Free · No account · Results never indexed
Was this made by AI?
Check an image, PDF or block of text. We read the file's own Content Credentials and metadata first — the only near-certain evidence there is — then fall back to detection models, and show you every signal behind the answer.
- of evidence
- 3 layersof evidence
- specialized tools
- 30+specialized tools
- sign-ups needed
- 0sign-ups needed
Drag a file here, or
JPEG, PNG, WebP, AVIF, PDF or .txt · up to 10 MB · you can also paste a screenshot
By uploading you agree to our Terms and Privacy Policy. Your file, IP address, approximate location and device details are stored.
No file to hand? See a real result:
Evidence, graded
Most detectors report one classifier's opinion as a percentage and call it a measurement. We grade evidence by how much it deserves to be trusted, and tell you which layer produced your answer.
- ANear-conclusive
What the file says about itself
C2PA Content Credentials, EXIF software tags, the IPTC synthetic-media marker, and the generation settings diffusion tools write straight into PNGs. When this is present it is near-conclusive — and we stop there without calling a model.
- BIndicative
Detection models
Two image classifiers and a text classifier, consulted only when the file carries no provenance. Useful, but this is the layer that degrades on generators it wasn't trained on.
- CWeak on its own
Statistical hints
Missing camera metadata, telltale output dimensions, uniform sentence lengths, assistant-typical phrasing. Each is weak on its own, so they nudge the estimate rather than decide it.
Where detection actually fails
Knowing the limits is what makes a result usable. These are ours, stated plainly.
A screenshot erases the best evidence
Provenance data lives in the file, not the pixels. Screenshot an AI image and the strongest signal is gone — which is why absence of evidence counts for very little here.
Classifiers age badly
They're trained on particular generators. Newer models produce artefacts the classifier has never seen, and accuracy drops sharply without any warning sign.
Text detection is biased
Detectors key on simple sentence structure and narrow vocabulary — which describes non-native English writing as much as generated text. False positives are not evenly distributed.
Editing defeats it
Paraphrasing or hand-editing generated text disrupts exactly the patterns detectors rely on. The people who edit most are the hardest to catch.
What you can check
Over 30 specialized detectors built for exact media and model fingerprints.
- AI ImagesPhotos, digital art & illustrations
- ChatGPT TextEssays, articles, emails & responses
- DeepSeek AIDeepSeek V3 & R1 reasoning checks
- Claude AIAnthropic Sonnet & Opus text models
- AI PlagiarismAcademic papers & student essays
- Turnitin ScorePre-submission academic checker
- PDF DocumentsExtracts & audits embedded text/meta
- MidjourneyDiscord & web app generation markers
- Flux ImagesBlack Forest Labs diffusion models
- DALL·E 3OpenAI provenance & metadata
- Bypass AIDetects humanized & spun AI content
- Content CredentialsC2PA cryptographic manifests
Looking for a specific model or file type?
Check out DeepSeek R1, Claude 3.5 Sonnet, Turnitin Score, AI Resumes, Flux, and more in our complete directory.
Common questions
How accurate is AI detection?
It depends entirely on what evidence the file carries. When an image contains Content Credentials or generation metadata, the answer is near-certain. When it doesn't, we fall back to detection models, which are typically 60–80% accurate and degrade on generators they weren't trained on. That is why every result shows a range rather than a single fabricated number.
Which AI models can this tool detect?
Our system detects output from all major frontier generators including ChatGPT (GPT-4o, o1, o3-mini), DeepSeek (V3 and reasoning R1), Anthropic Claude (3.5 Sonnet, Opus), Google Gemini, Midjourney (v5 & v6), DALL·E 3, Stable Diffusion (1.5, SDXL), Flux 1, and Sora video stills.
Can I use this to prove a student or employee cheated?
No. No AI detector is reliable enough to support an accusation on its own, and text detectors are documented to misclassify writing by non-native English speakers more often. Treat a result as a prompt to have a conversation, never as proof.
What are Content Credentials (C2PA)?
C2PA Content Credentials are a tamper-evident cryptographic record some tools attach to a file describing how it was made. OpenAI, Adobe Firefly, Google and several camera makers embed them. When present, they are far stronger evidence than any statistical classifier.
Why does the result say there isn't enough evidence?
Because that is often the honest answer. Screenshots, social-media re-encodes and edited photos strip the metadata that carries real signal, leaving only weak statistical hints. We would rather say so than invent a confident-looking number.
What happens to the files I upload?
They are stored in private cloud storage along with your IP address, approximate location and device details, which is described in the privacy policy and requires your consent before upload. Your file is also sent to Hugging Face for analysis. You can request deletion at any time through our 'Delete my data' page.
Is it free and do I need to create an account?
Yes. There are no accounts, no credit cards, and no payment walls. There is a generous limit of 100 checks per hour per visitor to keep the service free and available to everyone.