AI Image Detector

A free AI image detector that runs on your own device. Get a 0–100 score, a tile map, and the evidence behind it.

Drop a picture here

or click to browse — one file, or a whole folder's worth

Choose pictures
JPG · PNG · WEBP · AVIF · HEIC · GIF · BMP · TIFF

Twenty-five a batch · 60 MB each · fifty checks a day · nothing uploaded

Nothing is uploaded No account to create Roughly four seconds a picture
Three steps

How a check actually runs

There is no queue and no upload. Nothing downloads merely because you visit or hover over the tool. Only after you submit an image for a check does the analyser load into that browser tab; it is not installed or written to localStorage. From then on the work happens between your own processor and the page in front of you — which is why an answer arrives in seconds and why we never hold a copy of anything.

01Hand it a picture

Drag a file in, click to browse, paste from the clipboard, or drop in a web address. Twenty-five pictures can go in at once, and they queue up locally rather than on somebody's server.

02The frame is read twice

Once as a complete image, then again as a grid of overlapping tiles. The second pass is what gives a small patch of generated content, sitting inside an otherwise ordinary photograph, somewhere to show itself.

03You get a number and the reasoning

A 0–100 likelihood, the band it lands in, the tiles that pushed it there, and any Content Credentials the file happens to carry — laid out so you are equipped to disagree with it.

Three outcomes

The three shapes an answer takes

Pictures do not sort themselves into real and fake. A genuine photograph can hold one generated object; a fully synthetic frame can be clean enough to survive a glance. So the readout separates the three cases, because acting on them differently is the entire point.

AI-generated
94 typical score

Made end to end by a model

The synthetic signature is spread evenly across the frame — every tile tells the same story. This is what output straight from a generator looks like on the grid.

AI-edited
61 typical score

A real frame with a false patch

One tile climbs while its neighbours stay quiet: a face swapped in, an object generated over the top, a background replaced, a detail painted out.

No AI signal
08 typical score

Behaves like a capture

Neither the pixels nor the file point anywhere near a generator. Every tile stays low and the headline number settles at the bottom of the scale.

In the wild

Where fake pictures actually do damage

Not a threat list for its own sake. Fifteen situations people describe when they write in, each leaning on a different part of the readout.

  • Mis- and disinformation

    A fabricated scene attached to a real event, moving faster than anyone can source it.

  • Dating and social profiles

    A generated face on a borrowed body, used to open accounts and start conversations.

  • Insurance claims

    Damage that never happened, photographed convincingly enough to pay out.

  • Marketplace listings

    A part, a phone, a bike. Generated end to end, and the deposit is real.

  • Identity documents

    A genuine card with a swapped portrait, aimed at an automated identity check.

  • Rental and property

    A flat that photographs beautifully and does not exist at the address given.

  • Job applications

    A polished headshot for a candidate nobody can find anywhere else.

  • Fabricated evidence

    An incident staged in software and submitted as a photograph of an event.

  • Charity and disaster appeals

    A disaster scene invented to collect donations before anyone can verify it.

  • Reviews and testimonials

    A customer who does not exist, endorsing a product on camera.

  • Coursework and submissions

    A figure, a diagram or a field photograph handed in as original observation.

  • Campaign imagery

    A candidate placed somewhere they never stood, timed for the week before a vote.

  • Wire and press photos

    A picture offered to a newsroom on a deadline too short to check it.

  • Before-and-after claims

    A result generated rather than photographed, sold as what the treatment does.

  • Auctions and collectibles

    A lot photographed from every angle by a model that never held the object.

Reading the number

What the score says, and how firmly

What comes back is a likelihood, never a bare yes or no. A tool that announces “AI” with no figure attached is concealing how close the call was, and the distance between 66 and 96 is the distance between worth a second look and worth acting on.

The scale is cut into five fixed bands with the decision line at 65. Those cut points come from the detector itself rather than from this page, so a 72 next winter means exactly what a 72 means this afternoon.

The five bands, what each one means and what to do about it
Score Band What it means What to do next
0 – 20 No AI signal Nothing in the pixel statistics leans towards a generative model. Behaves like a genuine capture. Still worth checking where it came from.
20 – 45 Probably real A faint synthetic reading. Ordinary retouching and compression live here too. Take it as real unless something outside the picture says otherwise.
45 – 65 Inconclusive The honest middle. Screenshots, upscales and heavy filters collect in this band. Do not carry this result into a decision in either direction.
65 – 90 Likely AI Over the decision line, but near enough that aggressive resizing could have lifted it there. Corroborate with a second source before you rely on it.
90 – 100 AI-generated A strong reading holding steady right across the frame. Work on the basis that it is synthetic.

Beneath the number sits the signal list — every piece of evidence that fed into it. A whole-frame reading is always there. A tile-by-tile reading appears whenever the picture is large enough to divide sensibly. Content Credentials appear when the file still carries them.

The tile map is the part that makes the score checkable rather than merely believable. Each tile is scored in isolation, so you can see for yourself whether the signal covers the picture evenly or crowds into one corner — and those two situations call for very different responses.

Under the hood

What the model is actually looking at

Generators and cameras leave different fingerprints, and neither is visible to a person. A camera sensor produces noise with a particular texture, then the lens, the demosaicing step and the JPEG encoder each stamp their own regularities on top. A diffusion model never touches any of that machinery; it assembles a picture out of learned statistics, and the residue it leaves behind is smooth and coherent in places a real optical chain never is.

A vision transformer trained on both populations is what separates them. It reads the frame at full size for the headline number, then walks a grid of overlapping tiles for the map — as many as nine of them on a picture with room for it. Tiles overlap on purpose, so an edit that straddles a boundary is not quietly split in half and diluted into two innocent-looking readings.

Alongside the pixels, the file itself is opened. Content Credentials — the C2PA manifest some cameras and some generators now embed — are verified against their signing chain right in the tab, which can turn a suspicion into a documented fact in one step. When the manifest and the pixels disagree, both are shown; the tool does not quietly pick a winner on your behalf.

None of this reads minds or reconstructs history. It measures how closely a frame resembles the things the model was trained to recognise, which is a strong signal and not a confession.

Why the file's own metadata cannot answer this

The obvious approach — read the EXIF, look for a generator's name — fails almost immediately in the wild. Uploading to nearly any social platform strips metadata for privacy and bandwidth. A screenshot creates a brand-new file with nothing of the original in it. And a text editor can add “Canon EOS R5” to a generated picture in about four seconds.

So metadata gets treated as it deserves: a bonus when it survives, never a foundation. The pixel signal is the only evidence in the chain that is still present after a picture has been posted, re-saved, screenshotted and sent on again — which describes nearly every picture anybody actually needs to check.

Accuracy

Where it holds up, and where it slips

On a held-out benchmark of camera originals and current-generation model output, balanced accuracy is 91.3%. That figure was measured on clean files, and clean files are not what the internet is made of.

Push the same images through JPEG quality 60 and it settles at 87.3%; at quality 40 it reaches 84.6%. Compression erases exactly the fine texture the model reads, so a picture that has been through three platforms is a harder problem than the same picture straight off the card. Publishing the decline is more useful than publishing only the best number.

What pushes a result towards the middle

  • Heavy compression, and every re-save that stacks on top of it
  • Screenshots and photos taken of a screen
  • Aggressive downscaling, or upscaling that invents texture
  • Strong denoise, beauty filters and phone night modes
  • Very small crops, where there is little left to read
  • Illustration, painting and 3D render, none of which are photographs to begin with

Two kinds of mistake are possible and they are not equally expensive. A false positive calls a real photograph synthetic, which can damage a person who did nothing wrong. A false negative lets a generated image through, which usually means the check simply failed to help.

The decision line sits at 65 rather than 50 for that reason: it deliberately trades a little sensitivity for a lower rate of accusing genuine work. Anything that would move the headline number by tightening that trade is watched as two separate figures, not averaged into one flattering one.

Read the result as evidence, never as a verdict. It is one input into a decision that should also involve where the picture came from, who benefits from it being believed, and what a reverse image search turns up.

Coverage

Trained against thousands of generators

A detector is only as good as the range of things it has seen. The model behind the deepfake page was trained on 2.7 million pictures from 4,803 distinct generators, which is what gives it a chance against one it has never met.

What this table is not: an attribution list. Neither detector will tell you which generator made a picture, and any tool that claims to is guessing at something the pixels do not carry. These are the families the training corpus covers — coverage raises the odds that a new model's output still looks familiar, and it is not a guarantee about any single file.

Generator family Made by Architecture
DALL·E OpenAI Diffusion
GPT image generation OpenAI Autoregressive
Midjourney Midjourney Diffusion
Stable Diffusion Stability AI Latent diffusion
Flux Black Forest Labs Rectified flow
Imagen Google Diffusion
Firefly Adobe Diffusion
Ideogram Ideogram Diffusion
Grok Imagine xAI Diffusion
Qwen-VL Alibaba Autoregressive
Recraft Recraft Diffusion
Seedream ByteDance Diffusion
StyleGAN NVIDIA GAN
Latent consistency models Various Distilled diffusion
Open-weight fine-tunes Community Mixed

The training set covers 4,803 generators in total; the families above are the ones large enough to name. The long tail — fine-tunes, LoRAs and one-off checkpoints — is the majority of that number and the reason the coverage matters.

In practice

What people actually bring here

Three situations account for most of the pictures dropped into this tool, and each one leans on a different part of the readout: the whole-frame number, a single lit tile, or the file evidence sitting underneath both.

Listings you are about to send money to A rental that photographs beautifully and does not exist. A part nobody manufactured. When every tile in the grid reads high at once, that is the cheapest warning available before a deposit leaves your account.
The person on the other end of a conversation Dating profiles, recruiter messages, a supplier's team page. The face can be generated while the room behind it is a real photograph — one tile burning hot against eight quiet ones is precisely that pattern.
Pictures arriving faster than they can be sourced Breaking events attract fabricated imagery within the hour. Before a picture is reposted or published, a tile map takes seconds and shows whether the whole frame is invented or one detail was added to a real one.
What you get

Everything here, at no cost, with nothing held back

Every capability of this AI image detector is unlocked: no watermark on the report, no result held behind a sign-up wall, no feature reserved for a tier that does not exist. The reason is structural rather than charitable — the model runs on your hardware, so a check costs us nothing to serve.

A map, not just a number

Drag the overlay across the picture to see where the signal concentrates, or flip to per-tile figures and read the values yourself.

Twenty-five at a time

Queue a whole batch. Every card carries its own score and verdict, and any card opens into the full breakdown behind it.

Works from a link

Paste a picture address or an ordinary page and choose from whatever it exposes. If a host refuses your browser, save the file and add it directly.

Your files stay yours

Pictures you choose are decoded and scored inside your own tab. They are never uploaded, stored, queued, logged or trained on.

Something you can hand over

Copy a written summary, save a printable PDF, or export JSON carrying the score, every tile value and the reasoning.

Calibrated, not vibes

Scores map onto published bands with a fixed decision line, so the same number always carries the same meaning.

Credentials checked properly

C2PA Content Credentials are verified against their signing chain in the browser, so a valid claim is proven rather than assumed.

A history only you can read

Past checks are kept in your own browser storage so you can return to one. The pictures themselves are never part of it.

Links that carry no picture

Share a result and the analysis travels inside the address itself. Whoever opens it reads the reasoning, never your original file.

Free

$0no card, no expiry
  • Fifty checks a day
  • No account, ever
  • Twenty-five per batch
  • 60 MB a file
  • Tile map and scores
  • Content Credentials
  • Shareable result links
  • PDF and JSON export

Paid

Not yet

No paid tier exists and none is dated. If one arrives it will be about reporting and team workflow, and everything listed beside it stays free. Nothing here is a trial with a clock on it.

What a paid tier would cover

Batches stop at twenty-five pictures and files at 60 MB, which is roughly what a browser tab handles without complaint. The daily allowance is counted on your own device and rolls over at midnight in your time zone.

Formats, limits and requirements

Formats read
JPG, PNG, WebP, AVIF, HEIC, HEIF, GIF, TIFF, BMP, JPEG XL
Largest file
60 MB per picture
Batch size
25 pictures in one pass
From a link
Direct picture addresses and public pages
On a phone
Any current mobile browser, camera roll included
App or extension
Neither — there is nothing to install
Needs
JavaScript and WebAssembly switched on
Privacy

Where your picture goes: nowhere

This is the part where most AI image detection tools ask you to trust a policy. A file you choose here is decoded and scored inside the tab and never crosses the network, which means there is no retention window to describe, no deletion request to file, and no training set it could ever end up in. This is not a policy promise layered over an upload — there is no upload for a policy to govern. One exception is worth stating plainly: the address box does fetch a remote picture over the network, and when a host blocks your browser from reading it directly, that single request is routed through a public proxy.

Chosen The page reads the file straight off your device.
Scored The model runs on your own hardware, out of our reach.
Gone Close the tab and nothing of it remains anywhere.

Read the privacy policy

By eye

What to look at before you even run a check

Running an AI image detector is the reliable path, but a trained eye catches a fair share on its own — and knowing what the model is chasing makes its answers easier to weigh. Newer generators have fixed several of these, so treat the list as clues rather than as a checklist.

  1. Hands and teeth, still the most expensive details to get right
  2. Jewellery and glasses whose frames change thickness or fail to meet
  3. Text on signs, packaging and screens dissolving into near-letters
  4. Shadows falling in directions that contradict each other
  5. Reflections that omit something standing right in front of the mirror
  6. Patterns — brick, tiling, fabric weave — that drift out of alignment
  7. Ears, earrings and collars that do not match on the left and the right
  8. Backgrounds where a crowd blurs into people with no faces
  9. Skin lit evenly everywhere, with none of a real lens's falloff
  10. Hair that merges into the background instead of ending in strands

Every one of these can appear in an ordinary photograph, and a competent generator can avoid all ten. Use them to decide what deserves a check, never to reach a conclusion on their own.

Side by side

Against the usual free web checker

Most free AI image detector sites follow one pattern: take the upload, return a percentage, ask for an email before showing the rest. Row by row, here is what that approach leaves on the table.

Capability Original or AI Typical free checker
Runs on your device, nothing transmitted Yes No Your file goes to a server
Finds a generated patch inside a real photo Yes No Whole frame only
Tile-by-tile map with individual scores Yes up to 9 No
Bands and decision line published Yes Partly A number, no thresholds
C2PA Content Credentials verified Yes signing chain No
Whole batch in one pass Yes 25 files Partly Usually one at a time
Printable PDF report Yes No
Machine-readable JSON export Yes No
Reads HEIC, AVIF and TIFF Yes Partly JPG and PNG only
Accuracy figures and failure modes stated Yes No
Daily allowance without paying Yes 50 a day Partly 10–20 in total
Usable without an account Yes No Sign up to see the result
Result link that carries no picture Yes No Hosted on their server
Works offline after the first load Yes model is cached No

“Typical free checker” describes the pattern shared by the free web tools we have used, not any single named product. Individual tools differ, and the honest comparison is with the category rather than with a competitor picked to lose.

More than pictures

Six more detectors, all running now

Synthetic media did not stop at pictures. Deepfakes, video, voice, music, written text and copied text all have their own detector here, under the same rules as this one: processing stays on your device, evidence is visible, and the limits are stated plainly.

Pictures and video

Deepfake and video detectors

A face read against 4,803 generators of training coverage, and a clip sampled at twelve points onto a timeline rather than flattened into one average.

Sound

Voice and music detectors

Cloned speech read from the spectral texture of a recording, and finished tracks checked for the residue neural music generators leave behind.

Writing

AI content detector and plagiarism checker

An essay scored passage by passage rather than as one average, and a document matched against the sources you supply. Both take PDF and Word files.

All three beta tools publish their limits beside the result. Their models load only after you choose a file, and none of the media is uploaded.

Questions

The things people ask before trusting it

Using it

Is Original or AI free?
Yes, and without an asterisk. It is a free AI image detector with no account, no card, no trial clock and no locked feature. Fifty checks a day is a fair-use line rather than a paywall, and it resets at midnight in your own time zone. Because the model runs on your device, a check costs us nothing to serve, which is what makes free sustainable rather than promotional.
Do I have to sign up?
No. There is nothing to create an account for. Your saved results live in your own browser storage, so they follow the browser rather than a profile — clearing site data clears them, and they never reach us in the first place.
Does it work on a phone?
Yes. It works as an AI image detector online in any current mobile browser, and the file picker reaches your camera roll and camera directly — there is no app and no extension to install. The first check on a new device downloads the model, so give that one a little longer over a slow connection; after that it is cached.
Can I check a picture I only have a link to?
Yes. Paste it into the address box — a direct picture link works, and so does an ordinary page — the tool gathers whichever pictures that host lets your browser read and lets you choose. Some sites block cross-origin reads; when that happens, save the picture and add the file instead.
How many can I do at once?
Twenty-five per batch, 60 MB a file. Each one gets its own card with a score and a verdict, and any card opens into the full breakdown. The limits are about what a browser tab handles comfortably, not about upselling you.

Trusting the result

How accurate is it, really?
Balanced accuracy is 91.3% on a held-out benchmark of camera originals and current model output. Push the same set through JPEG quality 60 and it is 87.3%; at quality 40, 84.6%. Compression removes the fine texture the model reads, so a picture that has crossed three platforms is genuinely harder than the same picture straight off a memory card.
Why did it flag my real photograph?
Usually because something in the processing chain smoothed away the texture that marks a capture. Heavy denoise, phone night mode, beauty filters, upscaling and repeated re-saving all push a genuine photograph up the scale. If you have the original file rather than a shared copy, check that — it frequently lands several bands lower.
Can a generated picture slip through?
Yes. Someone who understands what the detector reads can suppress the signal — hard compression, added grain, a screenshot, a round trip through a photo editor. No pixel-based AI image detector is immune to that, and any tool claiming otherwise is overselling. It is a strong filter, not a lock.
What does the tile map add?
It is what makes the score checkable. Every tile is scored on its own, so an even spread across the frame — the fully-generated pattern — is visibly different from one hot tile among quiet neighbours, which is the inserted-object pattern. Two pictures can share a headline number and mean entirely different things.
Can I use this as proof?
No, and please do not. A score is a likelihood produced by a statistical model, not a forensic examination, and it should never be the sole basis of a legal, insurance, academic, employment or disciplinary decision. Treat it as one input alongside provenance, context and a reverse image search.

Privacy and scope

Do my pictures get uploaded?
No. Files you choose are decoded and scored inside the tab and never cross the network. There is no upload endpoint for them to reach. The single exception is stated openly: the address box fetches a remote picture, and when a host blocks direct browser access, that request travels through a public proxy — a URL, never a file from your device.
Are my pictures used for training?
No, and they could not be. They never leave your device, so there is no copy anywhere for a training set to draw on. That is a property of the architecture rather than a promise about our conduct, which is the stronger of the two.
Does it detect deepfake video?
Yes, through the video detector. Its browser beta samples up to twelve frames from a clip of one minute or less and shows them on a timeline. It does not inspect every frame, track a face through motion or listen to the soundtrack, so treat it as a way to find moments for review rather than proof about the whole clip.
What about AI-upscaled or restored photos?
They tend to land in the inconclusive band, which is the correct answer rather than a failure. An upscaler invents plausible texture, and invented texture is exactly what the model reacts to — even though the underlying photograph is a real capture. Where you can, check the version before the upscale.
Which generators does it recognise?
It is not trained to name a product, and any tool claiming to identify the specific generator behind a picture is guessing more than it admits. What it reads are the statistical traces common to diffusion and transformer image models as a class, which is why it also flags output from tools that did not exist when the model was trained.

Settle it in about four seconds

Drop in the picture you have been squinting at and read the evidence yourself. A free AI image detector, no account, and the file never leaves the tab it landed in.

No sign-up. No upload. Fifty a day.

Also worth reading