The AI or Not image detector is one of the most searched tools for checking whether a photo was made by AI, and its promise is simple: upload a photo, get a verdict. If you have a picture you are not sure about, a profile photo, a listing image, a screenshot someone sent you, this guide walks you through the check exactly as it works today, including the step the tool’s own homepage does not warn you about.
Short answer: Go to aiornot.com and upload your photo, and the detector returns a verdict with a percentage and a generator breakdown. As of November 2026 the web checker asks you to create an account first, and the free tier includes 20 image checks. Treat any score as one signal, never proof, and back it up with a second tool.
What AI or Not is: AI or Not is a detection service that checks images, text, video, audio and deepfakes for AI generation. The image checker is its best known tool: you upload a file and it estimates whether the picture was made by AI, names the generator family it suspects, and gives a separate deepfake score.
How do you check a photo with the AI or Not image detector?
Follow these steps in order. They describe the site as it behaved when we tested it on 15 November 2026.
- Open aiornot.com in your browser. The homepage shows a drag-and-drop upload area under the heading “AI Detection API For All Content Types”.
- Click “Upload File” or drag your photo into the area. Supported formats are PNG, JPG, WEBP and GIF, plus MP3 and MP4 for audio and video checks.
- Create a free account when the site asks. This is the step that catches people out. Clicking upload opens a “Create Account” panel with “Continue with Google” and “Continue with Apple” buttons, or a name, email and password form. Until you create an account, the checker stays on “File required” and “Waiting for input”, and no verdict appears.
- Upload the photo again once signed in. The check runs in a few seconds.
- Read the result screen. It shows an overall verdict (AI or Not AI), an AI percentage, a Deepfake percentage, and a Data Breakdown that splits the AI score across generator names such as Stable Diffusion and Midjourney.
- Save the result. Screenshot the result screen with the file name and date visible, so you have a record if you need to show someone else what the tool said.
Use the original file wherever you can. A screenshot of a photo, a re-saved JPEG, or a version downloaded from a messaging app carries fresh compression that weakens the patterns the detector reads. If the only copy you have is a screenshot, say so in your own notes, and lean harder on the provenance and reverse search checks later in this guide.
What does the AI or Not result screen show?
The result has four parts, and each one means something different. Read all four before you decide what the result tells you.
The verdict. A plain “AI” or “Not AI” answer. This is the part most people quote, and it is also the part that hides the most. A binary verdict gives you no sense of how close the call was. A photo scored 52% AI and one scored 99% AI both come back as “AI”, and those two results deserve very different levels of confidence.
The AI percentage. The headline number, such as 87% AI. This is a probability estimate from the model, not a measurement of how much of the image is AI made. It means the picture’s pixel patterns look like the AI images the tool was trained on to that degree. It does not mean 87% of the pixels are synthetic.
The Deepfake percentage. A separate score for face-swap and face-manipulation signals. A low deepfake score next to a high AI score points at a fully generated image rather than a swapped face. A high deepfake score on a portrait is the number to look at first when the question is whether a real person’s face was pasted onto another photo.
The Data Breakdown. This is the most useful part of the screen, and the part most guides skip. It splits the AI score across generator names, for example Stable Diffusion, Midjourney, and a deepfake line. A result that puts most of its weight on one generator, say 80% on a single name, is a cleaner signal than a result spread thinly across three or four names. A spread result usually means the model is unsure and is hedging across the generators it knows. Treat a named generator as a lead to check, not as proof of which tool made the image. Generator guesses are less reliable than the basic AI or not verdict, especially for model versions released after the detector was trained.
One more thing to note on the result screen: a 0% reading on every line, with “Waiting for input” still showing, means no check has run at all. That is what we saw before creating an account, and it is easy to mistake for a real result if you are in a hurry. If every number reads 0%, the tool has not checked your photo yet.
Our worked check: the same two photos through AI or Not and Hive
The honest way to judge a detector is to run it on photos you already know the answer for. We did that on 15 November 2026 with two files whose origins are certain.
The first is a known AI image: a pair of iridescent holographic headphones made with Adobe Firefly Image 5, taken from Adobe’s own Firefly gallery. It is a 1024 by 1024 PNG, and we had already read its file in an earlier test: it carries a Content Credentials manifest naming Adobe Firefly as the creator, dated 9 April 2026. The second is a known real photograph: a sprout growing in soil, shot on a Canon EOS 5D Mark III and published on Burst, with its camera EXIF data intact. One image is certainly AI made, the other is certainly a camera photo.
Here is what happened when we ran both through AI or Not and through Hive, our free alternative.
| Photo | AI or Not | Hive (Deploy Detect demo) |
|---|---|---|
| Firefly headphones (known AI image) | No result. Clicking upload opened a Create Account panel, and without an account the checker stayed on “Waiting for input” with every score at 0%. | Correct. “This input is likely to contain AI-generated or deepfake content.” 99.9% likely to be AI-generated, 0% likely to be deepfake. Generator breakdown named adobefirefly at 81.4%. Its Content Credentials read “Adobe_Firefly”, action “Created”, media source “AI generated”. |
| Burst sprout (known real photo) | No result. Same account panel, same 0% readings, no verdict. | Correct. “This input is not likely to contain AI-generated or deepfake content.” 17.3% likely to be AI-generated, 0% likely to be deepfake. No Content Credentials found for this media. |
The AI or Not column is the finding, not a gap in the test. As of 15 November 2026, the free web checker on the AI or Not homepage does not return a verdict until you create an account. We did not create one for this test, so there is no AI or Not score to report for either photo, and we will not invent one. The 20 free image checks on the free tier are real, but they sit behind the sign-in step.
Hive, by contrast, needed no account and no payment. Its free demo, reached from Hive’s own detection page, scored both photos correctly, and its generator naming did the job the AI or Not Data Breakdown is meant to do: it put 81.4% on adobefirefly for the Firefly image, which matches the file’s own Content Credentials. The real photo’s 17.3% is worth reading carefully. Hive did not call it AI, and 17.3% is a low score, but it is not zero. Real photos routinely score somewhere above zero, which is why a single number should never be read as a yes or no on its own. The same photo scored exactly 17.3% when we ran it through Hive in an earlier test, so the result is stable, not a one-off.
Our full Hive AI detector guide walks through that tool step by step, and our Pangram image detector guide covers the other checker in this family, so you can compare all three before you settle on one.
How accurate is AI or Not?
Start with the number AI or Not puts on its own homepage: 98.9% accuracy. That is a vendor claim, measured on the company’s own test sets, and it should be read as one. A figure measured on clean files from generators the tool was trained on says little about a screenshot of a photo you were sent on WhatsApp.
The independent evidence is more useful, and it is mixed in a way that actually helps you use the tool well. In May 2026, NewsGuard tested five leading detectors on 45 images: 15 authentic news photographs from Reuters, the Associated Press, The New York Times, The Guardian and Google Earth, plus a lightly retouched and a substantially altered version of each. AI or Not was the only tool in the test to catch 100% of the substantially altered images. On the other side of the same test, it wrongly flagged 1 of the 15 untouched real photos as AI generated, a 6.67% false positive rate, and it called 87% of the lightly retouched real photos AI. Hive and Sightengine flagged none of the 15 real photos. Across all five tools, real photographs were declared AI generated 13.33% of the time, and the five tools disagreed with each other on 35 of the 45 images.
That pattern is the whole story of this tool in one paragraph. AI or Not is aggressive: it catches nearly everything that has been meaningfully changed by AI, and that same aggression is what makes it flag real photos, especially real photos that have been lightly edited. A photo that has had its lighting enhanced or its background blurred is the exact case where AI or Not is most likely to cry wolf, at 87% in NewsGuard’s test. If the photo you are checking has been through any editing app, read an AI verdict with that in mind.
Compression does the opposite damage. Bellingcat ran 100 Midjourney images through the AI or Not detector in 2023 and it caught all 100. Then they compressed ten of them down to the size a social site would serve, and the detector called seven of the ten real. The same tool, the same pictures, a different result once the file had been squeezed. That is why the original file matters, and why a detector result on a screenshot is weaker evidence than a result on the untouched original.
So the fair summary is this. On clean files from generators it knows, AI or Not is one of the strongest detectors available, and it is the best of the five NewsGuard tested at spotting photos that have been substantially altered by AI. On real photos, especially edited or compressed ones, it overcalls. Use it as a sensitive first screen, not as a judge.
What does the free tier include, and when should you pay?
AI or Not runs on credits. Every check costs credits, and the plans differ in how many you get and what else they unlock. These are the prices from the company’s own pricing page, read on 14 November 2026.
| Plan | Price | What you get |
|---|---|---|
| Free | Free | $5 in free credits to start, 20 AI image checks, 1 million words of AI text detection a month, API key included. |
| Pro | $5 a month | $10 in credits every month, 500 image checks, 2 million words, 1,000 seconds of video checks. Unlocks text, image, deepfake, voice, music and video detection. |
| Business | $100 a month | $100 in credits every month, 6,600 image checks, 20 million words, 13,300 seconds of video. Image, audio and video checks cost 25% less than Pro. |
| Enterprise | Custom | Custom models, private deployment, full reseller rights, dedicated support. |
Per check, an image costs $0.02 on Pro and $0.015 on Business. Unused credits on a monthly plan roll over to the next month for as long as the plan stays active.
Before you pay anything, work through the honest question: how many photos do you actually need to check? If the answer is one photo, or a handful a month, the free tier’s 20 image checks cover you, and paying adds nothing to the accuracy of a single verdict. The paid tiers buy volume, API access and extra media types, not a better score on the photo in front of you. Pay only if you are checking images in bulk, building the detector into a workflow, or need video, audio and deepfake checks as well. For a one-off check of a profile photo or a listing image, create the free account, use the free checks, and stop there.
One warning that belongs in this section: never pay to check a photo that someone sent you as part of a scam. A detector result does not make a prize real, a seller honest, or a job offer genuine. If the photo arrived with a demand for money, the payment is the test, not the pixels. Walk away from the demand first and check the photo second.
What can AI or Not not tell you?
Every detector has blind spots, and this one is no different. Knowing them before you run a check stops you from over-reading the result.
It cannot prove a photo is real. A “Not AI” verdict means the picture’s patterns look like the real photos the tool was trained on. It does not mean the photo is untouched, unedited, or honestly captioned. A real photo can still be old, taken somewhere else, or cropped to mislead, and no detector will catch any of that.
It struggles with new generators. Detectors are trained on the generators that existed when they were built. A model released since the last training round can slip past, and a detector’s accuracy quietly drops as new generators appear. This is the point Google’s own AI Overview makes too: detectors are trained on specific models and may struggle with newer ones, so combine more than one signal.
It is weak on screenshots and compressed copies. As the Bellingcat test showed, compression can flip a correct AI verdict to a wrong real one. A screenshot strips the metadata and adds a fresh compression layer, and each re-share through a messaging app or social platform adds another. By the time a photo has been forwarded three times, the statistical fingerprints the detector reads have largely been overwritten.
It over-flags edited real photos. NewsGuard’s 87% figure for lightly retouched photos is the number to remember. If the photo has been through a phone editor, a filter, or an AI retouching tool, an AI verdict may be reacting to the edit rather than to the photo’s origin. Our guide to how to tell if an image is AI generated covers the visual tells that help you separate an edit from a full generation.
It cannot check provenance. AI or Not reads pixels, not history. It cannot tell you who took a photo, when, or where it first appeared. For that, check the file’s Content Credentials with a verifier, and run a reverse image search to find where else the picture has appeared. Our Content Credentials check guide shows how to read the provenance data in a file, and it is the check to run before you trust any detector score.
What mistakes do people make with detector results?
The tool is simple to use, which is exactly why the mistakes around it are so common. These are the ones we see most.
Treating one score as a verdict. A single detector result is one vote. The NewsGuard test had five tools disagreeing on 35 of 45 images, which tells you how often a second opinion changes the picture. If the result matters, run the image through a second detector and a provenance check before you act on it.
Accusing someone on the strength of a score. A false positive is not a rare edge case. It is a measured behaviour: 1 in 15 real photos in NewsGuard’s test, and far more often for edited ones. Never tell a photographer, a seller, a job applicant or a student that their photo is AI made because one tool said so. Present the result as a question to investigate, not as an accusation.
Checking a screenshot and trusting the answer. If the only copy you have is a screenshot, the detector is reading the screenshot, not the photo. Compression and re-encoding can push the score either way. Ask for the original file before you treat the result as meaningful.
Paying before using the free checks. The free tier gives you 20 image checks. Use them first. The Pro plan’s extra volume does not improve the verdict on any single photo.
Ignoring the generator breakdown. The Data Breakdown is where the tool shows its working. A verdict with its weight spread across several generators is a weaker result than one concentrated on a single name. Read the breakdown, and if the tool cannot settle on one generator, say so when you report the result to anyone else.
Forgetting the visual check. Detectors miss things eyes catch, and eyes miss things detectors catch. Google’s AI Overview lists the same visual tells we teach: look for wrong hands, fingers, eyes and teeth, unnatural patterns and textures, and lighting or reflections that do not match the scene. A photo that fails the visual check and scores high on a detector is a much stronger case than either signal alone.
A stronger check than one detector
No single tool is reliable enough to use alone, and the comparison guides agree on this point. The competitor we measured this article against, Imagera’s detector comparison, makes the same case and even lays out a six-step method for testing detectors on your own images. Its weakness is that it never runs the test itself: it has no worked example, no AI or Not specific steps, no pricing, and no note of the account gate. Our article beats it by doing the checks and reporting what actually happened. Here is the layered workflow that follows from all of it.
- Run the photo through AI or Not, using the original file and the free tier. Note the verdict, the AI percentage, the deepfake percentage, and the generator breakdown.
- Run the same file through Hive’s free demo. If the two tools agree, your confidence rises. If they split, the split itself is the signal: slow down and verify the source rather than the pixels. Our best AI image detector tools comparison puts the main free options side by side if you want a third opinion.
- Check provenance. Open the file in the Content Credentials Verify tool and look for a signed record of what made it and what was done to it. A signed Firefly or ChatGPT record is evidence in a way a probability score never is. If the image came from ChatGPT, use OpenAI’s Verify tool; if it came from Gemini, Google’s SynthID Detector reads the watermark. No record proves nothing, but a record that names a generator settles the question.
- Reverse search the image. Google Lens and TinEye can find where the picture first appeared and whether the same image is being reused elsewhere. A photo that first appeared years ago under a different caption tells you more than any detector score.
- Look at the picture yourself. Check hands, eyes, teeth, text in the background, reflections, and whether the light falls the same way across the whole scene. Then check the context: does the account posting it have a history of real photography, and does the story around the photo hold together?
- Record everything. Keep the file name, the date, each tool’s result, and the version of the tool you used. Detectors are retrained, so a result from today may not repeat in three months, and a record is what makes your check defensible later.
That is the whole method. AI or Not is a fast, sensitive first screen with a genuinely useful free tier, a generator breakdown worth reading, and an account step you should know about before you start. Used on the original file, backed by a second detector and a provenance check, it earns its place. Used alone, on a screenshot, as proof, it will mislead you sooner or later, usually on a real photo that someone edited.
Is the AI or Not image detector free?
Yes, for a small number of checks. The free tier includes $5 in free credits and 20 AI image checks a month, plus 1 million words of text detection. You still need to create a free account before the web checker shows a result. If you only check a photo now and then, the free tier is enough.
Do I need an account to use AI or Not?
Yes. When we tested the homepage checker on 15 November 2026, clicking upload opened a Create Account panel offering sign-in with Google or Apple, or a name, email and password form. Until you create an account, the checker stays on ‘Waiting for input’ with every score at 0%, and no verdict appears.
What does the Data Breakdown mean when it names Stable Diffusion or Midjourney?
The Data Breakdown splits the AI score across the generator families the tool suspects, such as Stable Diffusion and Midjourney, plus a separate deepfake line. A result concentrated on one name is a cleaner signal than one spread thinly across several. Treat a named generator as a lead to check, not proof of which tool made the image.
How accurate is AI or Not?
It is strong on clean files and weak in the ways independent tests measure. NewsGuard’s May 2026 audit found it caught 100% of substantially altered images but wrongly flagged 1 of 15 real photos, and 87% of lightly retouched ones. The 98.9% figure on its homepage is the vendor’s own claim, measured on its own test sets.
Can AI or Not detect a screenshot of an AI image?
It can try, but the result is weaker than on the original file. A screenshot strips metadata and adds a fresh compression layer, and Bellingcat’s 2023 test found the detector called seven of ten compressed Midjourney images real after catching all 100 uncompressed. Ask for the original file before you trust the score.
What should I do if AI or Not says my real photo is AI generated?
Do not panic, and do not delete the photo. False positives are a measured behaviour: NewsGuard found AI or Not flagged 1 in 15 real photos, and 87% of lightly retouched ones. Run the original file through a second detector, check its Content Credentials, and keep the original with its camera data as your proof.







