Tool 01 — Image
Is this image AI?The free AI image detector that reads evidence, not vibes.
A free AI image detector that reads what the file provably carries: C2PA Content Credentials, Stable Diffusion parameter chunks, ComfyUI workflows, XMP generator markers and EXIF — the trails left by Midjourney, DALL·E, Firefly, Imagen and 50+ other tools. Provenance over pixel guesswork.
Aipurity is a free AI image detector that shows you the file’s actual evidence — C2PA manifests, generator metadata, EXIF history — instead of a black-box percentage.
- Runs 100% in your browser
- Nothing is uploaded — works offline once loaded
- Every verdict lists its evidence
Drop a file, or click to choose
Analysed locally in your browser. Nothing is uploaded, ever.
What this AI image detector actually reads.
Content Credentials (C2PA)
The cryptographic provenance standard from Adobe, OpenAI, Google and camera makers. From August 2026 the EU AI Act requires AI outputs to carry machine-readable marks — we verify the signed manifest and its edit history.
Stable Diffusion & ComfyUI chunks
Local generators write their full recipe into PNG text chunks: prompt, sampler, seed, model hash. A1111 WebUI, ComfyUI, SDXL, InvokeAI, NovelAI — if it’s there, the image testified against itself.
XMP & IPTC generator markers
Midjourney, DALL·E, Firefly, Recraft, Leonardo and others stamp XMP metadata, including the IPTC digitalSourceType “trainedAlgorithmicMedia” — the industry’s official “made by AI” label.
EXIF capture trail
Real cameras write lens, exposure and timestamp data. Its presence weakly supports a real capture; its absence means a stripped, edited or generated file.
How to check if an image is AI-generated.
- 01
Your file stays on your machine
The check runs entirely in your browser. No upload, no server — no one harvests your photos.
- 02
We parse structure, not pixels
PNG chunks, JPEG segments, XMP packets and C2PA manifests are read byte by byte, matched against 50+ known generator fingerprints.
- 03
You get evidence, not a vibe
Every verdict lists the exact signals behind it. “Detected” means the file names its maker; “inconclusive” means no honest tool could say more.
Which AI generators can it recognize?
Different generators leave different evidence. This is exactly what we read — and where every detector, including this one, goes blind.
| Evidence signal | Typically written by | What you’ll see |
|---|---|---|
| C2PA Content Credentials | Adobe Firefly & Photoshop, OpenAI DALL·E & Sora exports, Google Gemini / Imagen, Leica and Sony cameras | A cryptographically signed manifest naming the maker and edit history |
| PNG parameter chunks | Stable Diffusion WebUI (A1111), ComfyUI, SDXL, InvokeAI, NovelAI | The full generation recipe: prompt, model, sampler, seed |
| XMP / IPTC markers | DALL·E, Firefly, Recraft, Leonardo and tools that stamp digitalSourceType “trainedAlgorithmicMedia” | The industry’s machine-readable “made by AI” label |
| No reliable marker | Midjourney web downloads, screenshots, most social-media re-shares | An honest “inconclusive” — the file carries no proof either way |
We’d rather tell you Midjourney usually can’t be proven from the file alone than invent a confident-looking score.
Can you still spot an AI image by eye?
Sometimes. Before any AI-generated image checker existed, these visual tells were the only way — still worth checking, just know they fade with every model generation.
Hands, teeth and text
Extra fingers, fused teeth and warped lettering were reliable giveaways in 2023–24. Current models get them right most of the time.
Physics and reflections
Mirrors, shadows and liquid that don’t agree with the scene still trip generators — worth a close look.
Texture that’s too clean
Waxy skin and hyper-smooth gradients suggest synthesis, but heavy beauty filters produce the same look on real photos.
The context test
Reverse-search the image, find the earliest posting, and ask who benefits from you believing it. Context outlives pixels.
Peer-reviewed benchmarks keep finding that pixel classifiers collapse on fresh generators and in-the-wild images — which is why this tool reads evidence the file actually carries instead.
When a check like this matters
Roughly 34 million AI images enter the world every day by published estimates — sooner or later one lands in your inbox, your marketplace chat, or your news feed.
Marketplace & rental listings
Too-perfect product shots and apartment photos are cheap to generate. Provenance metadata from a real camera is the counter-evidence a buyer can actually check.
Dating & social profiles
AI profile pictures carry no capture trail. A stripped, metadata-free portrait isn’t proof of fakery — but a C2PA-signed camera photo is strong evidence of a real one.
Newsrooms & fact-checkers
Before amplifying citizen footage, check what the file admits: generator markers end the question, and a clean-but-stripped file tells you to trace the original source.
Insurance & disputes
Claim photos with intact EXIF and edit history read very differently from files that were run through an editor and exported clean.
Stock platforms & contests
Submissions billed as human-made can be spot-checked for Stable Diffusion parameter chunks and trainedAlgorithmicMedia labels in seconds.
Everyday “is this real?” moments
A viral photo in the family group chat deserves ten seconds of evidence-reading before it gets forwarded again.
Generator roster: what each one leaves behind
Public export behaviour of the major generators — and the verdict you should expect from an evidence-based check.
| Generator | Typical export evidence | Expected verdict |
|---|---|---|
| Adobe Firefly / Photoshop AI | C2PA Content Credentials on standard exports | Detected, with signed manifest shown |
| OpenAI DALL·E / ChatGPT images | C2PA manifest on downloads | Detected while the manifest survives |
| Google Imagen / Gemini images | SynthID watermark; C2PA on some paths | Detected when C2PA present; SynthID needs Google’s verifier |
| Stable Diffusion (A1111 / ComfyUI / SDXL) | Full parameter recipe in PNG text chunks | Detected, prompt and model shown |
| InvokeAI / NovelAI | Generation metadata in PNG chunks | Detected when chunks intact |
| Recraft / Leonardo | XMP generator markers on many exports | Detected when markers survive |
| Midjourney (web / Discord download) | Usually no reliable machine-readable marker | Inconclusive — we say so instead of guessing |
| Flux (hosted UIs vary) | Depends on the UI: local runs write chunks, hosted exports often clean | Detected or inconclusive by export path |
| Any generator, after a screenshot or social re-upload | Metadata stripped by the platform | Inconclusive — trace the earliest file |
What a verdict looks like
“Detected” — the file named its maker
A C2PA manifest, a Stable Diffusion parameter chunk, or a trainedAlgorithmicMedia label was found. Every signal is listed, so you can verify the reasoning rather than trust a number.
“Inconclusive” — no proof either way
The honest verdict for most screenshots and social-media downloads. It means the evidence was stripped, not that the image is human-made.
Capture-side signals
Intact EXIF (lens, exposure, timestamp) or a camera-signed C2PA manifest weakly-to-strongly supports a real capture — shown with the same caveats we apply to everything else.
The EU AI Act makes provenance the default
From 2 August 2026, Article 50 of the EU AI Act requires providers of generative AI systems to mark their outputs in a machine-readable way, and deployers of deepfakes to disclose them — with fines that make ignoring it expensive. C2PA Content Credentials, backed by Adobe, OpenAI, Google, Microsoft and major camera makers, is the leading implementation, which means the share of AI images carrying verifiable provenance is about to climb sharply. This check verifies those marks today, so “where did this image come from?” gets an evidence-based answer, not a guess — and as marking becomes law rather than courtesy, evidence-reading beats pixel-guessing by a wider margin every month.
Provenance, in plain words
The six terms this page keeps using — defined once, honestly:
C2PA / Content Credentials
A cryptographically signed manifest inside the file recording who made it, with what tool, and how it was edited. Signed means verifiable — a stripped or altered manifest fails validation rather than lying.
EXIF
The metadata block real cameras write: lens, exposure, timestamp, sometimes GPS. Easy to strip, hard to fake coherently.
XMP & IPTC
Extensible metadata used by editing software and AI generators alike — including the IPTC digitalSourceType field, whose value “trainedAlgorithmicMedia” is the industry’s official machine-readable “made by AI” label.
PNG text chunk
A section of a PNG file where local generators like Stable Diffusion WebUI and ComfyUI store the full generation recipe — prompt, model, sampler, seed.
SynthID
Google’s invisible watermark for AI output. Robust to some edits, but only Google’s own tools can verify it — we report what the open standards in the file show.
“Inconclusive”
Our verdict when a file carries no provable evidence either way. Every honest detector has this state; most just hide it behind a confident-looking percentage.
Honest limits
What it can’t tell you.
A screenshot, a re-save, or a pass through WhatsApp/WeChat strips metadata — a stripped AI image returns “inconclusive”, not “human”. Independent 2024–26 benchmarks show pixel-only detectors collapse to near coin-flip on in-the-wild images, so we refuse to fake a confident number.
Common questions.
Why did my obviously-AI image come back “inconclusive”?+
Its provenance was stripped — usually by a screenshot, a platform re-encode, or an export. “Inconclusive” means the file carries no proof either way; treat the source and context as your evidence.
Does it work if the image has no metadata — a screenshot or a social-media download?+
It will tell you, honestly, that the evidence is gone: screenshots and most platform downloads carry no provenance. That blind spot applies to every detector — most just hide it behind a made-up percentage. Trace the earliest version of the image you can find and check that file instead.
Which AI generators can this detect?+
Anything that writes evidence into the file: C2PA signers (Firefly, DALL·E, Gemini/Imagen exports), Stable Diffusion family tools (A1111, ComfyUI, SDXL, InvokeAI, NovelAI), and 50+ XMP/IPTC generator fingerprints including Recraft and Leonardo. Midjourney web downloads usually carry no reliable marker, so they come back “inconclusive” — we say so rather than guess.
How can I check if an image is AI-generated for free?+
Drop it into the free AI image checker above — it runs in your browser. It reads C2PA Content Credentials, Stable Diffusion parameter chunks and 50+ generator fingerprints to detect AI-generated images by the evidence they carry. If the file was stripped (screenshot, social re-encode), you get an honest “inconclusive” instead of a guess.
Can a detector just look at the pixels?+
Many claim to. Independent benchmarks show real-world accuracy collapses far below lab numbers and query-based attacks evade public detectors — the same applies to deepfake-image classifiers. Pixel guessing is an arms race the detector side is losing.
What is C2PA and why does it matter now?+
C2PA (Content Credentials) is a cryptographically signed manifest of who made a file and how. The EU AI Act transparency rules, applicable from 2 August 2026, require AI media to be machine-readably marked.
Can it detect Midjourney v7 images?+
Usually not from the file alone — Midjourney’s standard web and Discord downloads carry no reliable machine-readable maker label, so an evidence-based check returns “inconclusive”. Tools that claim confident Midjourney percentages are pixel-guessing; independent benchmarks show that collapses in the wild. Check the posting context and the earliest source instead.
What image formats work best?+
JPEG and PNG carry the richest evidence — C2PA manifests, XMP/IPTC markers, EXIF, and PNG parameter chunks. Anything your browser opens can be checked, but format conversions and screenshots typically strip provenance, so expect “inconclusive” on re-encoded copies.
Is there an API or batch mode?+
The check is browser-first by design — your files never touch a server. Batch checks and API access are part of the Pro plan for people who need volume.
Does it work on a phone?+
Yes — the check runs in mobile Safari and Chrome the same way it does on desktop, including the C2PA verification (the toolkit loads on demand). Share a photo from your camera roll straight into the page; it still never leaves the device.
Can schools or employers use this to check submissions?+
As evidence, yes — a Stable Diffusion recipe or a C2PA manifest in a submitted file is a fact you can point to. But treat the absence of evidence as exactly that: absence. A stripped file is “inconclusive”, not “guilty”, and no tool — including this one — can fairly punish someone over a missing metadata block.
Do you store my images?+
No. The analysis is plain JavaScript running in your tab. Disconnect from the internet after the page loads and it still works.
Sources & further reading
- RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Image Detectors (arXiv, 2025)
- The Deployment Gap in AI Media Detection (arXiv, 2026)
- C2PA — Coalition for Content Provenance and Authenticity
- EU AI Act, Article 50 — transparency obligations (applicable 2 Aug 2026)
- Everypixel Journal — AI image generation volume statistics