How to Tell if a Picture Is AI Generated: 3 Checks That Beat Your Eyes

ByProf. H
How to tell if a picture is AI generated: four men leaning over one print, unable to tell by looking
Honoré Daumier, The Print Collectors (c. 1860–63) — public domain. Four experts, one picture, and the only method they have is to lean in closer. That method is now the least reliable one available to you.

The three checks that actually work take about thirty seconds, cost nothing, and not one of them involves looking harder at the picture: open the file’s Content Credentials, ask Google where the image first appeared, and ask Gemini whether it carries an invisible SynthID watermark. Staring at it is the method that fails — in a Microsoft AI for Good Lab experiment covering roughly 287,000 judgments by more than 12,500 people, players were right only 62% of the time, and on two types of AI image they scored below 50%, which is worse than guessing.

Here is how it usually happens. A photograph arrives in a group message or turns up in your feed — a flooded street, a celebrity somewhere they should not be, a house for rent at a price that makes no sense. Someone asks whether it is real.

So you do what everyone says to do. You look at the hands. You count the fingers. You check the teeth and the ears and the background lettering.

That advice was written for 2023, and it is quietly expiring.

The thirty-second version. Do these in order and stop as soon as one of them answers your question.

  1. Drop the file on contentcredentials.org/verify. If it carries a credential, you will see who made it and with what.
  2. Search the picture itself with Google Lens, or open About this image from the three dots on a Google Images result, and see where it first showed up.
  3. Upload it in the Gemini app and ask whether it was made with Google AI. Gemini looks for a SynthID watermark you cannot see.
Three checks for whether a picture is AI generated: Content Credentials, where it first appeared, and a SynthID watermark, with what each one proves and misses
What each check proves, and where each one goes blind. None of the three is complete on its own, which is why the order matters.

Why squinting at the picture stopped working

In 2025, researchers at the Microsoft AI for Good Lab published results from an online game called Real or Not. Players were shown real photographs and AI-generated images and asked to sort them.

About 287,000 judgments came in from more than 12,500 people around the world. Roughly 110,000 of those judgments were wrong.

That is a success rate of 62% — a coin flip is 50%. On the AI images specifically, players correctly identified 121,735 out of 193,779, or 63%.

A nineteenth-century photograph of an older bearded man examining a manuscript through a magnifying glass
Antoine-Samuel Adam-Salomon, Bearded Man with Magnifying Glass Examining a Manuscript — Metropolitan Museum of Art, CC0. Close visual inspection is the oldest verification method there is. Against a modern image generator it now scores 62%.

The detail worth sitting with is which pictures fooled people most. The authors report that participants did best on human portraits — faces are what we are built to read — and struggled most with natural and urban landscapes.

And two categories beat the humans outright, scoring below 50%: faces made by older GAN tools, and inpaintings, where only part of a real photograph has been altered.

That last one matters more than it sounds. A picture does not have to be invented from nothing to mislead you. Someone can take a genuine photograph and change one object in it, and the rest of the frame will pass every visual test you know.

Chart showing people identified AI images correctly 62% of the time versus a detection tool scoring over 95%
Every figure here is stated directly in the paper. The study team’s own detector scored over 95% on both real and AI images — the gap between the red bars and the green one is the whole argument for not trusting your eyes.

The authors are blunt about where this is heading: because generators keep improving, they write that their results likely overestimate how well people can do today.

The honest summary. This is not a story about detectors being useless. It is the opposite. Machines are much better at this than people are — the study’s own detector scored above 95% — and the researchers still note that detectors make mistakes too. What has collapsed is the idea that a careful person can settle it by looking.

Hany Farid, who has spent his career on digital image forensics, on what actually holds up. TED — 27.8M subscribers, 1,346,048 views as of August 31, 2026.

Check 1: look for Content Credentials

Content Credentials are the closest thing we have to a nutrition label for a picture. The open standard behind them is called C2PA, maintained by the Coalition for Content Provenance and Authenticity, and the Content Authenticity Initiative that promotes it was established by Adobe.

When a credential is attached, it travels with the file and records what made the image and what was edited afterward. OpenAI attaches them to images its tools generate, Google’s Pixel Camera writes them into photos as they are taken, and in June 2025 Sony launched a C2PA-based Camera Verify system for press photographers and news editors.

To check one: go to contentcredentials.org/verify and drag the image file onto the page. There is no account and no limit.

Two press photographers filing pictures from the street outside a government building, one shielding a screen with a jacket
Photo: Katie Chan, Wikimedia Commons — CC BY-SA 4.0. A news picture’s chain of custody begins here, in the minutes after it is taken. Content Credentials are an attempt to make that chain travel with the file.

Now the part the tool vendors tend to skip. Most pictures you receive will have no credential at all, and a missing credential proves nothing.

Credentials are fragile. Screenshot a picture and the credential is gone. Save it, re-upload it, or pass it through an app that strips metadata, and it is gone. A photograph with no credential is simply a photograph we know nothing about — which is most of them.

Check 2: ask where the picture first appeared

This is the check that does the most work, and almost nobody does it. Instead of asking “is this image fake,” you ask “where has this image been before today?”

A photo of flooding that is presented as this week’s storm but was first indexed in 2019 is answered without any judgment about pixels at all.

Google’s About this image shows you when the image and similar ones were first indexed, where it may have first appeared, and where else it has been seen — including on news and fact-checking sites. Google’s own announcement gives the example of an image of a staged moon landing, where the surrounding coverage was what identified it as AI-generated.

On a computer, the official steps are:

  • Go to google.com, click Search by image, then Upload a file and choose your picture.
  • Or drag the image file straight into the Google search box.
  • Or, in Chrome, right-click any image on a page and choose Search with Google Lens.
  • On a Google Images result, click the three dots on the image to open About this image.
Google’s own walkthrough, published July 23, 2026. Google — 14.5M subscribers, 47,301 views as of August 31, 2026.

This is the same habit that protects you elsewhere. It is what we recommend for checking whether a Facebook ad is fake, and it is the visual cousin of working out whether a phone call is AI.

Check 3: ask Gemini about the watermark

Google embeds an invisible watermark called SynthID into images made with its own AI tools. It does not change how the picture looks, and you cannot see it — but Google’s systems can read it.

A watermark visible in the paper of a fifty euro banknote when held up to the light
Photo: public domain, Wikimedia Commons. A banknote watermark is the old version of the same idea: a mark built into the material itself rather than printed on top, invisible until something checks for it.

To use it: open the Gemini app, upload the picture, and ask whether it was created with Google AI. Gemini checks for a SynthID watermark and tells you if it finds one.

The limit here is strict and easy to misread. SynthID only marks Google’s own output. A picture from Midjourney, from a Chinese generator, or from any of the dozens of open tools carries no SynthID, and Gemini finding nothing does not mean the picture is real.

Are AI image detectors accurate?

Better than you, and not as good as their home pages suggest.

The Microsoft team’s own detector scored above 95% on both real and AI images, and the paper says detector success rates are consistent across image categories in a way human performance is not. The same paper adds the necessary caveat: detectors “too, will make mistakes.”

So a detector is a useful second opinion. It is a bad only opinion — and a very bad basis for accusing a person of faking something.

There is also a practical problem with the ones that rank highest when you search for a free checker. On August 31, 2026 we opened the home page of each and read what it offered.

Table comparing free AI image detector limits: five free checks a month, ten checks then five dollars a month, versus tools with no cap
Our own count, taken August 31, 2026. Every one of the four detectors advertises itself as free, and every one of the four has a cap or a paid tier on the same screen. The two green rows do not.

Is It AI offers five free detections a month and asks you to create an account. AI Photo Check gives ten free checks and then asks $5 a month. Isgen and ZeroGPT both run a free tier with pricing on the same screen.

None of that is dishonest. But if you check pictures more than a few times a month, the free path is the one built into tools you already have. We keep a running record of how we count things like this on our how we count page.

Why do AI images look glossy, yellow, or just off?

Because a generator has an average taste, and it returns to it.

Ask for a photograph and many models give you soft even lighting, warm skin, a slight golden cast, shallow depth of field and a suspicious absence of clutter. Real rooms have crumbs and cables. Real light comes from one bad direction.

This is a genuine clue, and the Microsoft researchers describe exactly this: people learn to recognize the average aesthetic of a generator rather than any real limit on what it can produce.

Which is precisely why the clue is fading. It reflects the default settings the model makers chose, not the ceiling of the technology — and defaults change with every release. The pictures that fooled people most in the study were the ones that looked realistic but not professional.

The reason this matters more after 60

Fake pictures are rarely the point. They are the setup.

A stolen or generated photograph is the opening move in a romance scam, and the FTC’s guidance on romance scams recommends the same reverse-image habit described above — searching the profile picture to see whether it belongs to someone else entirely.

The same logic runs through the fake charity appeal after a disaster and the too-good rental listing. We wrote about that pattern in detail in AI scams targeting seniors.

The one rule that survives all of this. If a picture is being used to get money, credentials, or urgency out of you, the image is not the thing to verify. Verify the person, on a phone number you already had. No amount of image analysis substitutes for that.

What to do with this

  • Bookmark contentcredentials.org/verify right now, before you need it.
  • Practice the Lens step once today on a picture you already know is real, so the steps are familiar when it matters.
  • Change the question you ask. Not “does this look fake” but “where was this picture before today.”
  • Do not accuse anyone based on a detector score. Treat it as one input.
  • If money is involved, stop and call. Report scams to the FTC at reportfraud.ftc.gov.

If you want to go deeper

Frequently asked questions

How can I tell if a picture is AI generated for free?

Use the two tools with no cap and no account: drop the file on contentcredentials.org/verify to look for a Content Credential, and search the image itself with Google Lens to see where it first appeared. Dedicated detectors also have free tiers, but the ones we checked on August 31, 2026 cap them — five checks a month at Is It AI, ten checks then $5 a month at AI Photo Check.

Are AI image detectors accurate?

More accurate than people. In the Microsoft AI for Good Lab study, human players scored 62% while the research team’s own detector scored above 95% on both real and AI images. The same paper warns that detectors make mistakes too, so treat a detector result as a second opinion rather than a verdict — especially before accusing anyone of anything.

Do AI images have hidden watermarks?

Some do. Google embeds an invisible SynthID watermark in images made with its own AI tools, and you can ask the Gemini app to check for it. It is not universal: images from other generators carry no SynthID, so finding no watermark does not mean a picture is real.

Why do AI pictures look glossy or yellow?

Because generators have default habits — even lighting, warm tones, shallow focus and unusually tidy scenes. Researchers note that people learn a generator’s average look rather than its actual limits, which is why this clue weakens with each new model release.

What is a Content Credential?

It is a record attached to an image file showing what made it and what was edited, built on an open standard called C2PA. Think of it as a nutrition label for a picture. It only helps when it is present, and it is stripped by screenshots and by many re-uploads.

Can I tell if a photo is AI just by looking at the hands?

Not reliably any more. Hands and teeth were useful tells in early image generators and are largely fixed in current ones. In the Microsoft study, the images people misjudged most were realistic but unpolished ones, and partly-edited real photographs scored below 50%.

Does a reverse image search work on a phone?

Yes. Use the Google app or Google Lens, choose the photo from your camera roll or take a picture of the screen, and Lens will show you where else that image appears online.

Someone sent me a photo asking for money. What should I do?

Treat the image as irrelevant and verify the person instead, using a phone number you already have — not one in the message. The FTC’s romance scam guidance recommends a reverse image search of any profile photo, and scams can be reported at reportfraud.ftc.gov.

Sources

Keep reading

Written by Prof. H. Every figure in this article was read from its primary source on August 31, 2026, and the detector limits were checked by hand the same day. Our method is described on how we count; corrections go to contact.

Similar Posts