You have a photo that looks almost too perfect, and you want a straight answer: was it made by AI? The Sightengine AI image detector is one of the tools built for exactly that question. It is a free online check from a company that has sold image analysis to marketplaces and dating apps for years, and it gives you more than a single yes or no. This guide runs it properly, on one known AI portrait and one real photograph, so you can see what its scores actually mean before you trust one.
The Sightengine AI image detector is a free check that reads the pixels of a photo and returns an AI probability plus a per-generator breakdown. Open the detector page, drop in an image up to 12 MB, and read the score. A high score means the image looks AI-made, but it is never proof alone.
What does the Sightengine AI image detector actually check?
Sightengine checks the pixels of the image itself, not its metadata. That matters, because metadata is stripped the moment an image is uploaded to most social networks, messaging apps or marketplaces. A photo with no EXIF data, no watermark and no C2PA record can still be checked, because the detector looks for the small patterns that synthetic images leave behind, such as odd textures, repeated details and the kind of smoothness that cameras rarely produce.
The result has two parts. The first is the GenAI score, an overall probability that the image was AI-generated. The second is the face manipulation score, which looks for face swaps and AI-driven changes to a face. Both run on the same upload, so one check covers both questions. Sightengine also lists a per-generator breakdown, which tries to say which generator made the image, from a list of more than 20 names that includes Midjourney, DALL-E, Stable Diffusion, Flux, Firefly, Ideogram, Imagen, GPT image generation, Grok Imagine, Kling and Sora for video.
What GenAI means: GenAI is Sightengine’s name for its AI-generated content model. A high GenAI score means the pixels look like the output of a text-to-image generator. A low score means they look like a camera photo. The face manipulation score is separate, and it can be zero even when the GenAI score is high, which is what happened in our own test below.
The tool is aimed at real jobs, not just curiosity. Sightengine lists misinformation, fake insurance claims, fake profiles, fake IDs, marketplace spam, fake photo evidence, impersonation and fake news as the problems its detection is built for. That is the context to keep in mind. It is a screening tool for people who handle images at scale, and a useful check for anyone else who wants a second opinion on one photo.
The short version: Sightengine reads pixels, not metadata, and returns an overall AI probability, a face manipulation score and a guess at which generator made the image.
How do you use the Sightengine AI image detector without an account?
There are two ways in, and they are not the same. The public detector page at sightengine.com/detect-ai-generated-images needs no sign-in at all. We used it on 20 November 2026 and it never asked for an account, a login or an email address during either check. The second way in is the Sightengine Detect dashboard, announced on 13 April 2026 as a no-code drag-and-drop workspace with video frame-by-frame checks and audio detection. That dashboard does need an account, so for a one-off check the public page is the simpler route.
Here is the check, step by step.
- Open the public detector page in your browser. You do not need to log in.
- Drag your image onto the page, or use the upload control to pick a file from your device. Images up to 12 MB are accepted, in JPEG, PNG, WebP or GIF format.
- Wait a few seconds while the image is analysed. The page shows the overall verdict first, with a percentage badge.
- Read the GenAI score and the face manipulation score underneath it, then scroll to the per-generator breakdown.
- Save a screenshot of the result if you need a record of it, because the page does not keep a history for you.
One warning before you upload anything. Sightengine says uploads are deleted after processing by default, and that no human reviewers look at your images. That is a reasonable privacy position for a one-off check, but it is still worth thinking about what you upload. Do not run a photo of a private person through any detector without a good reason, and never upload a photo that belongs to someone else just to test it. A detector score is not proof of who a person is, and a visual match is never proof of identity.
If the score leaves you unsure, run a second check a different way. A reverse image search of the same photo can show whether it has appeared online before, which often settles the question faster than any detector. Our guide to telling whether an image is AI-generated explains how the two methods fit together.
The short version: use the public detector page for a free check with no account, keep the file under 12 MB, and read the overall score before you look at the breakdown.
What did Sightengine say about a known AI portrait and a real photo?
A detector is only worth anything if it can tell a real photo from a fake one, so we ran two images through the public detector page on 20 November 2026, with no account and no sign-in. The first was a photorealistic portrait of a fictional woman, generated with an AI image tool that same day, saved as a 1280 by 1920 JPEG. The second was a real photograph: the official White House portrait of President-elect Barack Obama, taken on 13 January 2009 by Pete Souza on a Canon EOS 5D Mark II, saved as a 1916 by 2608 JPEG of 803,500 bytes. Both files were uploaded exactly as saved, with no edits, crops or filters applied first.
The detector got both of them right, and the gap between the two scores could hardly be wider.
| Image | Overall verdict | GenAI score | Face manipulation | Top generator scores |
|---|---|---|---|---|
| Known AI portrait of a fictional woman (1280 by 1920 JPEG) | Likely AI-generated | 99% | 0% | GPT 52%, Wan 45%, Stable Diffusion 1%, Qwen 1%, Grok 1%, all others 0% |
| Real official portrait of Barack Obama, 13 January 2009 (1916 by 2608 JPEG) | Not likely to be AI-generated or Deepfake | 0% | 0% | Every generator 0%, including Midjourney, DALL-E, Flux, Firefly, Ideogram, Imagen and Kling |
The AI portrait scored 99% GenAI with a red “Likely AI-generated” badge, and the real portrait scored 0% GenAI with a green “Not likely to be AI-generated or Deepfake” badge. Both images scored 0% for face manipulation, which is worth noting. The AI portrait is a fully generated image, not a real face that was swapped or edited, so a zero face manipulation score is the correct result for it. The two scores measure different things, and they should not be read as one number.
The per-generator breakdown is where the result gets interesting. On the AI portrait, Sightengine split its confidence between GPT at 52% and Wan at 45%, with Stable Diffusion, Qwen and Grok at 1% each and everything else at zero. That split is honest in its own way. It says the image looks like the output of a recent diffusion model, but it cannot pin down which one. The portrait was made with an AI image tool that is not one of the named generators in the list, so the breakdown is a nearest-match guess, not an identification. Treat it that way.
On the real Obama portrait, every generator scored zero, which is the cleanest possible result. A real photo taken on a camera in 2009, years before any of these generators existed, should score zero across the board, and it did. That is a useful baseline. When a detector gives a real photo a clean zero, it tells you the tool is not just calling everything AI.
The short version: in our two-image test on 20 November 2026, the known AI portrait scored 99% and the real 2009 photograph scored 0%, with the generator breakdown split across two names on the AI image.
How do you read the per-generator breakdown?
The breakdown is the most useful part of the Sightengine result, and also the easiest part to misread. It is grouped into three families. Diffusion covers the text-to-image generators, from Midjourney and DALL-E to Flux, Firefly, Ideogram, Imagen, GPT image generation, Grok Imagine, Kling, Wan, Qwen, Hunyuan, Seedream, Reve, Recraft, Krea, Higgsfield, HiDream, Cosmos, Vidu and Z-Image. GAN covers StyleGAN, the older family behind many AI faces. Other covers face manipulation.
Read the breakdown as a ranking of guesses, not as a fact. A score of 52% for GPT and 45% for Wan does not mean the image is 52% GPT and 45% Wan. It means the detector’s model found the image most similar to the outputs of those two generators, and it could not separate them cleanly. Generators share training data, architectures and styles, so their outputs overlap, and a close call between two names is normal. The honest reading of our test is “a recent diffusion model, probably GPT or Wan style”, not “made with GPT”.
That matters because the generator name is the part people quote. If you tell someone a photo was made with Midjourney, they hear a specific claim about a specific tool, and they may repeat it. The breakdown does not support that kind of claim. It supports a softer one: the image looks synthetic, and its style is closest to these generators. Keep the claim at that level and you will not overstate what the tool can do.
Compare that with a detector that gives only one number. Hive returns a single AI probability, which is simpler to read but tells you nothing about which generator might be behind the image. Sightengine’s breakdown adds that extra layer, and it is worth having when you are trying to work out whether an image came from a tool you recognise. Our roundup of the best AI image detector tools puts the two side by side, with the same honest-limits warning on each.
The short version: the breakdown ranks the most similar generators, so a split score means the tool cannot tell them apart, and the top name is a guess, not an identification.
How much does Sightengine cost, and what is free?
The public detector page is free to use for a one-off check, with no account and no payment. Behind it sits a paid service, and the pricing is worth understanding before you run a lot of images, because the free allowance is measured in operations, not in images. Sightengine’s pricing page, checked on 20 November 2026, sets the free tier at 2,000 operations a month, with a maximum of 500 operations a day. One AI image check costs 5 operations, so the free tier covers 400 checks a month, or 100 a day.
| Plan | Price | Operations | AI image checks at 5 operations each |
|---|---|---|---|
| Free | Free | 2,000 a month, max 500 a day | 400 a month, 100 a day |
| Starter | 29 US dollars a month | 10,000 a month | 2,000 a month |
| Pro | 99 US dollars a month | 40,000 a month | 8,000 a month |
For most readers, the free tier is far more than enough. Four hundred checks a month covers anyone checking the odd suspicious photo, and the daily cap of 100 checks only bites if you are screening images in bulk. If you do need bulk checks, the Starter plan at 29 US dollars a month buys 10,000 operations, and Pro at 99 US dollars a month buys 40,000. Prices are in US dollars on the pricing page, and they can change, so check the page itself before you buy.
One point to keep in mind is that an operation is not the same as a check. Sightengine sells operations, and different models cost different numbers of operations, so a check that also runs moderation or deepfake detection costs more than a GenAI check on its own. If you only ever use the public detector page, none of this matters, because that page does not ask you to count operations at all. It only starts to matter when you move to the dashboard or the API for repeat use.
The short version: the public page is free with no account, and the free tier behind it allows 400 AI image checks a month at 5 operations each.
How accurate is Sightengine, and where does it go wrong?
Sightengine says its detector is 98.3% accurate and ranked first in an independent benchmark by researchers at the University of Rochester and the University of Kansas, run on 80,000 images across several text-to-image generators. That is the vendor’s claim, and it should be read as a vendor claim, not as our finding. The published paper behind the benchmark, on the arXiv record, reports 98.72% accuracy on real images, which is close to the headline figure, but it also shows much weaker results on some edited images, including 25.12% on DreamStudio image-to-image edits and 43.76% on DALL-E image-to-image edits. A single accuracy number hides that spread.
An independent test by NewsGuard, on a low-cost tier of AI detectors, gives a more grounded picture. Sightengine flagged none of 15 authentic photos as AI, which is a clean result on real images, but it caught only 5 of 15 substantially edited images. That matches what our own two-image test showed. The tool is strong on clear cases, a fully generated portrait against a real camera photo, and weaker on the messy middle, where a real photo has been heavily edited or an AI image has been reworked.
That messy middle is where false positives live. Sightengine’s own FAQ says the detector may struggle with images from generators it has never seen, and that heavily edited images can be used to try to evade detection. The AI Overview for this keyword makes the same point, noting false positives on heavily edited real photos and weaker results on newer or custom diffusion models. Our own guide on what to do when an AI detector flags a real photo works through exactly that problem, and the advice there applies here too. A high score on a photo you know is real is a reason to check again, not a reason to accept the score.
There is also a timing problem that no detector can fix. Sightengine says it adds new generators typically within a few weeks of release, which means there is usually a short window after a model launches when detection accuracy on its outputs is lower. A photo made with a generator released last week is harder to catch than one made with a model from last year. That is worth remembering when a new image tool is in the news and its outputs start circulating.
What do users actually think of it? Opinion is mixed. Independent tests generally rate Sightengine well against other detectors. But on Reddit and in forums, users report mixed results, and some threads discuss ways to get an AI image past a detector. That split is the honest picture. A strong score on one image is useful, it is not proof.
The short version: the vendor claims 98.3% accuracy and first place in an 80,000-image benchmark, but the paper behind it shows far weaker results on some edited images, and an independent test caught only 5 of 15 substantial edits.
When should you trust the score, and when should you check further?
Trust the score most when it agrees with everything else you can see. In our test, a 99% score on a portrait we knew was AI-generated, and a 0% score on a real 2009 photograph, both lined up with what we already knew about the images. A score that agrees with the file’s history, its source and its context is a score you can rely on. A score that contradicts all of those things is a score to question.
Check further whenever the stakes are high. If you are about to accuse someone of faking a photo, reject a marketplace listing, or turn down an insurance claim, one detector score is not enough. Run the same image through a second detector, reverse search it to see whether it has appeared online before, and check for C2PA provenance or other records of where it came from. Our guide on how to tell if an image is AI-generated walks through that full sequence, and it is the right next step when a single score leaves you unsure.
Never use a detector score on its own to accuse a person of anything. A score is a probability, not a verdict, and Sightengine says so itself on its own page. Even a very high or very low score, such as 99% or 1%, should be weighed against the context, with human judgment making the final call. That is the vendor’s own wording, and it is the right rule. A photo can only suggest what it shows, and a detector can only suggest what a photo is. Neither one proves it.
The short version: trust a score that agrees with the image’s history and context, and check further with a second detector and a reverse search whenever the stakes are high.
FAQs
Is the Sightengine AI image detector free?
Yes, for one-off checks. The public detector page needs no account and no payment. Behind it, the free tier allows 2,000 operations a month, with a maximum of 500 a day. One AI image check costs 5 operations, so the free tier covers 400 checks a month, or 100 a day, which is plenty for most readers.
Do I need an account to use Sightengine?
No, not for a single check. The public detector page works without signing in, and we used it on 20 November 2026 without creating an account. The Sightengine Detect dashboard, which adds video checks and a saved workspace, does need an account, so use the public page if you only want to check one image.
What does a 99% Sightengine score mean?
It means the detector is very confident the image was AI-generated, based on the pixels it analysed. It is still a probability, not proof. Sightengine says even a very high or very low score should be weighed against the context, with human judgment making the final call, so treat 99% as strong evidence, not a verdict.
Can Sightengine tell which AI generator made an image?
Only as a best guess. The per-generator breakdown ranks the generators whose outputs look most similar to your image, and it often splits its confidence between two names. In our test it gave GPT 52% and Wan 45% for one portrait. Treat the top name as a hint about the style, never as proof of which tool was used.
Can Sightengine be wrong about a real photo?
Yes. No detector is 100% reliable, and Sightengine says so on its own page. Heavily edited real photos can score as AI, and images from brand-new generators can slip through. An independent NewsGuard test found Sightengine flagged none of 15 authentic photos, but it caught only 5 of 15 substantially edited images, so check a surprising result twice.
Is Sightengine better than Hive or AI or Not?
It depends on what you need. Sightengine adds a per-generator breakdown that Hive does not offer, and it ranked first in one independent benchmark, but every detector has blind spots. The honest approach is to run a suspicious image through two detectors, reverse search it, and compare the results, rather than trusting any single tool on its own.







