The honest hub
AI image detector — provenance edition
Most “AI image detectors” look at pixels and guess. That approach produces false accusations. The better question is: what did the model leave in the file? Here's the state of every major vendor's marking, and a reader for the parts that are publicly verifiable.
Runs entirely in your browser — your file is never uploaded.
Drop an image here, or browse files
JPEG · PNG · WebP — up to 25 MB. Files are read locally and never uploaded.
Quick answer
As of September 2026, provenance-based “detection” beats pixel-guessing wherever vendors attach signed marks: Claude (files), OpenAI (images), Google (images), Adobe (Firefly) — all C2PA-territory. Midjourney and most open-source pipelines still leave nothing. EU AI Act transparency obligations are pushing every provider toward embedded marks, so this map keeps moving — dated, sourced, and honest.Marking status by model family
| Model family | Mark on images | Mark on text | Third-party verification |
|---|---|---|---|
| Anthropic Claude | Signed C2PA credential on .png/.jpg/.svg outputs | Invisible SynthID-Text-derived watermark | Files: yes, anywhere · Text: Anthropic key only (API in private preview) |
| OpenAI GPT Image / DALL·E | C2PA metadata in generations | — | Images: yes (C2PA) |
| Google Gemini / Imagen | SynthID (pixels) + C2PA metadata | SynthID-Text | C2PA: yes · SynthID: Google's gates only |
| Adobe Firefly | Content Credentials | — | Yes (C2PA) |
| Midjourney | None by default | — | No provenance layer exists to check |
| Open-source pipelines (SD/Flux/ComfyUI) | Only if the pipeline author adds it | — | Incorrect often; check per-file |
Why we don't offer a pixel-guessing “AI detector”
Statistical pixel classifiers answer a different question (“looks AI-ish”) and are notorious for flagging human photos. Provenance answers questions with receipts: “this signed record says who made it, when, and what was done after.” Where receipts are absent we say so plainly instead of guessing — that's the product decision. For the raw-data view, use the C2PA reader, and for text work the hidden-character scanner.
Sources
- Anthropic — How Claude's text watermark works (2026) — accessed 2026-09-29
- OpenAI — C2PA provenance announcement (Feb 2024) — accessed 2026-09-29
- Google DeepMind — SynthID — accessed 2026-09-29
- C2PA — Technical specification — accessed 2026-09-29