A detector's 99% accuracy claim is the vendor's own measurement

Ircam Amplify's AIMD product page states an accuracy figure with no published method or independent audit.

Site publication

September 18, 2026

What happened

Ircam Amplify, a company the site states was founded in 2019 to translate research from the Paris institute IRCAM into commercial audio tools, offers a product it calls AIMD, an AI Music Detector. The about page describes the company's mission as providing 'transparency', meaning 'reliable, science-based information about what a given audio file contains', ahead of a second goal of attribution. The product page, as retrieved on 16 September 2026, markets AIMD to distributors, labels, collecting societies and publishers as a way to flag tracks generated by named systems including Suno, Udio, Sonauto and ElevenLabs.

What the documents say

The product page states AIMD achieves '99% accuracy' with 'less than 1% false positives', delivered as a RESTful API kept 'always up-to-date with the latest AI models'. It does not publish the detection method, the evaluation set used to produce that figure, or how the claim behaves against generators released after the model was last updated. Neither page discloses an independent audit of the accuracy claim, and no third-party benchmark for AIMD was located among the sources opened for this record. A vendor-reported accuracy figure describes performance the vendor measured on data the vendor chose.

Why it matters for makers

Detector accuracy is a property of a specific test set against specific generators at a specific moment, not a permanent fact about a product. A high headline figure from the company selling the detector says nothing about how it performs on a generator released after the claim was measured, on adversarial post-processing such as re-recording or pitch-shifting, or on genres under-represented in whatever set produced the number. A distributor or collecting society relying on a single vendor's detector to gate payouts is making a policy choice about acceptable error, not applying a settled scientific measurement.

What to check before you use it

Ask any platform using AIMD, or a comparable detector, what happens on a flagged false positive, since the vendor's own 'less than 1%' figure implies genuine human recordings will occasionally be caught. This is an editorial reading beyond the source documents: treat a single vendor's accuracy claim as marketing until an independent, published test against current generators is available, and check whether your distributor discloses which detector, if any, it runs before release.

• What test set and generators produced the vendor's accuracy figure, and has it been reproduced independently?
• What happens to your track, and to your payout, if a detector flags a false positive?
• Does the detector's stated scope cover the specific generator you used, or only the ones named on the vendor's page?

A detection product answering to its own maker's benchmark is not the same as an audited standard. Until independent testing exists, a stated accuracy figure should be read as a claim to verify, not a settled measurement to build policy on.

Sources & reading trail

States the AIMD product's claimed 99% accuracy and under-1% false-positive figure and named target generators.

Source published: Not established · Retrieved: 16 September 2026

States the company's 2019 founding and its stated mission of transparency and attribution.

Source published: Not established · Retrieved: 16 September 2026

Papers, terms and official documents establish the record; the maker reading and the checks are Signal to Song editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• DeepMind watermarked its own generated music, not music generally
• A broadcast standard specifies how to blind-test audio
• A provenance standard records edits, not the truth of a file
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.