An AI image detector returns a probability that a picture was generated. Ours does that, free and with no account, and it is not the strongest method available to you. That ranking is below, and a classifier comes third.
The reason is worth knowing before you rely on one. Generated images carry statistical traces, and ordinary handling destroys them. A screenshot, a re-encode, a social platform resize or a crop can strip the signal a detector depends on, which means a low score on a heavily-processed image tells you very little.
A classifier is third. That ordering is the useful part of this page.
| Approach | What it establishes | When it fails |
|---|---|---|
| Content credentials / provenance | A signed record of how the file was made | Absent, or stripped by re-upload |
| Reverse image search | Where else the image exists, and since when | Genuinely new images |
| A trained classifier score | A probability, and no provenance | Screenshots, re-encodes, crops |
| Camera and file metadata | Weak corroboration | Routinely stripped by platforms |
| Looking for visual artifacts | Sometimes decisive, often nothing | Current models, mostly |
Provenance is first because it is the only one that is evidence rather than inference. A cryptographically signed credential describing how a file was created answers the question directly. Its weakness is availability: most images do not carry one, and re-uploading through a platform frequently strips what was there. Absence of a credential proves nothing at all.
Reverse image search is second and chronically under-used. If the image has been online since 2015, it was not generated last week. It is free, it takes thirty seconds, and it resolves a surprising share of real questions outright.
A classifier is third. It is instant and works on any image, which is why it is the thing everyone reaches for first, and it returns a probability about pixels rather than a fact about origin.
Visual artifacts are last on purpose. The hand-and-teeth advice circulating online was written against models from several generations ago. Current generators mostly do not make those mistakes, so "the hands look fine" is not reassurance.
Narrower than this category usually advertises, so here it is precisely.
Upload an image and you get back one AI-likelihood score, a human score and a verdict, plus the filename and size. Detection runs through Sightengine, a specialist third-party provider, and the result is labelled with the detector that produced it. Three scans a day, no account, no card.
Three things it does not give you, stated plainly because they are commonly advertised in this category:
It does not tell you which generator made the image. The scan returns a single probability. It does not distinguish one image model from another, and a result naming the specific generator is not something our scan produces.
It does not localise. There is no pixel-level overlay marking which regions read as generated. One image, one number.
It does not read provenance. It does not parse content credentials or camera metadata, which is why those sit above it in the ranking as separate things for you to check.
An accuracy figure. We have not run a benchmark over a labelled set of generated and photographic images, and a percentage we did not measure is exactly the kind of number this category is full of. If you want to know how it behaves on your material, three free scans on images where you already know the answer will tell you more than any vendor claim.
The failure mode most likely to mislead you in real use.
Classifiers read statistical regularities left by the generation process. Those regularities live in fine detail, and fine detail is the first thing to go when an image is re-encoded, resized, cropped, screenshotted or passed through a messaging app. Each step degrades the evidence.
So the practical rule: the result is only as good as the file you have. A score on the original file means something. The same score on a screenshot of a screenshot from a group chat means considerably less, and it will skew toward "human" rather than flagging uncertainty, because the traces are simply gone.
Which leads to the one piece of advice worth more than the score. Get the original file if you can. Ask the sender for the file rather than a screenshot. That single step improves any method here, including the two ranked above the classifier, because provenance data and metadata survive in originals and die in screenshots.
Establish that a photograph is real. These tools estimate whether an image looks generated; a low score is not a certificate of authenticity, and a photograph that has been heavily edited can read oddly. And none of them can tell you whether an image is misleading, which is usually the actual question: a genuine photograph with a false caption is a real photograph.
Separately, and a different job entirely: whether text was AI-written is the text detector, and it shares nothing with this beyond the name.
The daily count gates text scans and rewrites; image scanning has its own count.
| Free | Starter | Pro | |
|---|---|---|---|
| Price | $0 | $9.99/mo · $7.49 yearly | $19.99/mo · $14.99 yearly |
| Actions a day (shared) | 3 | 20 | Unlimited |
| Words per scan | 1,500 | 5,000 | 7,500 |
| Words per humanization | 260 | 5,000 | 10,000 |
| Detection a month | No cap | No cap | No cap |
| Humanizer a month | No cap | 40,000 words | 80,000 words |
One action is one scan, one file upload or one rewrite, from a single shared daily count. That daily number is the gate that matters: there is no monthly ceiling on any plan, so what separates the tiers is how often you can work and how long a single piece can be, not how much you get through in a month.
Read the free row together rather than line by line. No monthly cap sounds unlimited and is not: three actions a day at up to 260 words a rewrite is roughly 780 rewritten words in a day. It is a real allowance for checking and fixing specific passages, and it is not a plan for processing a backlog.
Current prices, the Business and Enterprise tiers and regional pricing are on the plans page, which is the page we keep up to date.
Image scanning has its own count too. It is 3 scans a day with no account, 5 on a free account and 25 on Starter, so the row above does not cap it. Images are not text: there is no word or line limit, only the number of scans.
Judge by method rather than brand, because a classifier is not the strongest option available to you. Content credentials and provenance metadata come first, since a signed record of how a file was made is evidence rather than inference. Reverse image search is second and badly under-used: if the picture has been online for years it was not generated last week. A trained classifier, which is what our tool is, comes third, because it returns a probability about pixels rather than a fact about origin.
Ours cannot, and we are stating that because it is widely advertised. Our scan returns a single AI-likelihood score, a human score and a verdict. It does not distinguish one image model from another, and it does not produce a pixel-level overlay of which regions read as generated. If naming the specific generator is what you need, treat any claim to do it as something to verify rather than assume.
Because the signal was destroyed before you scanned it. Classifiers read fine-detail statistical regularities from the generation process, and re-encoding, resizing, cropping, screenshotting or sending through a messaging app degrades exactly that detail. The result skews toward human rather than flagging uncertainty, because the traces are gone rather than ambiguous. Always scan the original file if you can get it.
We are not publishing a figure, because we have not run a benchmark over a labelled set of generated and photographic images, and an unmeasured percentage is the kind of number this category is already full of. What we can tell you is the method: detection runs through Sightengine, a specialist third-party provider, and the result is labelled with the detector that produced it. Three free scans on images where you already know the answer will tell you more about its behaviour than any vendor claim.
No. These tools estimate whether an image looks generated, which is not the same as certifying authenticity, and a heavily edited photograph can read oddly for reasons unconnected to AI. The more important limit is that no detector can tell you whether an image is misleading, which is usually the real question: a genuine photograph with a false caption is still a genuine photograph, and every detector will say so.
Yes: three scans a day with no account and no card, five on a free account and twenty-five on Starter. Because images are not text there is no word or line limit, only the number of scans. These are counted separately from the shared daily actions that gate text scans and rewrites, so an image scan does not spend one of those.
Three scans a day, no account, no card.
Open the tool →The different job: was the writing AI-generated.
Open the detector →Whether what an AI wrote is true, not whether a picture is real.
See the comparison →Twenty-five tools, and what each one actually does.
Browse the tools →Three scans a day, no account. Try images where you already know the answer and see how it behaves.