Calabi Labs · Guide · 2026-06-15
Want to know if an image is an ai deepfake or not use this openai tool
Here's the page — clean HTML, ~850 words, structured to match the user's search intent while honestly explaining both what OpenAI's tool does and where Calabi fits.
```html
Does OpenAI's Verify Tool Actually Tell You If an Image Is AI-Generated?
What OpenAI's Verify Tool Actually Checks
When you upload an image to Verify at openai.com/research/verify/, it looks for two specific signals OpenAI embeds in content it generates:
- C2PA metadata — a structured data block stored in the image file that carries cryptographic provenance information. For OpenAI exports, this includes JUMBF atoms with the OpenAI generator signature, Content Credentials, and fields like
DigitalSourceType: trainedAlgorithmicMediain the XMP namespace. Verify checks whether those atoms are present and intact.
Verify returns one of three verdicts: the image shows signs of OpenAI origin, it shows signs of non-OpenAI AI generation, or no AI generation signals are detected. The tool is transparent about what it's checking — it publishes its methodology and limits openly.
Why OpenAI Verify Can't Detect Most AI Images
The core limitation is scope: Verify is built to serve OpenAI's content provenance initiative, not to be a universal deepfake detector. It only reliably flags images generated by DALL·E and tools within OpenAI's ecosystem. Here's where it falls short for the "is this image AI?" question most people are asking:
- It only finds OpenAI signals. A Midjourney export, a Runway Gen-3 video, or a Flux Schnell image will almost certainly come back as "no AI signals detected" — even though those are 100% AI-generated.
- Metadata is fragile. Upload an AI image to Instagram, download it, screenshot it, or transcode it through almost any tool, and the C2PA metadata is gone. Verify relies on that layer surviving.
- It doesn't detect deepfakes. Verify checks provenance metadata — it has no opinion on whether a photo depicts a real person or a合成假的 face. That requires an entirely different class of perceptual analysis.
So if someone sends you a suspicious image and you want to know "is this AI?" — Verify is a narrow, OpenAI-specific tool. It answers "did DALL·E make this?" not "is this image synthetic?"
The Honest Alternative: Checking What Platforms Actually Scan
If you're an AI creator trying to post on Instagram, TikTok, YouTube, or Reddit without getting flagged, the question isn't really "is this AI?" — it's "does this file look like AI to the platform's automated scanners?" Those systems don't use OpenAI's provenance tool. They scan for:
- C2PA / Content Credentials — the cryptographic manifest that says "this was AI-generated." Platforms like YouTube and Reddit check for this automatically.
- XMP AI flags — fields like
DigitalSourceType: trainedAlgorithmicMediain image metadata that explicitly label the content. - Encoder fingerprints — Lavc (libavcodec), x264 SEI messages, and similar codec signatures in video files that identify them as machine-encoded rather than phone-captured.
- Missing capture signals — no GPS coordinates, no real device Make/Model, no capture timestamp. Authentic phone photos carry these. AI exports don't.
This is exactly what Calabi handles — the inverse of detection. Instead of scanning for AI signals, Calabi removes them. Upload your AI-generated image or video, and Calabi strips the C2PA atoms, XMP AI flags, and encoder fingerprints in one pass, then injects authentic phone-capture identity (real device profiles: iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra — including GPS, timestamp, and real-phone encoder names). You get a forensic proof card before download showing exactly what was stripped and what was injected.
When to Use Which Tool
| Scenario | OpenAI Verify | Calabi |
|---|---|---|
| Checking if a friend/colleague made something in DALL·E | Works if C2PA metadata survived | Not the right tool |
| Posting AI video to TikTok/Instagram without auto-flagging | Doesn't help — only checks OpenAI content | Strips the signals platforms scan for |
| Verifying a screenshot or re-uploaded file | Metadata likely gone, invisible watermarks likely degraded | Handles the file-level signals regardless of history |
| Detecting a deepfake face in a photo | Not designed for this | Not designed for this — Calabi is for provenance, not facial analysis |
Frequently Asked Questions
Does Calabi work on video as well as images?
Yes. Calabi processes both images and video, stripping encoder fingerprints like Lavc and x264 SEI messages that are common in AI-generated video files and that platforms use to flag synthetic content.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```