PARAS MADAN
All writing

Latest in AI

ElevenLabs Watermark API: Check Audio Provenance Without Overclaiming

ElevenLabs can now check uploaded audio for its watermark. Learn what a positive result means, what a negative result cannot prove and how to build a review queue.

ElevenLabs added a beta Content Provenance API in its November 17 update. It checks an uploaded audio file for an ElevenLabs watermark. That is useful for a newsroom, moderation team or creative operation trying to understand an asset's origin. It is not a general-purpose answer to whether a recording is fake.

The distinction belongs in the product interface before the first file is checked. A result labelled “human audio” would promise something this endpoint does not establish.

Read detection and metadata separately

The official provenance guide documents POST /v1/content-provenance/watermark. Its response separates is_detected from details. A detected watermark may lack recoverable generation details, particularly for short or heavily changed audio. Available details can include product, model, generation time and a generation identifier, with nullable fields. The endpoint does not disclose the prompt or the account behind the audio.

I would present three outcomes in a review tool: watermark detected with details, watermark detected without details, and no watermark detected. Keep request failures separate from all three. An upload error is not a negative finding.

That gives a reviewer a useful statement about the check without turning missing metadata into a claim about the speaker. A positive result alone does not establish consent, ownership or whether the surrounding story is accurate. A negative result does not establish that the recording is human-made.

Build a small intake queue, not a universal scanner

The documented limits are one audio file per request, up to 50 MB, lasting between 0.5 seconds and five minutes. Each API key user gets 25 checks per 24 hours. Authentication requires an API key with the watermark permission; OAuth tokens are not accepted.

For an initial workflow, I would queue only assets that need an origin check for a specific decision. Record why each file was submitted and who will review it. Keep the original file intact, then attach the response and check time to that record.

Check size, duration and file type before submission so avoidable errors do not become a manual cleanup task. Do not quietly shorten or re-encode a file and present the result as though the original was checked. If you need to inspect an excerpt, record its boundaries and explain that limitation to the reviewer.

Before uploading, make sure your team has permission to send that recording to ElevenLabs. Do not use sensitive customer calls as convenient test fixtures merely because they are already available in your storage.

Keep failures visible

The guide distinguishes rate-limit responses from invalid requests: 429 can be retried later, while 400 or 422 requires a request change. Build a queue state for each rather than endlessly retrying every error.

My first test set would use permitted files whose origin the team already knows. Check how the interface handles absent details, a rejected upload and an exhausted quota. The point is to validate your handling of the response, not to manufacture an accuracy claim from a few examples.

I would also give reviewers a way to record evidence outside the API, such as the source of the file and the creator's explanation. Keep those notes distinct from the machine result so someone reading the record later can tell what was measured and what was inferred.

Make the conclusion narrow enough to trust

The useful output is a provenance signal attached to a file, with its limitations visible. It should help a person decide what to investigate next, not automatically accuse a creator or clear a recording as authentic.

That restraint makes the integration more useful. Reviewers get a concrete result they can act on without inheriting a false promise from the label on the button.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain new capabilities through practical technical writing and clear product stories.

Book a 30-minute call.