AI Detector for Images, Video & Documents

What does the AI detector look at?

Two different things, shown separately. File evidence is data written into the file by whatever made it — a camera, an editing app or an AI service — and it can name the exact tool. The visual check and the text check judge the content itself, with classifiers trained to tell generated images from real ones and AI writing from human writing. File evidence is specific but often missing; the checks don’t need it, but can only estimate.

How is the verdict decided?

The headline leads with the strongest evidence found: a signed statement from the tool that made the file outranks an unsigned label, and both outrank a reading of the pixels or the words. Each label tells you what it rests on.

Why doesn’t missing file evidence mean the file is real?

OpenAI (ChatGPT, DALL·E and Sora), Adobe Firefly, Microsoft, Google and others attach Content Credentials to many outputs. But platforms and messaging apps often re-save uploads without them, screenshots never carry them, and anyone with a metadata tool can strip them. An empty panel means the file is silent about its origin, nothing more.

Images, video and documents: what changes?

• Images get the file-evidence and visual checks in full; the AI image detector goes deeper into reading the score.
• Videos are checked for Content Credentials, which Sora, Veo and Adobe Firefly video can carry, plus the visual check on 8 sampled frames — labeled experimental, since the model was trained on stills; see the AI video detector.
• PDFs are checked for Content Credentials, XMP and creator and producer tags; their text gets the text check and their embedded photos the visual check. A scanned PDF has no text layer, so only its file evidence and photos are checked.
• Word files and plain text, as .docx, .txt or .md files or pasted in, get the text check, and a Word file’s creator app is read as file evidence. Under 80 words, the verdict is "Too short to judge"; the AI text detector explains the text verdicts and their limits.

How accurate are the checks?

The visual classifier is a fine-tuned version of Community Forensics, a detector from a CVPR 2025 paper. On its developer’s own development data it reached 90–96% balanced accuracy, flagged about 3–5% of real photos as AI, and caught roughly 66–93% of web-compressed images from recent generators — the developer’s figures, not an independent benchmark. It scores the center and four corners and reports the lowest, so an image is called AI only when every region looks generated: fewer false alarms on real photos, more misses on partial edits.

The text classifier is a RoBERTa-base model fine-tuned on RAID, an AI-text benchmark from ACL 2024. On RAID’s leaderboard its developer reports an AUROC of 99.28% and 90.17% of AI text caught at a 1% false-positive rate; real-world accuracy is lower. It is English-focused, and human writing that is short, formal or AI-polished, or by a non-native English speaker, is likelier to be flagged — so Toolvy refuses texts under 80 words and counts a passage only at a 97% score or higher.

Is my file uploaded?

No. The metadata reader and both classifiers run in your browser tab, and DevTools → Network shows no request carrying your file or your text. Each model downloads once, on first use, and is then cached: the image model, a ViT-Small of about 44 MB, and the text model, about 126 MB from Hugging Face’s CDN, both under the MIT license.

Related tools

See every field a file carries in the metadata viewer, or learn what a manifest records and what its signature proves in what C2PA is.