AI Image Detection: 3 Free Tools to Spot AI Photos

The first tool to try is sakuttoAI's free AI Image Detector. Drop an image in and it reads any AI-generation records stored in the file's C2PA, XMP, EXIF or PNG data, without the image ever leaving your browser.

Check AI images with free tools before you look closer

Three free tools for AI image detection

Tool	What it checks	What it can miss
sakuttoAI AI Image Detector	Records saved inside the file	Images whose records were stripped by an upload or a screenshot
SynthID Detector	A watermark hidden in the pixels	Images from AI models that don't add the watermark
Hive	Patterns in the pixels that suggest AI	Cropped images and real photos with a small AI edit

AI image detection means working out whether a photo or illustration was made by AI, and the easiest way to do it is to let free tools do the looking. The three tools in the table each check something different, so put the same image through all three and see whether they agree.

Image generators now produce photo-quality fakes

Tools like Midjourney, Stable Diffusion and Firefly can make images that look like real photographs, which leaves most people struggling to tell the two apart. In June 2024, a Northwestern University team led by Negar Kamali published a guide to spotting the difference.

Even that guide admits some AI images carry no visible flaws at all. Counting fingers and checking shadows one by one takes time and still leaves you without an answer when there's nothing to find, so let the tools make the call and treat what you see as the reason to run them.

Three free tools that each check something different

Because each tool looks at a different kind of evidence, an image that slips past one can still be caught by another. sakuttoAI's AI Image Detector reads records written into the file, and SynthID Detector looks for a watermark hidden in the pixels. Hive's free tool estimates how AI-like an image is even when there's no record or watermark at all.

Each one gives you a result seconds after you drop the image in, so running all three is still quicker than going over the picture by hand. If they agree, you have your answer. If they don't, that's when it's worth tracing where the image came from.

Read the file's records with the AI Image Detector

The five results the AI Image Detector can show

Result	What it means
Recorded as AI-generated	The file says it was created with generative AI
Recorded as partly edited with AI	The file says part of a photo or artwork was replaced using generative AI
Recorded as a camera capture	The file says a camera took it. Records can be edited, so this isn't proof
No AI record found	The file has metadata, but none of it mentions AI
No records, so we can't tell	The file carries no creation or capture records

Besides the picture itself, an image file has room for notes on how it was made and with what software. Content Credentials, defined by the industry standard C2PA, are one such record: they can list when and where a file was created, how it was edited, and whether AI was involved. The AI Image Detector reads these notes and gives you one of the five results above.

What the AI Image Detector reads and shows

To use it, drag the images you want to check onto the AI Image Detector page. Each one is read the moment you drop it, and a result such as "Recorded as AI-generated" appears next to it. It handles JPEG, PNG and WebP, and you can drop in as many files as you like.

The tool looks at four kinds of records: C2PA Content Credentials, the IPTC DigitalSourceType field in XMP, EXIF fields such as Software, and the generation settings some tools save in PNG files. Open "Records found" to see what each one says, such as the name of the AI tool. Your files are read inside the browser and are never sent to sakutto.ai or to any AI service.

Use Content Credentials Verify to check the signature

The AI Image Detector shows what a record says, but it doesn't verify the C2PA digital signature. Records can be edited after the fact, so if you need to know whether one has been tampered with, use Content Credentials Verify instead.

Content Credentials are tied to the file with cryptography, so changing any part of the record leaves a detectable trace. Drop an image into Verify and it shows the record along with its edit history. Verify also takes formats the AI Image Detector doesn't support yet, including iPhone HEIC photos.

No record doesn't mean a person took the photo

"No records, so we can't tell" is not evidence that a camera took the picture. Many social platforms strip metadata on upload, and a screenshot or a re-save wipes it too. The Northwestern guide makes the same point about screenshots.

The C2PA itself confirms that the record can be removed. When the file comes up empty, move on to SynthID Detector and look for a watermark.

Check for a SynthID watermark

A watermark, in this sense, is a signal hidden in the pixels that people can't see but software can read. Unlike a record saved alongside the image, it lives in the picture itself. The SynthID Detector site checks whether an image carries one.

SynthID marks images from Google and its partners

SynthID is the invisible watermark Google DeepMind built for AI-generated content, and NVIDIA, OpenAI and Kakao have adopted it too. It goes in when the image is created and is designed to survive cropping, filters and lossy compression. If synthid.com finds it, that's strong evidence an AI tool made the image.

You can check images from any of those four companies' models on the site. Our walkthrough covers the steps: SynthID Detector: How to Check for AI Watermarks.

A "not detected" result doesn't clear an image

A "not detected" result doesn't mean the image is real. The picture may have come from a model that doesn't use SynthID, or heavy editing may have weakened the signal.

Some other apps add hidden marks of their own, such as the one described in Windows Paint Invisible Watermark: A Server GUID. When neither a record nor a watermark turns up, Hive's free tool is the next step.

Get an AI likelihood score from Hive

Hive's free tool takes an uploaded file or a URL and returns a score for how likely the content is to be AI-generated. It doesn't look for records or watermarks. It judges the pixels, which is why it still works on images that have neither.

A detector score is a probability

The "ai_generated" number Hive shows is a model's estimate of likelihood, not proof that AI made the image. The sample results on its page list one score per image.

Treat a high score as a reason for more suspicion, and weigh it against what the records and watermark check turned up. The Reuters Institute adds that a detector's accuracy depends on the data it was trained on.

Cropped or edited images can fool detectors

In a February 2024 test of eight publicly available detectors, the Reuters Institute showed that small edits can flip a verdict. In May 2023, a fake image of an explosion at the Pentagon spread online. Detectors flagged it as AI-generated, but once the right-hand side was cropped and scaled up, a detector suggested it was real.

AI edits to real photos slipped through as well. The testers took a real photo of Ukrainian refugees, used DALL-E 2 to replace a car with a tank, and got back "not likely to be AI-generated." That's why you shouldn't rely on Hive's score without the records and watermark checks.

How are watermark checks different from AI detectors?

A watermark check looks for a signal planted on purpose when the image was made, while an AI detector looks for leftover patterns and makes an estimate. A found watermark is far more reliable, but it only covers images from models that add the mark. Detectors reach images without a watermark too, at the cost of some reliability.

So if a record or watermark turns up, trust it over the score; if neither does, use Hive's number as one input. The Reuters Institute's advice is the same: detectors aren't a one-stop answer and need careful use.

When the results disagree, trace the image's source

A reverse image search uses the picture itself as the query and finds other pages that show the same or a similar image. Turn to it when the three tools disagree, or when there's no record or watermark and Hive's score sits in the middle. Once you know where an image has appeared, you can see who posted it first and what they said about it.

Search by image with Google Lens

Google Lens turns an image into a search and returns similar images plus the websites that use it. On a computer, go to google.com, click the Search by image icon, and then do one of the following:

• Upload a file
• Drag the image into the search box
• Paste the image's URL

It works in most browsers, including Chrome, Firefox, Safari and Edge. If a news outlet ran the same picture, its caption adds detail the outlet has checked, which a viral repost rarely carries.

Check news outlets, official accounts and comments

Look for the same picture in news reports and on official accounts, and read what people say under the post. The Northwestern guide recommends checking reputable news sources and official accounts, and reading captions and comments on social media for context.

If an image has gone viral, a fact-checker may already have looked into it. Search the subject of the image together with Snopes, FactCheck.org or Reuters Fact Check to see whether anyone has published a verdict.

Treat visual signs as the cue to run the tools

Visual signs that should send an image to the tools

Area	What looks off
Anatomy	Missing, extra or merged fingers; overlapping teeth
Text	Made-up letters, invented words, misspellings
Light	Shadows pointing different ways; reflections that don't match the scene
Texture	Skin that looks waxy, shiny or plastic

Visual signs tell you when to run a picture through the tools, not whether it's AI. If something about the anatomy, text, light or texture looks off, run the image through all three. A picture that looks fine on all four counts still isn't proven real.

Hands, teeth and other anatomy errors

Hands are where AI-generated people most often go wrong: fingers go missing, multiply or fuse together, fingernails disappear, and the hand may be the wrong size for the body. The guide also flags overlapping teeth and odd artifacts between them. The more people in a scene, the more chances the generator has to slip.

If the picture claims to show someone you can find other photos of, such as a public figure, compare their ears, nose and mouth with those photos. These features differ from person to person, so a mismatch is a good reason for doubt.

Garbled text on signs and clothing

Any writing in the picture is worth reading closely, whether it's on a storefront, a book cover or a T-shirt. AI-made lettering can look like an alphabet that doesn't exist, spell words that aren't real, or string together sentences that make no sense.

Generators are getting better at text quickly, though, and the errors can be subtle. Clean, readable lettering is no longer a sign that a photo is genuine.

Shadows and reflections that don't match the light

Shadows cast by a single light source such as the sun should run roughly parallel, so shadows that point in different directions are a warning sign. So are shadows whose shape doesn't fit the object casting them.

Mirrors, water and glass deserve a second look too. One example in the guide shows a man whose reflection is wearing a scarf he doesn't have on. Check that whatever appears in a reflection actually exists in the rest of the scene.

Skin that looks waxy or plastic

AI-generated skin is often described as waxy, shiny, glossy or cartoon-like. A face that looks airbrushed all over is worth a closer look.

Texture is the most subjective of the four signs, since people judge smoothness differently. If a face bothers you, let the tools decide, not your gut.

Run AI image detection in four steps

The quickest way to handle AI image detection is to run the image through free tools before you go over it by eye. Use these steps next time a picture makes you pause:

1. Drop it into sakuttoAI's AI Image Detector to read any records in the file
2. Check it on synthid.com for a SynthID watermark
3. Get a likelihood score from Hive's free tool
4. If the results disagree, run a reverse image search with Google Lens

Every tool can miss, and a missing record or watermark is not proof that a camera took the picture. Because the three look for different things, using all of them catches more fakes while still taking only a few minutes. And when the tools leave you unsure, holding off on sharing the picture is the safer choice.