Glossary

AI Act (article 50)

Article 50 of Regulation (EU) 2024/1689, known as the EU AI Act, sets out transparency requirements. Generative AI providers must mark their audio, image, video and text outputs in a machine-readable format, and professional users must disclose deepfakes. The article has applied since 02/08/2026, although the deadline for marking outputs from systems already on the market has been deferred to 02/12/2026. It does not, however, require platforms to tell listeners which tracks are AI-generated. See Standards and regulation.

AI-assisted

Describes a human-led work in which an AI tool contributed to part of the result, such as mastering, a generated sound or an initial draft of lyrics. The musical vision, composition and direction remain the artist's own. It is the middle category of the declaration scale the manifesto calls for, and Traxsource has used it as a label since 01/07/2026.

AI-generated

Describes a track produced by a generative model from a prompt or other starting material, without a human composing or performing it. According to Deezer, fully AI-generated tracks made up more than half of the tracks delivered to its service each day at the June 2026 peak, around 90,000 a day, but accounted for only 1 to 3% of streams (figures supplied by the platform on 21/07/2026).

Appeal

A way to challenge a decision that labels, blocks or demonetises a track. With no appeal procedure common to all platforms, artists go through their distributor and provide session files and dated drafts, while the platform has the final say. The manifesto states that artists must be told why the decision was made, have a named person they can contact, know what evidence will be accepted and be given a deadline for a response.

Artificial streams

Plays generated by bots, click farms or accounts left running on a loop, rather than by listeners, with the aim of collecting royalties. Platforms exclude streams they identify as fraudulent when calculating royalty payments. Deezer reported that up to 85% of streams of fully AI-generated tracks were fraudulent in 2025, depending on the month. The platform published this figure on 29/01/2026 and repeated it on 21/07/2026. These streams take money from the shared pool from which every artist is paid.

C2PA / Content Credentials

An open technical standard that allows a signed record of provenance, known as "Content Credentials", to be attached to a media file. The record identifies the tool that created or modified the file and when this happened; the slightest alteration to the file breaks the signature. The current version, 2.4, dates from April 2026. The standard attests to the signer's declaration about the file's origin, which is not enough to prove that it was made by humans.

DDEX

The organisation that standardises the metadata record labels and distributors send to streaming platforms. Version 4.3.2 of its ERN format allows release metadata to state that a recording or contribution was made wholly or partly with generative AI. This information is self-reported, and Spotify has displayed it in track credits since launching a beta version on 16/04/2026.

Demonetisation

A platform's decision to stop paying royalties on a track while leaving it online. TIDAL, for instance, announced that, since 29/06/2026, it no longer pays royalties on tracks it identifies as fully AI-generated. The track remains available to listen to, so the loss of income can go unnoticed. The manifesto therefore includes demonetisation among the decisions artists must be able to challenge.

Detector

Software designed to estimate whether an audio excerpt was produced by a generative model, usually by recognising the statistical traces the model leaves in the sound. The providers of detectors used by platforms and distributors report high accuracy rates, although, to our knowledge, none of these figures has been independently audited. See Detectors.

Dilution

The effect of mass-produced music on artists' share of streaming revenue. Most platforms pay royalties from a single pool, distributed in proportion to streams. Each stream attracted by an automatically produced track, or generated through related fraud, therefore reduces the share going to human artists. This effect arises from volume alone and does not depend on tracks being mislabelled.

Distributor

An intermediary, such as DistroKid, Believe, TuneCore or CD Baby, that delivers releases and their metadata to streaming platforms and collects payments from them. Distributors now also act as the first layer of screening, with some blocking tracks from unlicensed generators before delivery. It is also through their distributor that artists challenge an AI label applied by a platform.

DSP

Short for digital service provider, the industry's term for a streaming or download platform such as Spotify, Apple Music, Deezer, TIDAL, YouTube Music or Traxsource. Each DSP has its own AI policy, which determines how tracks are labelled and recommended, and how royalties are paid.

False positive

A track made by humans that a detector incorrectly classifies as AI-generated. Platforms say such errors are rare. Deezer reports fewer than 1 human track misclassified in 10,000, while Ircam Amplify reports a rate below 1%. These figures come from the providers and have not been independently audited. The right of appeal the manifesto calls for does not depend on how often such cases occur, since any decision that labels, blocks or demonetises a track must be open to challenge.

Human provenance

The documented fact that a work was composed, performed and produced by people. The term describes where a work comes from and makes no judgement about its quality or value. This site therefore prefers it to words such as "real" or "genuine". Artists can document that provenance using material they already have, such as session files, stems or dated drafts.

Hybrid creation

A work combining human creativity and AI tools in varying proportions, from a song made by a human with just one AI-generated sound to a human vocal over AI-generated backing. Yet platforms generally use binary classifications, while detecting AI-generated stems within hybrid works remains an open research question (arXiv 2607.26874, 07/2026). The manifesto states that artists must be able to declare the role of AI in their work without having a verdict imposed on them.

ISRC

Short for International Standard Recording Code, the unique identifier assigned to each recording so that its streams and royalties can be tracked from one platform to another. Appeal procedures and third-party services refer to tracks by their ISRC, but the identifier says nothing about how the recording was made.

Labelling / tagging

Identifying a track as AI-generated or AI-assisted in a platform's interface. Deezer applies the label at album level. Apple Music will display "Made With AI" labels based on tags supplied by content providers, while TIDAL and Traxsource base their labels on their own detection systems. A label may therefore reflect the uploader's declaration or a detector's verdict, two sources that do not always agree.

Metadata

Information that accompanies a recording, including titles, credits, identifiers and rights information, and now also AI disclosures. In practice, this is where information about provenance is recorded, whether as a DDEX declaration or a C2PA manifest. Metadata is no more reliable than the person who wrote it, and it can be lost when a file is re-encoded.

Open licence

A licence that allows anyone to use and examine a tool, specification or dataset under terms set out in advance. The manifesto commits VERIMUSIC to publishing its tools, standards and protocols for documenting human provenance under an open licence, so that anyone can examine them without having to take VERIMUSIC's word for it.

Stem

A single element of a mix, exported as a separate audio file, such as drums, bass, vocals or keyboards. Because they show how a track was built, stems are among the evidence distributors request when a decision about a track is challenged. Research continues into detecting AI-generated stems in mixes otherwise made by humans.

Training data and consent

The music used to train a generative model. Whether this use requires the consent of the works' creators is being debated in courts in several countries, while some developers of music generators are now signing licensing agreements with labels and distributors. The manifesto states that works must not be used to train AI without their creators' consent.

Voice cloning

Generating singing or speech in the voice of a specific person, using recordings of that voice. Under the AI Act, any cloned voice presented as that person's own counts as a deepfake and must be disclosed. Spotify allows vocal imitation only with the consent of the artist being imitated, and the manifesto states that no one's voice may be cloned, released or monetised without their consent.

Voiceprint

The set of measurable characteristics by which a person's voice can be identified. As biometric data, a voiceprint falls within the special categories of data defined by the GDPR (General Data Protection Regulation), so that collecting it, storing it or matching it against recordings requires a lawful basis and informed consent. The manifesto states that everyone must uphold this protection, including services that claim to verify voices.

Watermark

An inaudible signal deliberately embedded by a generator in its audio, designed to withstand compression and allow the output to be recognised later. One example is SynthID, which Google uses with its own models. A watermark makes it possible to identify output from the tool that embedded it, but its absence proves nothing about the file.