Before you start
You need one picture you doubt, a browser and about ten minutes. Nothing to install. If you want a practice round, save two pictures, one you took yourself and one an AI image tool made for you, and run every check on both.
Five checks for a picture you doubt
- Find out who posted it and when. A picture with no author, date or place attached is a claim with nothing behind it. Open the account that shared it and read its other posts.
- Run a reverse image search with Google Images or Lens, TinEye, Bing or Yandex Images. Wikipedia says Google's Search by Image takes an uploaded picture or a copied image address. Open every page that shows the picture and note the earliest one.
- Zoom in and look at hands, writing on signs, shadows, skin and hair. Use the table below as your list.
- Look for a signed credential or a watermark. Some AI tools attach one, and lesson 3 shows how to read it. A missing one tells you nothing.
- Add up what you found. An earlier copy under a different story, an account with no history or a credential that names an AI tool are strong signs. One odd finger is a weak one. If the checks do not settle it, call the picture unverified.
How often people get it wrong
A 2023 paper tested the eye directly. Its authors gave 50 volunteers 100 pictures each, half real and half made with Midjourney V5, and asked which were AI. The volunteers averaged 61.3% correct, so 38.7% of their answers were wrong. A separate set of 1,085 crowdsourced questionnaires averaged 49.9%, which is a coin flip. The authors also had annotators remove low-quality AI images that were obvious fakes at first glance, so the test used hand-picked pictures.
The kind of picture mattered. Groups of people were the easiest to sort. Everyday objects were the hardest, and volunteers scored 50.8% on them.
The authors also trained a detection model that got 13% of the same kind of test wrong, which looks like a win for software. The Deepfake-Eval-2024 benchmark took the question out of the lab. Its authors collected fakes from social media and from users of a deepfake detection platform in 2024, and open-source detectors lost 45% of their AUC on images compared with earlier benchmarks. AUC is a score for how well a detector separates fake from real, so a drop that size is large. None of the commercial detectors it evaluated reached 90% accuracy.
For you, that makes an online detector's percentage one clue among five and never the verdict.
Source: Hugging Face Papers: human and model perception of AI-generated images (HPBench), table 5 (read November 6, 2026)
The flaws people looked for, and why they fade
The 2023 volunteers named the flaws they used, and they could tick more than one reason per answer. Of all the reasons ticked after a correct answer, 28% were problems with fine detail and 17% were skin that looked too smooth. The paper lists nine kinds of defect, and the table shows six of them.
Read it as a record of 2023 pictures. Every AI image in that test came from one generator, Midjourney V5, and newer generators have appeared since. We did not find a newer human test of today's tools on the pages we read, so treat these flaws as clues, and treat a clean hand as proving nothing.
| Look at | What the paper counted as a flaw | What to do |
|---|---|---|
| Hands | Fingers that overlap or take unreasonable shapes | Zoom in and count the fingers and joints |
| Writing | Incorrect signs and text | Read every sign, label and badge letter by letter |
| Light and shadow | Unnatural light sources and shadows | Check that each shadow points away from the same light |
| Skin and expression | Overly smooth skin and unrealistic expressions | Look for pores and creases, and for a face that fits the moment |
| Daubed areas | Rough or uneven colors and textures, with details lost | Look for patches that seem painted on, and compare one surface in two parts of the frame |
| Fine detail | Problems in wrinkles and hair details | Follow single hairs and wrinkles to see where they end |
What a reverse image search can tell you
Reverse image search is a content-based image retrieval technique. You give the system a picture, and it looks for matching pictures instead of matching keywords. Wikipedia names Google Images and Lens, TinEye, Bing and Yandex Images among the services.
Use it for origin. Wikipedia says it can locate an image's source, find higher-resolution versions, show the pages where an image appears and identify the creator. It can also reveal manipulated versions and derivative works. A news photo that turns up years earlier under another caption has been reused, and you need no pixel analysis to say so.
Do not read it as an AI test. An empty result means the engine found no match, and a brand-new AI picture and a brand-new photo both give one. TinEye matches heavily edited copies of a picture, but the same article says it "will not usually return similar images in the results".
A video on the same topic
The video below comes from New Scientist's YouTube channel and is titled "How to spot deepfakes and AI-generated images". We checked only its title and channel, not what it shows, so use it as extra viewing and rely on the five checks above.
Common mistakes
- Treating a perfect hand as proof that the picture is real. The 2023 flaws are clues from one generator, and the checks that settle a picture are its source and its earlier copies.
- Trusting an "AI or not" percentage. Detectors that looked good on older benchmarks did far worse on fakes collected online in 2024, so where a picture came from stays your best clue.
- Reading an empty reverse image search as "so it is AI". It means no match was found, and a fresh photo gives the same result.
- Checking only the pixels. The post a picture arrived in, the account behind it and the date carry as much information as the image does.
- Forgetting that a real photo can be AI-edited. Removing an object or a person leaves the rest real, and the editing lesson shows what such tools do.
Try it yourself
- Pick three pictures from your feed that made you stop. Run checks 1 and 2 on each and write down the earliest date you can find.
- Make one picture with an AI image tool of your choice. View it at full size, go through the six rows of the table and count the flaws you can find. Then ask a friend to pick the AI picture from it and a real photo.
- Take a photo you have never posted and run it through a reverse image search. Note what comes back, so you know what an empty result looks like.
Next: deepfake video and cloned voices
A picture is the easiest fake to check, because you can zoom in and search for it. Moving pictures and voices give you less to hold on to. Lesson 2 shows the checks that still work there, including one that does not depend on your eyes or ears at all.
Check yourself
In the 2023 study, how often were volunteers wrong when sorting real photos from Midjourney V5 images?
Show answer
B. About 39 answers in 100 The volunteers averaged 61.3% correct, so 38.7% of their answers were wrong.
What does a reverse image search mainly tell you?
Show answer
A. Where else the picture has appeared and where it came from Wikipedia says it can locate a picture's source, find other versions and show the pages where it appears.
A reverse image search for a viral photo returns nothing. What does that show?
Show answer
C. Only that the search engine found no match An empty result means no match, and a brand-new real photo gives the same result as a brand-new AI picture.
An online detector says a picture is 97% likely to be AI. What is the best use of that number?
Show answer
D. Treat it as one clue next to the picture's source and earlier copies The Deepfake-Eval-2024 benchmark found that no commercial detector it evaluated reached 90% accuracy, so a score alone settles nothing.
Frequently asked questions
Can I spot an AI image just by looking at it for a few seconds?
Rarely. The 2023 volunteers had no time limit and still scored close to a coin flip on some kinds of picture. Use a few seconds to find the post the picture came from and the oldest copy of it, because a search can settle that and a glance cannot.
Does reverse image search show whether a picture is AI-made?
No. By Wikipedia's account it finds matching or edited copies and the pages that show a picture, so it tells you where a picture has been and not how it was made. Finding the same picture years earlier is what exposes a reused photo.
Are AI image detectors accurate?
Not enough to rely on alone. On the fakes collected for Deepfake-Eval-2024, the best image model its authors fine-tuned reached 63% accuracy, and none of the commercial detectors they evaluated reached 90%.
Do AI images still have extra fingers?
Sometimes, but do not count on it. Hands were one of the flaws volunteers named in the 2023 study of Midjourney V5 images, and we found no newer measurement. A normal hand is no proof of a real photo.
What if the photo is real but was edited with AI?
Then parts of it are real and parts are not, and the five checks can miss that. A reverse image search that finds the original lets you lay the two side by side and see what changed.
Sources
- Hugging Face Papers: human and model perception of AI-generated images (Fake2M and HPBench), paper 2304.13023 https://huggingface.co/papers/2304.13023
- Hugging Face Papers: Deepfake-Eval-2024, a multi-modal in-the-wild benchmark of deepfakes circulated in 2024 https://huggingface.co/papers/2503.02857
- Wikipedia: Reverse image search https://en.wikipedia.org/wiki/Reverse_image_search
- YouTube: New Scientist, How to spot deepfakes and AI-generated images https://www.youtube.com/watch?v=7akzhpx0EIU