What can official AI verification tools prove, and what do they miss?

Vendor tools are useful for checking their own watermarks and provenance signals, but coverage, sign-in requirements, and watermark strength differ. Understand what each one confirms before handing an original file to another provider.

Tools covered
OpenAI · Google · Microsoft · C2PA
Last updated

Decide which vendor, if any, can answer your question.

If you need to know whether a file carries a specific vendor signal, go to that vendor. If you do not yet know where the signal came from, start with GenDetect for cross-vendor triage.

Keep the original file before checking

Official tools depend on metadata or watermarks, and screenshots, transcodes, and recompression can weaken those signals. Save the original file and its source link before uploading. A platform copy is not the best evidence.

A positive result only shows a signal

A C2PA manifest or watermark shows that a supported signal is present. It does not prove the content is accurate, unedited, or presented in the right context. Likewise, a missing signal is not proof of human authorship.

Four entry points, with different coverage.

Each tool is operated by its vendor, and its terms, sign-in requirements, and file limits apply. External links open in a new tab so the current GenDetect result stays available.

OpenAI Verify

Checks images and audio generated or exported by OpenAI tools for OpenAI C2PA manifests and SynthID watermarks. It does not detect other vendors, and it may not cover older or heavily processed OpenAI content.

Open OpenAI Verify

Google SynthID Detector

Checks images, video, and audio for SynthID watermarks from Google and partners including OpenAI, NVIDIA, and Kakao, with Apple support still rolling out. Sign-in with Google, Apple, or ChatGPT is required, and files are limited to about 100MB.

Open SynthID Detector

Microsoft Content Provenance Detection

Checks Microsoft-issued C2PA manifests and watermarks in images, audio, and video up to 100MB. It covers Microsoft AI systems only and cannot identify content generated by another vendor.

Open the Microsoft check

Content Credentials Verify

Verifies the signature chain, signer, and edit history of cross-vendor C2PA Content Credentials. It is not an AI detector, does not read SynthID or other watermarks, and cannot prove human authorship just because credentials are absent.

Open Content Credentials Verify

Official tools still cannot provide a complete answer.

Use these entry points as supporting evidence, not as a final verdict. The limits below apply to most vendor-run detectors.

They recognize only their own or partner signals

OpenAI and Microsoft cover their own systems, Google covers Google and selected partners, and Content Credentials verifies C2PA only. A negative result means that no supported signal was found, not that the content is human-made.

Watermarks can be weakened by editing

Cropping, compression, screenshots, and transcoding can reduce watermark detectability, and vendors acknowledge that heavy cropping or legacy content may be missed. Meta Content Seal was reported to miss cropped images and covers only its own model, so it is not listed as a fixed entry point here.

Uploads go to third parties under different terms

Official checks require uploading the file to the vendor, and some tools require sign-in. Retention, purpose, and cross-border processing are governed by that vendor’s privacy terms. Review them for sensitive material, or use only the necessary cropped portion.

GenDetect adds cross-vendor triage.

GenDetect does not replace any official entry point. It gathers source clues on one page so you can decide which vendor check, if any, comes next.