AI Watermark Detector: What Exists and What Doesn't
Search for an AI watermark detector and you will find a dozen tools. Almost none of them detect watermarks. They run classifiers and return a probability, which is a different thing with different failure modes.
- Only two real audio watermark detectors exist publicly, and neither works on Suno
- IRCAM Amplify, ACRCloud, Deezer and SubmitHub are classifiers that output probabilities
- A commercial detector false-flagged 4.7% of human-made tracks in peer-reviewed testing
- A pass on any public checker does not predict a distributor outcome
Most tools called an AI watermark detector are not detecting watermarks
If you search for an AI watermark detector, you get a list of products that appear to do the same job. They do not. They split into two groups that work on entirely different principles, and the group almost everyone lands on is the one that does not read watermarks at all.
A watermark detector decodes a signal that a generator deliberately embedded. It holds a key, it looks for a known payload, and it returns present or absent. False positives are essentially zero, because either the payload is there or it is not.
A classifier looks at your audio, compares its characteristics to patterns it learned from training data, and returns a probability. Nothing was embedded. It is reading incidental byproducts of how the model synthesized the sound, and it is guessing, in a rigorous statistical sense of the word.
Both get marketed with the same vocabulary. The difference determines whether the result means anything.
The two real watermark detectors, and why neither helps you
Google SynthID Detector. Real, documented, and rolling out to early testers since May 2025. It scans uploaded media for a SynthID watermark and, for audio, pinpoints the specific segments where the mark appears. Two limits: access is waitlisted and aimed at journalists, media professionals and researchers rather than the public, and it only finds Google's watermark. SynthID covers Lyria and NotebookLM audio. If your track came from Suno, SynthID Detector has nothing to find.
Meta AudioSeal. Open source, peer reviewed at ICML 2024, and available on GitHub. A generator and detector trained jointly on the Encodec architecture, with sample-level localization, so it can find a watermarked fragment spliced into a longer recording. It was built for speech and voice cloning rather than music, and it detects AudioSeal's own watermark. It will not find Suno's.
That is the complete list of public watermark detectors for audio. Both are keyed to their own creator's watermark. Neither is a general-purpose "is this AI" tool, and neither has any relationship to Suno or Udio.
There is no Suno watermark detector
Suno's v3 announcement in March 2024 said the company had developed proprietary, inaudible watermarking technology that can detect whether a song was created using Suno.
That single sentence is the entire public record. Since then: no published algorithm, no payload specification, no robustness data, no public detector, and no confirmation that the watermark is still present in current model versions. Suno's own terms of service do not mention watermarks.
So when a tool advertises Suno watermark detection, one of two things is true. Either it is running a statistical classifier and using "watermark" as marketing vocabulary, or it is making it up. There is no third option, because the information required to build a real Suno watermark detector has never been released.
Our full detector comparison covers how these tools behave in practice. This page is about what they are.
What the classifiers actually are
| Tool | What it really is | What it outputs |
|---|---|---|
| IRCAM Amplify | Transformer classifier over mel-spectrogram windows | Probability 0 to 1 |
| ACRCloud | Trained classification model | Probability, with an explicit disclaimer |
| Deezer | Patented classifier ("signatures") | Internal tag, now sold to third parties |
| SubmitHub AI Song Checker | Free consumer classifier | Probability |
| SynthID Detector | Real watermark decoder (Google only) | Present or absent, plus segments |
| AudioSeal | Real watermark decoder (AudioSeal only) | Present or absent, sample-level |
IRCAM Amplify runs a transformer audio classifier of the CLAP or AST family over mel-spectrogram windows, with a binary classification head and a confidence output. It was trained on human music against output from Suno, Udio, Stable Audio and MusicGen. It reads artifacts. We cover it in more depth in our IRCAM Amplify review.
ACRCloud publishes the most honest disclaimer in the industry. Its own documentation says detection results should be interpreted as model outputs based on statistical patterns in the audio, rather than as definitive proof of origin, and should not be the sole basis for legal or enforcement decisions without additional review. It claims 99.98% precision on internal test data, which is the important qualifier.
Deezer detects what it calls unique signatures distinguishing synthetic from authentic content, patented in December 2024. It tagged over 13.4 million AI tracks in 2025 and now sells the detector to third parties. By April 2026 it was seeing roughly 75,000 fully AI-generated tracks per day, more than 44% of daily uploads.
SubmitHub's checker is free and consumer-facing. Same category, different training data. Our SubmitHub checker review goes through what it catches.
Peer-reviewed testing showed a commercial detector could be defeated by resampling alone. Classifiers are brittle to signal processing in exactly the way watermarks are not. Undetectr processes the spectral characteristics these models read, which is the only mechanism with evidence behind it.
Try Undetectr → from $19 · $39 lifetimeThe peer-reviewed reality check
The TISMIR study "The AI Music Arms Race" is the most rigorous public evaluation available. The dataset: 30,000 tracks, 1,770 hours, combining 10,000 human tracks from the Million Song Dataset with 20,000 generated by Suno and Udio.
What it found:
- 0.988 F1 identifying AI tracks, 0.976 on non-AI. Strong performance.
- 4.7% of human-made tracks were mislabeled as AI. On a catalogue of any size, that is a lot of real musicians being wrongly accused.
- Resampling to 22.05 kHz defeated the detector.
That third finding is the one to remember, and it settles the argument about what these tools are. A watermark is engineered to survive resampling. SynthID explicitly claims robustness to speed changes, noise, and MP3 compression. If a simple resample breaks a detector, that detector is not reading a watermark. It is reading artifacts.
It also explains the gap between vendor numbers and lived experience. Reported accuracy comes from internal test sets on unprocessed audio. Real tracks have been through mastering, encoding, and format conversion, and performance degrades accordingly.
The false positive problem nobody advertises
The 4.7% figure deserves its own section, because it cuts against the interest of everyone selling detection.
A classifier that flags 4.7% of human music as AI is not a neutral referee. Applied to a distributor's upload queue, it means a meaningful number of entirely human recordings get held or rejected on a model's guess. There is no watermark to check against and no appeal to a ground truth, because no ground truth is in the file. The evidence is a probability.
This is why ACRCloud's disclaimer exists, and why it is worth taking seriously rather than reading past. The industry is aware its own tooling is not proof.
What a score does and does not tell you
Every classifier returns a number, and the number is routinely over-read.
What it means: the model's estimated probability that the audio was machine-generated, given its training data.
What it does not mean: that anything was measured in your file. There is no physical quantity being read. A 0.91 is not 91% of a watermark. It is a model's confidence, and confidence is a property of the model, not of your track.
Two practical consequences. Scores near the threshold are much less reliable than scores at the extremes, and the threshold is a vendor decision rather than a fact of nature. And detectors disagree with each other constantly, which is exactly what you would expect from independent models trained on different data.
A pass on a public checker does not predict a distributor outcome. No published correlation exists between any public detector's output and any distributor's internal decision. They are different models with different training data and different cutoffs. Treat a public pass as weak evidence at best.
What to do with this
If you are trying to release AI music, the useful conclusion is that you have been aiming at the wrong target. There is no watermark you can check, no detector that reads it, and no evidence any distributor decodes one. What exists is a set of classifiers reading the spectral character of your audio and returning a guess.
That guess is what stands between you and a release, and the peer-reviewed evidence says those guesses are fragile against signal processing. This is the mechanism Undetectr is built around, and it is the reason we recommend it over anything marketed as a watermark stripper. A watermark stripper targets something nobody has ever publicly demonstrated finding. Processing targets the thing that actually produces the flag.
None of which is about evading a ban, because none exists. Spotify, Deezer and DistroKid all permit AI music and ask for disclosure instead. The DistroKid AI screening guide covers what its process actually checks, and the Suno commercial use guide covers the rights you need before any of this matters.
Frequently asked questions
Two exist publicly. Google's SynthID Detector finds SynthID watermarks in audio from Lyria and NotebookLM, and it is waitlisted rather than open to the public. Meta's AudioSeal is open source and detects AudioSeal watermarks in speech. Neither works on Suno or Udio, because neither company's watermark is what those tools look for.
There is no public tool that does this. Suno stated in March 2024 that it embeds proprietary inaudible watermarking, but it has never published a detector, an algorithm, or a specification. Any tool claiming to detect the Suno watermark is either running a statistical classifier and calling it watermark detection, or it is guessing.
It is a classifier, not a watermark decoder. It runs a transformer audio model over mel-spectrogram windows and outputs a probability from 0 to 1 that a track is AI-generated. It was trained on human music versus output from Suno, Udio, Stable Audio and MusicGen. It reads generation artifacts, not an embedded signal.
In peer-reviewed TISMIR testing across 30,000 tracks, a commercial detector reached 0.988 F1 on AI tracks. It also mislabeled 4.7% of human-made tracks as AI, and resampling audio to 22.05 kHz was enough to fool it. Vendor-published accuracy figures come from internal test sets and degrade under real-world processing.
No. Every public checker uses different training data and different thresholds than any distributor's internal screening. There is no published correlation between a public detector's output and a distributor's decision. Treat a pass as weak evidence, not a guarantee.
It is the model's estimated probability that the audio is machine-generated, given the patterns it learned in training. It is not a measurement of anything physically present in the file. Scores near the threshold are substantially less reliable than scores at the extremes, and the threshold itself is a vendor choice.
Because they are different models trained on different data with different cutoffs. They are each estimating a probability rather than reading a fact. Disagreement is the expected behaviour of statistical classifiers, and it is a good reason not to treat any single score as authoritative.
Yes, and it is the only streaming platform that tags AI tracks in its interface. Its detector is patented and has been deployed since early 2025, tagging over 13.4 million AI tracks in 2025. Deezer does not ban AI music. It labels it and demonetizes the fraudulent streaming that surrounds it.
Ready to release your Suno tracks?
Undetectr was the only tool that passed every distributor in our testing. Clean your first track in under 60 seconds.