Skip to content
FREN
Menu

Detectors

Detectors and tools

How detectors distinguish AI-generated tracks from recordings made by people, who uses them, and what it would take to check the figures their makers report. The page also describes small services that offer to attest to music's human provenance, without endorsing them.

How AI music detection works

An AI music detector is a piece of software trained on two sets of tracks. One contains tracks known to have been produced by generators such as Suno or Udio; the other contains tracks known to have been recorded by people. During training, it compares these tracks and learns to recognise patterns that generators leave in the sound, usually too subtle to hear. When given a new track, it calculates the probability that it belongs to the first group. Some tools work differently, searching the file for a signature deliberately embedded by the generator, such as a digital watermark or Content Credentials. None of these tools assesses the talent, emotion or intent behind a piece of music. Their work is limited to comparing statistics, which is why a detector that performs well with today's generators may need to be retrained when the next generation arrives.

Who detects what

Tool Operator Use and users Reported accuracy Independent audit? Public access
Deezer AI detector Deezer (in-house tool) Tags tracks in Deezer's catalogue; licensed to SACEM (01/2026) and EJI (03/2026) 99.8% accuracy; misses about 2 AI-generated tracks in 1,000; flags fewer than 1 track recorded by people in 10,000 (Source: Deezer, 21/07/2026) None known Free playlist scanner since 11/06/2026, covering 20 platforms and 27 languages
AI Music Detector (AIMD) Ircam Amplify Marketed to streaming platforms (DSPs), distributors, record labels and collecting societies 99% accuracy, under 1% false positives (vendor's figures, no independent audit published; no public test set) None known API and SDK on request only; no self-service access or published pricing
Believe / TuneCore generative AI detection Believe (source of the technology not specified) Screens tracks submitted for distribution through Believe and TuneCore; blocks tracks from unlicensed generators, a practice already in place when announced on 30/04/2026 "99% reliable" at identifying the model that generated a track (reported by Believe) None known None; no documented appeal procedure
SH Labs and SoundPatrol Third-party providers Used for Traxsource's "Human-Made" / "AI-Assisted" labels since 01/07/2026; SH Labs is also reportedly used by Bandcamp and GEMA Not published None known No known public access
Apple Music in-house detection Apple Supports AI Transparency Tags, which are still based on declarations by content providers Not published None known None
TIDAL detection TIDAL, alongside declarations by distributors Fully AI-generated tracks excluded from royalty payments since 29/06/2026 and labelled as AI since mid-July 2026 Not published None known None; appeals through the TIDAL support form
Suno Content Credentials check Suno Anyone who wants to check whether a file carries Suno Content Credentials Not applicable (reads a signature rather than classifying the track) Not applicable Free check by uploading a file or submitting a link

The figures in this table come from the operators themselves. We have not found an independent audit of any of these tools. Sources are listed at the bottom of the page.

When a figure can be checked

The manifesto's second demand is for detectors open to independent scrutiny, so that their detection rates can be checked by people other than the companies behind them. For that scrutiny to be possible, several conditions must be met, and none of the tools described above currently meets them all.

  1. A published method must explain what the model analyses, how it was trained and how "AI-generated" is defined in its training data, whether that category covers only fully generated tracks or also AI-assisted stems. Ircam Amplify and Deezer describe their approaches in general terms, but neither publishes a reproducible protocol.
  2. Error rates must also be published and measured on a publicly available test set. Both types of error matter, whether a detector misses an AI-generated track or flags one recorded by people. Deezer publishes these rates, but not the set of tracks used to measure them, so others cannot reproduce the reported result.
  3. The test must be run by a third party independent of the vendor, using tracks the vendor did not choose. To our knowledge, no detector used by a platform or distributor has yet undergone such an audit.
  4. Testing must also cover less widely represented musical genres and traditions, as well as small-budget productions. A detection model learns only from the examples it is given, and a rate measured mainly on mainstream pop tells us little about a Gnawa recording, a folk demo recorded at home or a heavily processed club track. Testers of Ircam Amplify's tool report that it makes the most errors on densely produced electronic music, an observation that has not itself been audited (AI Musicpreneur, 10/08/2026). Only results published by genre, language and production budget would allow this question to be settled.
  5. Performance must also be assessed over time. Generators change every few months, and a result from June says nothing about December unless the test is repeated. A paper on detector robustness accepted at the ISMIR 2026 conference acknowledges that the model it proposes, trained on three generators, does not generalise to others and must be retrained to cover them (arXiv 2607.27454, 07/2026). Detecting AI-generated stems within hybrid works, meanwhile, remains an open research question, on which another paper presents what its authors believe to be the first study (arXiv 2607.26874, 07/2026). The first paper is co-authored by four authors from the École Centrale de Lyon and two researchers from Deezer Research, the second by four Deezer Research scientists, the research team of a platform that runs and licenses its own detector. Both shed light on the state of research, but neither can stand in for the independent evaluation described in point 3.

A sixth condition concerns how decisions are handled. When someone challenges a decision based on a detector's output, they must be told why it was made and have a person they can contact. Yet there is currently no standardised appeal procedure. Artists are advised to contact their distributor and provide their digital audio workstation (DAW) sessions, stems and dated drafts, although the final decision rests with the platform (iMusician, 11/08/2026).

Services that offer to attest to music's human provenance

Alongside the platforms, a handful of small companies offer to attest that a track was made by humans. The accounts below reflect what each says about itself. We do not endorse any of these services and have not verified their methods.

AudioIntell (audiointell.ai) is a division of Electropoint LLC. Artists submit a track title, artist name, ISRC (International Standard Recording Code) and an MP3 file, which "proprietary detection systems" analyse before a manual review. For US$2 per track, they receive a "Human-Made Music Certification" (HMCC) number, a badge and a listing in a searchable public database. The terms of service state that the service "does not guarantee the absolute originality or ownership of the track" and disclaim liability for errors arising from "undetectable AI influences". Neither the detection method nor the error rates are published (audiointell.ai, accessed on 02/10/2026).

Humanable (humanable.com) was founded in Nashville in 2024 by Lili McGrady. It offers "AI-Free" certification for songs and extended the scheme in March 2026 to podcasts, short videos, audiobooks, films and photography. Press coverage describes a process based on a sworn declaration by the creator rather than audio analysis, and the company's own site does not explain the method. The site gives no prices, names no partners and does not say how many works the scheme covers. The figures quoted in the press were supplied by the company itself (MusicRow, 06/03/2026; humanable.com, accessed on 02/10/2026).

HumanMade Certified (humanmade-certified.com) describes itself as an initiative of the British record label AudioWave Records and Banny Decks. It issues a numbered certificate linked to the creator and the track, which can be checked online. The pages we could access gave no details of the verification method, prices or the operator's identity. On 02/10/2026, the label's page pointed to an inactive, parked domain. We therefore regard this service as unverified (humanmade-certified.com, accessed on 02/10/2026).

Verified Human Inc (verifiedhumancert.com), a Delaware company, offers certification for individual music tracks and podcasts. It is free for recordings with ISRCs dating from before 2022, and the company offers packages for record labels without publishing their prices. Its site describes the mark as "the federal certification mark". However, the USPTO (the United States Patent and Trademark Office) register records an application filed on 24/02/2026 under serial number 99666944 that is still under examination and has not resulted in registration. A second non-final office action was issued on 21/09/2026 (USPTO TSDR, accessed on 29/09/2026; verifiedhumancert.com).

VerifiedHuman.info is a separate organisation from Verified Human Inc. Its process relies on creators' own declarations, and the site describes the service as "not an audio analysis tool". Creators answer questions about their process and tools and are then assigned one of five levels. Only the three highest levels qualify for certification, from VH5, fully human, to VH3, human-led and AI-assisted. Users pay what they choose, which can be nothing. The organisation claims "hundreds of certified creators across 25+ countries" since April 2023, a figure it reports itself, and has been a contributor member of C2PA since 01/2026 (verifiedhuman.info, accessed on 02/10/2026).

None of these five services publishes an audited method, and no streaming platform or distributor recognises their attestations as sufficient grounds to reverse a decision. Where published, prices range from US$0 to US$2 per track. The existence of these services shows a demand for ways to document human provenance, although it does not yet show that those making the decisions trust this documentation.

Sources

Updated 13/10/2026