94
typical score
Made by a model
Synthetic texture reads consistently across the whole frame. Typical of an image that came straight out of a generator.
Free No account, 50 checks a day
AI Image Detector
Find out in seconds whether a picture was AI-generated or AI-edited, with a 0-100 score and a map of the areas that triggered it.
Drop images here
or click to choose — one file or a whole batch
Choose imagesJPG · PNG · WEBP · AVIF · HEIC · GIF · BMP · TIFF
Paste a direct image link, or a public page — a blog post, a marketplace listing, a product page — and the tool collects the images the host lets your browser read.
Up to 25 images at once, 60 MB each, 50 a day.
No sign-up Your files stay in your browser Results in ~4 seconds
Three steps
How to check if an image is AI generated
Drop in a photo or paste a link to detect AI-generated images in about four seconds. The free AI image detector returns a 0-100 score, a plain verdict and a map of the regions that triggered it. JPG, PNG, WebP, AVIF, HEIC, GIF, TIFF and BMP all work, and nothing leaves your device.
Step 1Upload or paste the image
Drag a file in, click to browse, or paste a link to an image or a public page. JPG, PNG, WebP, AVIF, HEIC, GIF, TIFF and BMP all work, up to 60 MB per file.
Step 2Let the analysis run
Scoring takes about four seconds on a typical laptop. Everything runs inside your browser tab, so the file you picked is never uploaded and nothing is stored once you close the page.
Step 3Read the result
You get a verdict, a 0–100 probability, the band it lands in, a breakdown of every signal that contributed, and a region map showing which parts of the frame carried the signal.
In the wild
The checks people run first
Three situations account for most of the images dropped into this tool. Each one needs a different part of the result: a whole-frame score, a single lit region, or the file evidence underneath.
What it detects
Fully generated, partly edited, or genuinely shot
Most checkers give you one number and leave you guessing. This one separates the three cases that actually matter, because they mean very different things.
61
typical score
Real photo, fake part
One area lights up while the rest stays clean: a swapped background, an inserted object, a generative fill or an erased detail.
08
typical score
Looks genuinely shot
Nothing in the pixels or the file points to a model. Every region stays low, and the score sits at the bottom of the scale.
Reading the result
What the score means, and how sure the tool is
The tool returns a probability, never a flat yes or no. A detector that says “AI” with no number attached is hiding how close the call was, and the gap between 66 and 96 is the gap between a second look and something you can act on.
The scale splits into five fixed bands with the decision line at 65. A 72 today means what a 72 meant last month.
| Score | Band | What it means | What to do next |
|---|---|---|---|
| 0 – 20 | No AI signal | Nothing in the pixels points to a generative model. | Reads as a genuine capture. Check the context anyway. |
| 20 – 45 | Probably real | A low synthetic signal. Light editing and compression sit here too. | Treat as real unless something else contradicts it. |
| 45 – 65 | Inconclusive | The uncertain middle. Screenshots, upscaling and filters land here. | Don't treat the result as proof either way. |
| 65 – 90 | Likely AI | Above the decision line, but close enough that heavy resizing could have moved it. | Verify against a second source before you act. |
| 90 – 100 | AI-generated | A strong signal spread across the whole frame. | Treat as synthetic. |
Under the score sits the signal list: every piece of evidence that contributed. A whole-frame score always appears, a region reading when the image is big enough to divide, and content credentials when the file carries them.
The region map is where the score becomes checkable. Each tile is scored on its own, so you can see whether the signal covers the frame evenly or sits in one corner.
Features
Built to be believed, not just used
Every part of the result is explained, sourced and honest about its limits. That is the difference between an AI image detector you can cite and one you have to take on trust.
Region heatmap
See exactly where the synthetic signal sits. Slide the overlay, or switch to per-tile scores to read the numbers directly.
Batch checking
Queue up to 25 images at once. Every card shows its score out of 100 and its verdict; open any one for the full breakdown.
Works with image links
Paste an image or public page and choose from the images it exposes. If a host blocks browser access, save the image and add the file directly.
Private by default
Files you choose are read and scored inside your own browser. They are never uploaded, stored, queued or logged anywhere.
Evidence you can share
Copy a written summary or download a full JSON report with the score, every region value and the reasoning behind the verdict.
Calibrated, not guessed
Scores map to published bands with a fixed decision line, so 72 always means the same thing, and the limits are stated rather than buried.
Shareable reports
Send a result to a client or a colleague as a link. The analysis travels inside the link itself, so the picture is never uploaded and the person opening it sees the reasoning, not your file.
The short version
What an AI image detector actually does
An AI image detector is a tool that reads the pixels of a picture and estimates how likely it is that a generative model produced them. It looks at the statistical texture of the image itself, not the file's metadata, then returns a probability, a confidence band and the evidence behind the call.
Your eye stopped being a reliable test around 2024. The tells people were taught to look for — six-fingered hands, melted text, glassy skin — are the ones the newest models fixed first.
So this AI picture detector works below the level you can see. A model trained on millions of real and generated frames reads how each pixel sits against its neighbours, and that relationship differs between a camera sensor and a diffusion model in ways that survive a screenshot. A stripped EXIF block doesn't blind it.
0–100Confidence score
25Images per batch
9Regions mapped
0Files sent to a server
The hard cases
The hard images other detectors miss
Fully generated pictures are the easy case. Almost any AI image checker catches a clean Midjourney render straight out of the generator. The trouble starts with mixed frames, recompressed files, and pictures that have been through four platforms since they were made.
Real photos with AI edits
An object added, a person removed, a background swapped. Only part of the frame is synthetic, so one whole-image score averages the fake region away against the real pixels around it.
AI-upscaled and retouched photos
The capture was real. The detail wasn't. Upscalers and retouch models invent pore texture, hair strands and fabric weave that no sensor recorded, which pushes a genuine photograph up the scale.
Screenshots and re-uploads
Every platform re-encodes what you post and throws the file data away. Metadata checks die there. The pixel signal degrades but survives, which is why this reads the frame rather than the header.
Face swaps and synthetic portraits
A real body with a generated face, or a face belonging to nobody. The swapped region scores far above the rest of the frame, so it shows up as one hot area rather than an even wash.
Stylized art, anime and illustration
Not photographic is not the same as not real, and confusing the two is how false accusations start. A hand-drawn illustration has no sensor noise either, so an AI art detector has to weigh pattern over style.
Images edited to dodge detection
Noise stacked on, filters piled up, the file resaved a dozen times. Edits that bury generator traces also visibly damage the picture, and heavy damage pushes a result into the inconclusive band.
Accuracy
How accurate the AI image detector is
On the benchmark set, the model reaches 91.3% balanced accuracy on clean images. Re-encode those same images as JPEG at quality 60 and it holds 87.3%. Drop to quality 40 and it holds 84.6%. Balanced accuracy means real and generated images are weighted equally, so the figure isn't inflated by a lopsided test set.
Those figures come from the model's published benchmark, not an internal test. They describe the pixel lane on that dataset, and your image is not that dataset.
What lowers reliability
- Very small images, and anything under 576 pixels on the shorter side, which is too small to divide into regions
- Screenshots of screenshots, where the frame has been re-encoded several times over
- Output from a generator released in the last few weeks
- Heavy filters, thick borders, collages and text overlaid on the image
- Mixed frames where only a small region is synthetic and the rest is a genuine capture
A false positive is a real photograph scored above the line. Heavy retouching, upscaling and aggressive compression are the usual causes, and uploading the original rather than a screenshot fixes most of them.
A false negative is a generated image scored below the line. A new generator, a small synthetic patch inside a real frame, or a heavily laundered file will do it.
A result is evidence, not a verdict. Weigh it with everything else you know about where the image came from.
Under the hood
How the detector reads an image
Two lanes run on every image, kept separate on purpose. One reads the pixels. One reads the file. Neither silently overrules the other, because they fail in different ways.
The pixel lane is a vision transformer trained on the Community Forensics dataset — millions of real photographs and generated images across a wide spread of models. It reads the frame at 384 pixels square and returns a raw likelihood, calibrated so the number maps onto the published bands. What it picks up is the statistical residue of generation: diffusion models and generative adversarial networks (GANs) leave generator-specific signatures in how neighbouring pixels relate, measurably different from what a sensor and a lens produce.
The same model then runs across the frame as a grid of overlapping tiles, each scored on its own. This catches partial edits, because a swapped face lights up its own tiles while the untouched photograph around it stays low. An image needs 576 pixels on its shorter side to divide into enough regions.
The file lane reads provenance. It checks content provenance and authenticity (C2PA) credentials and verifies the signature cryptographically, then scans for embedded generation parameters, generator names and SynthID watermark declarations. A signed credential bound to these exact pixels is strong. A loose text marker is weak, since anyone can copy it in, and the two are graded differently.
The lanes combine into one verdict, and disagreement is reported rather than hidden. A low whole-frame score with most regions above the line returns a conflicting-evidence result that says exactly that.
What the tool does not do: it doesn't run error level analysis, sensor pattern noise matching, or shadow-and-reflection geometry checks. Those are real forensic techniques and they're not in this build. Claiming them would be easy and it would be false.
Why pixels
Why metadata and watermarks fail on real-world images
Metadata is stripped the moment an image is posted to almost any platform, so a metadata-based check fails on precisely the images people need to verify. The picture a stranger sent you has been through a compression pipeline with no file data left to read.
EXIF holds camera make, lens, exposure and sometimes location. It survives a straight file copy. It does not survive an upload to a social platform, a messaging app, a screenshot, or most re-saves.
Content credentials are a real improvement, and this tool verifies them cryptographically when present. They only hold when every tool in the chain preserves them, and one save in an editor without support breaks the chain. Adoption is thin, so most images carry none, and their absence proves nothing. Invisible watermarks work when the generator applied one and nothing since disturbed it; a crop or a re-encode can remove the mark, and many generators never apply one.
That leaves pixel-level analysis as the only method with anything left to read after a screenshot. It isn't perfect, and the sections below are honest about where it slips. It is the only approach that survives how images actually travel.
Coverage
Which generators it recognises
Two things happen here, and conflating them is how tools overclaim. The pixel model doesn't identify brands: it separates generated texture from camera texture in general, which lets it generalise to a model it has never seen. The file scan is separate, and it can name the exact tool.
What the pixel analysis does
It returns a probability that the frame is synthetic, whatever made it. No brand is attached to that number. A generator released in the last few weeks sits in a window where accuracy on its output is lower, and that holds for every detector.
What the file scan can name
When an image still carries credentials or generation parameters, the scan names the tool. These markers survive far less often than people expect.
- Midjourney
- OpenAI DALL·E and GPT image
- Stable Diffusion and SDXL
- Adobe Firefly
- FLUX (Black Forest Labs)
- Google Imagen and Gemini
- Leonardo.Ai
- Ideogram
- Recraft
- xAI Grok
- Runway
- SynthID watermark declarations
Product names belong to their owners. A listing here indicates detection coverage, not affiliation.
The skeptic's question
Can an AI image be edited to slip past detection?
Yes, people try, and it works often enough to be worth understanding. Detection is a moving target and any page claiming otherwise is selling something.
The attempts fall into a few families: adding noise, stacking heavy filters, re-saving repeatedly, cropping tight, screenshotting, and passing the image through a second model. Each degrades the generator traces the detector reads.
Here's the part that gets left out. Those same edits degrade the picture. Push hard enough to bury the signal and you've visibly damaged the image, and that damage is itself measurable. A heavily laundered file usually lands in the inconclusive band rather than sailing through as real.
The model was trained on post-processed and adversarially edited samples, not only clean output, so moderate compression and ordinary filters don't defeat it. The honest limit: an image pushed far enough ends up inconclusive, and the tool says so rather than guessing.
This section describes categories of attempt because people search for them and deserve a straight answer. It deliberately gives no settings, no steps and no tooling for defeating detection.
By eye
Signs an image was made by AI
Before you upload anything, eight things are worth a look. None of them is decisive on its own.
- Count the fingers, then look at how hands meet objects.
- Check teeth, ears and jewellery, where repeated small shapes go wrong.
- Read any text on signs, labels and packaging.
- Look into reflections in eyes, mirrors and glass.
- Trace shadow direction back to the light source.
- Scan crowds, fabric and foliage for repeating patterns.
- Follow the edge where hair or fur meets the background.
- Look for skin with no pores, no asymmetry and no blemishes.
Now the uncomfortable part. The newest generators clear most of this list. Hands are largely fixed, text is close, and skin texture convinces at full resolution. These tells still catch older output, which is a lot of what circulates, but a clean pass through all eight proves nothing. That gap is what the pixel analysis fills.
In practice
What people check with it
Product photos and marketplace listings
A generated product shot sells an item nobody made. Check the listing image before you pay, especially when one photo looks cleaner than the seller's others.
Receipts, invoices and proof of payment
A fabricated receipt opens most refund fraud. Run the image and watch the region map, since these are usually real templates with edited numbers.
ID documents and verification selfies
Synthetic identity photos and edited documents reach onboarding queues daily. The region map matters more than the score, because the forged part is often a small patch.
Profile pictures and dating app photos
A face that belongs to nobody is cheap to make and hard to spot. Check a profile picture before you trust the person, particularly when there's only one.
Insurance and damage claim photos
Generated damage and edited evidence both show up in claims. A partial edit on a genuine photograph is the usual pattern, so read the tiles, not the headline.
Delivery and proof-of-completion photos
Proof-of-delivery images are easy to fake and rarely examined. Check them when a package is disputed or a job was signed off from a photo alone.
News photographs and viral posts
A convincing fake moves fastest in the first hour of a breaking story. Verify before you publish or repost, and keep the report with the file.
Artwork, illustrations and portfolios
Both directions cause harm: passing off generated work, and accusing a human artist wrongly. Read the evidence rather than the verdict, since a stylised look proves nothing.
Property, rental and travel listings
Generated interiors and enhanced exteriors show a place that doesn't exist. Check listing photos before you send a deposit for somewhere you haven't visited.
Charity appeals and fundraising images
Emotive generated imagery drives donation scams. Check the photo behind an urgent appeal before you give, and before you share it onward.
Consequences
What goes wrong when a fake image gets through
The item never existed
A seller lists furniture with photos generated to order, takes payment from a dozen buyers, then closes the account. Each buyer has a bank dispute and a photo that matches nothing.
A face that belongs to nobody
Synthetic portraits let one operator run many personas without the reverse-image-search risk of stolen photos. Victims lose money over months, and the usual advice to check whether the picture appears elsewhere quietly stops working.
Onboarding accepts a person who isn't there
Generated selfies and edited documents pass know-your-customer queues built for human forgery. The account that opens is used for laundering, and the real person whose details were assembled finds out much later.
Damage that was added afterwards
A genuine photograph of a real car gets a generative fill of a dent that was never there. Whole-image scoring reads it as mostly authentic, because it mostly is, and the claim pays out.
The first hour belongs to the fake
During a developing story a fabricated image spreads while verification catches up. Corrections reach a fraction of the people who saw the original, and the picture circulates for years afterwards.
An artist competes with an imitation
Work generated in a living artist's style undercuts commissions and floods the same tags. The reverse lands too: artists accused of generating their own work, on nothing but someone's impression.
Who it's for
Who this is built for
Everyday users
Someone sent you a picture and it feels off. Drop it in, read the number, move on. No account, nothing to install, and the file stays in your browser.
Sellers and buyers
You're about to pay for something you've only seen photographed. The region map matters most, since edited listing photos are usually real pictures with one invented detail.
Trust and safety teams
You're triaging a queue, not examining one image. Batch handles 25 at a time with a per-image verdict, and every result exports as JSON you can attach to a case.
Journalists and fact-checkers
You need to defend the call to an editor. The signal breakdown shows what contributed and how strongly, the bands are published, and the limits sit on the page rather than buried.
Educators and students
You're teaching people to weigh evidence rather than trust a verdict. The region map makes the reasoning visible, which beats a bare percentage in a classroom.
Artists and designers
You're checking a submission, or defending your own work against an accusation. Stylised images are the hardest case, and the tool shows the uncertainty rather than papering over it.
Use cases
Who checks images, and why
Newsrooms
Verify a submitted photo before it runs, and keep the report with the story file.
Marketplaces
Catch generated product shots and fake listing photos before they go live.
Dating & social safety
Check a profile picture before you trust the person behind it.
Education
Review submitted work and show students what the evidence actually looks like.
Insurance & claims
Screen claim photographs for generated damage or edited evidence.
Brand & agencies
Audit supplier and stock imagery so nothing generated slips into a campaign.
Next steps
You found out it is AI. Now what?
-
Save the evidence first
Download the JSON report and keep the original file. Posts get deleted and accounts close while you're deciding what to do.
-
Search for the earliest copy
Run a reverse image search for the earliest copy. An image online since 2019 is unlikely to be diffusion output, whatever the pixels suggest.
-
Look at the account behind it
Check the seller, profile or page for other signs: account age, whether the other photos share the same signature, and whether the picture appears under different names.
-
Report it where you found it
Most platforms now have a synthetic media policy. Report under that specific rule rather than a generic spam category, and attach the score if the form allows it.
-
Don't re-share it unlabelled
Reposting a fake to warn people spreads it further, and the warning separates from the image within a couple of shares. Label it in the post itself.
Alternatives
How this compares to other ways of checking
| Method | Works after a screenshot | Explains the result | Handles partial edits | Cost |
|---|---|---|---|---|
| Checking with your own eyes | Yes | No | Rarely | Free |
| Metadata or EXIF viewers | No | Partly | No | Free |
| Watermark or credential checks | Only if preserved | Partly | No | Free |
| Reverse image search | Sometimes | No | No | Free |
| Asking a general AI chatbot | Unreliable | Sounds confident, often wrong | No | Varies |
| This detector | Yes | Yes | Yes | Free |
Asking a chatbot is the most common mistake. A general language model produces a fluent, confident paragraph with no forensic analysis behind it: no calibrated threshold, no benchmark, no way to show which pixels drove the answer. It guesses in complete sentences, and the fluency is what persuades.
Feature by feature
How it compares to a typical online checker
Most free checkers do one thing: take your upload and hand back a number. This is what that leaves out, row by row.
| Feature | Best AI Image Detector | Typical online checker |
|---|---|---|
| Runs in your browser, nothing uploaded | Yes | No Files sent to a server |
| Finds AI edits inside a real photo | Yes | No Whole image only |
| Per-region heatmap with tile scores | Yes 9 tiles | No |
| Published score bands and decision line | Yes | Partly Score only, bands unstated |
| Content Credentials (C2PA) verification | Yes | No |
| Batch checking in one pass | Yes Up to 25 | Partly Often one at a time |
| Printable PDF report | Yes | No |
| Machine-readable JSON export | Yes | No |
| HEIC, AVIF and TIFF input | Yes | Partly JPG and PNG only |
| Accuracy figures and limits published | Yes | No |
| Free daily allowance | Yes 50 images | Partly 10 to 20 in total |
| No account needed | Yes | No Sign-up to see a result |
| Shareable report link | Yes no upload | Partly Hosted on their server |
| Ad-free | Yes | No |
The right column describes the pattern common to free web-based checkers, not any one product. Individual tools do better on individual rows, and the comparison is about the common case rather than the best case.
Access
Free to use, no account needed
Every check is free. You don't need an account, an email address or a payment method, and no trial expires. The reason is structural rather than generous: the model runs inside your browser on your own device, so there's no per-check server cost to recover. The allowance of 50 images a day is a fair-use line rather than a paywall, and it resets overnight. Detection in this free AI image detector stays free.
Free
$0no card, no expiry- 50 checks a day
- No account
- 25 images per batch
- 60 MB per file
- Region map
- Content credential verification
- Shareable report links
- PDF and JSON export
Paid
Not yetThere is no paid tier and no date for one. If it arrives it will cover reporting and team workflow, and everything listed beside it stays free. Nothing here is a trial that runs out.
What a paid tier would addBatches cap at 25 images and files at 60 MB, the limits of what a browser tab handles comfortably. The daily allowance is counted on your own device and resets at midnight in your time zone.
Privacy
What happens to the image you upload
Nothing is uploaded. A file you choose is read and scored inside your browser tab and never travels to a server, so there's no retention period to describe and nothing to delete. Your images are never used for training, because they never arrive anywhere they could be. One exception: the URL box fetches a remote image over the network, and when a host blocks direct access that request routes through a public proxy.
Choose
The file is read from your device by the page.
Score
The model runs on your own hardware, offline to us.
Discard
Close the tab and nothing remains.
Compatibility
Formats and devices it works on
It runs in any modern browser on any device with nothing to install. There's no app and no extension, which is usually what people searching for one want: an AI image detector online that works on a phone without a download. The file picker reaches your camera and photo library. Layered formats like PSD aren't read directly, so export a flattened PNG or JPG.
- Formats
- JPG, PNG, WebP, AVIF, HEIC, HEIF, GIF, TIFF, BMP, JPEG XL
- Maximum file size
- 60 MB per image
- Batch size
- 25 images at a time
- URL input
- Direct image links and public pages
- Mobile
- Any modern mobile browser
- App or extension
- Neither — nothing to install
- Requirements
- JavaScript and WebAssembly enabled
At volume
Checking images in bulk
Batch takes 25 images at a time. Each is scored in turn and its card updates the moment it finishes, so results appear while the rest run. Any card opens the full report, and the batch exports as one JSON file with the score, band, region statistics and signals for each image.
There is no API yet
No hosted endpoint exists at the moment, and describing one that doesn't ship would waste your time. The detector runs entirely client-side, so an integration today means embedding the page rather than calling a service. If you need programmatic access, say so through the contact page — that demand is what would justify building it.
Compared with a public checkpoint
The pixel model is an open research model, and its licence and origin are stated rather than hidden. A maintained deployment adds calibration, the tiled region pass, provenance reading and published bands around the raw classifier output.
FAQ
Questions people ask
Using the tool
What is an AI image detector and how does it work?
An AI image detector estimates whether a picture was produced by a generative model, by reading the pixels rather than the file data. A vision model trained on millions of real and generated frames measures how each pixel sits against its neighbours, since sensors and diffusion models leave different traces. You get a 0-100 probability, a band, the contributing signals, and a region map.
How do I check if an image is AI generated?
Drag a file onto the page, click to browse, or paste a link to an image or a public web page. Scoring takes about four seconds. Read the verdict and probability first, then open the region map to see whether the signal covers the whole frame or sits in one area.
Is this AI image detector free to use?
Yes. There is no trial period, no payment method and no account. The free allowance is 50 images a day, which resets overnight and covers almost everyone who is not doing this for a living. Batches cap at 25 images and files at 60 MB, which are practical limits of running a model inside a browser tab rather than commercial ones. No paid tier exists today.
Do I need an account to use it?
No. There's no sign-up, no email address and no login. Open the page and check an image. Because the analysis runs on your own device rather than a server, there's no usage to meter and no reason to identify you.
Can I check an image from a URL instead of uploading a file?
Yes. Paste a direct image link, or paste a public page and pick from the images it exposes. Fetching a remote image is a network request, and when a host blocks browser access the link routes through a public proxy. For sensitive work, download the image and add the file instead.
Can I use it on a phone, and is there an app?
Yes to the phone, no to the app. The site runs in any modern mobile browser with nothing to install, and the file picker reaches your camera and photo library. There's no native app and no extension. The first check downloads the model, so it takes longer than later ones.
Accuracy and trust
How accurate is this AI image detector?
The model reaches 91.3% balanced accuracy on clean benchmark images, 87.3% after JPEG compression at quality 60, and 84.6% at quality 40. Balanced accuracy weights real and generated images equally, so a skewed test set can't inflate it. Small files, repeated screenshots, heavy filters and brand-new generators all lower reliability.
How can I tell if a photo is AI generated when it looks completely real?
You mostly can't, which is the problem this solves. The visual tells people learned — bad hands, garbled text, plastic skin — are the ones the newest models fixed first. The analysis works below what your eye resolves, measuring pixel statistics that differ between a sensor and a generative model.
Can it detect an image where only part of the photo was AI edited?
Yes, and this is the case most detectors handle worst. One whole-image score averages a small synthetic patch against the genuine pixels around it, so the edit vanishes into the average. The tool scores the frame again as overlapping tiles. One hot region inside a clean frame signals a generative fill or a swapped face.
Does it work on AI art, anime and illustrations?
Yes, with more uncertainty than on photographs. A stylised look is not evidence of AI origin, and treating it that way is how human artists get falsely accused. A hand-drawn illustration has no sensor noise either, so read the region pattern and the confidence band rather than the headline number.
Why did it flag my real photo as possibly AI?
Heavy retouching, AI upscaling, strong filters and aggressive recompression all push genuine photographs up the scale, because each replaces recorded detail with invented or averaged detail. Small images and screenshots do the same. Upload the original file rather than a screenshot, and the score usually drops.
Can someone edit an AI image so the detector misses it?
Yes, and the attempts are well known: added noise, stacked filters, repeated re-saving, tight crops, screenshots, and a pass through a second model. The trade-off is that these edits degrade the picture, and heavy degradation lands a result in the inconclusive band rather than clearing it as real.
Is this better than asking a general AI chatbot whether an image is AI?
Yes, and the gap is larger than it looks. A chatbot produces a fluent, confident paragraph with no forensic analysis behind it: no calibrated threshold, no benchmark, no way to show which pixels drove the answer. A classifier trained to separate generated from captured images returns a number you can check.
Can it detect deepfakes and face swaps?
Yes, and the region map is what makes it work. A deepfake differs from a fully generated image: the body, background and lighting are a real capture, and only the face was replaced. Tile-level scoring surfaces the swapped area as a hot region while the surrounding photograph stays low.
Is an AI detection result proof?
No. It is evidence, and it should be weighed with everything else you know about the image. No detector reaches certainty, this one included, and any tool claiming to be definitive is misleading you. A high score with a clear region pattern is strong. A borderline score on a compressed screenshot is not.
Technical
Which AI image generators can it detect?
The pixel model is generator-agnostic: it separates generated texture from camera texture in general, which lets it generalise to models it has never seen rather than matching a fixed list. Separately, the file scan can name Midjourney, DALL·E, Stable Diffusion, Firefly, FLUX, Imagen, Leonardo.Ai, Ideogram, Recraft, Grok and Runway when markers survive.
Does it work without metadata or a watermark?
Yes. The analysis reads the image content itself, so a stripped EXIF block, a missing watermark or a screenshot doesn't blind it. Metadata and content credentials are read as an extra lane when present, and treated as supporting evidence rather than the deciding factor.
What image formats and file sizes are supported?
Supported formats include JPG, PNG, WebP, AVIF, HEIC, HEIF, GIF, TIFF, BMP and JPEG XL, up to 60 MB per file and 25 files per batch. Layered formats such as PSD aren't read directly; export a flattened PNG or JPG instead. Anything your browser decodes as an image generally works.
Is there an API for AI image detection?
No. There's no hosted endpoint at present. The detector runs entirely in the browser, so there's no server to call. If you need programmatic access for a moderation or verification pipeline, get in touch through the contact page.
Can I check several images at once?
Yes, 25 per batch. Each image is scored in turn and its card updates as soon as it's done, so results appear while the rest still run. Open any card for the full report, or export the whole batch as a single JSON file.
Can it detect AI-generated text inside an image, or tell me the prompt behind it?
No to both. The tool scores whether the pixels look generated. It doesn't read text in the image, doesn't judge whether wording was machine-written, and cannot recover the prompt. Prompt recovery isn't possible from pixels alone. When a file carries embedded generation parameters, the scan reports that they're present.
How is this different from an open-source detector on GitHub or Hugging Face?
The underlying pixel model here is an open research model, and that's stated rather than obscured. What's built around it is the difference: calibration onto published bands, the tiled region pass that catches partial edits, cryptographic credential verification, and a stated decision line. A raw checkpoint hands you an uncalibrated number.
Can it check videos as well as images?
No. Still images only. Video isn't supported in any form, including animated GIFs, where only the first frame is read. For a video you'd need to export a frame as an image and check that, which tells you about that frame alone.
Can it estimate when an image was created, or which model version made it?
No, not from the pixels. The model returns a probability that the frame is synthetic, with no date and no model version attached. When a file still carries content credentials or generation parameters, the scan reports what those claim, but that's the file describing itself, and it can be edited.
Privacy and access
Do you store the images I check?
No. Files you choose from your device are never uploaded, so there's nothing stored, nothing queued and nothing to delete. They're read and scored inside your browser tab and discarded when you close it. Your images are never used for training. The one exception is the URL box.
Check an image now
Drop in the picture you're unsure about and read the evidence yourself. This AI image detector is free to use, and nothing leaves your browser.
No account. No upload. Fifty checks a day.