 judgements found people sorted real and AI images correctly only 62 percent of the time, on 2024-era generators. Teaching broad qualities of AI faces nearly doubled accuracy in a 2026 Australian National University study, but even that is a probability, not a proof.
What is a reverse image search, and how does it help?
It is a search by the image itself instead of by words: you upload the picture or paste its address, and the search engine shows where else it appears. A real photograph usually has a trail — a news site, a portfolio, an earlier date. An AI image often has no trail, or one that leads to an AI gallery. It also catches old real photos being presented as new.
What are Content Credentials and SynthID?
They are provenance labels that travel with the image. C2PA Content Credentials embed a signed record of what made the image and whether AI was involved, readable in a browser or app. Google's SynthID embeds an invisible watermark in images made by Google's AI tools. Both only help when the image still carries the label — social media platforms often strip it when they recompress uploads.

