Framework, definitions and methodology
Most public debates conflate different realities; without a taxonomy, the 'percentages of AI music' become slogans. Primary sources for 2020-2026 mostly allow us to measure: (i) upload flows (new tracks delivered to DSPs) and (ii) fraud/streaming on detected tracks.
Operational definitions used here
Traditional music (baseline): works whose main audio comes from a human performance/production (even if ML/AI tools assist editing, mixing, mastering, or cleanup).
AI-generated music (in the 'generative' sense): works where at least one expressive component is generated by a model (lyrics, melody/structure, instrumentals, vocals). This category is broad and rarely measured robustly by DSPs.
Fully AI-generated music: works where the audio file (instrumental + vocals where applicable) is produced 'end-to-end' by a generator (text-to-audio / audio-to-audio) with human intervention limited to prompting/curation/superficial editing, with no capture of human performance as the primary source. This is the axis that Deezer publicly claims to detect and quantify.
Methodology of the 2020-2026 numbers
Direct measurements (high confidence): Deezer publications on the share of 'fully AI-generated' uploads at several dates (Jan 2025, Apr 2025, Nov 2025, Jan 2026) and related information (daily volumes, anti-recommendation policies, fraud).
Corroborating indicators (medium confidence): Spotify announcements on the removal of 'spammy tracks' and the rollout of AI rules (these do not directly measure '% AI', but signal operational pressure on the distribution chain).
2020-2024 estimate (low to medium): there is no public and stable equivalent to the Deezer series for 2020-2024; any global annual quantification is therefore an exercise in estimation. The bounds proposed rely on the chronology of consumer access to generators (acceleration in 2023-2024) and the presence of already-active 'AI music' platforms (e.g., Boomy) but without evidence that their production dominated DSP flows before 2025.
Main primary sources used
They focus on: DSP disclosures (Deezer, Spotify), institutional reports (US Copyright Office), legislative documents (DSM Directive), major-label press releases (UMG, WMG), court decisions (Beijing Internet Court), and standards documents (DDEX via WIPO).
Measurements and estimates of the share of AI-generated music from 2020 to 2026
What can be measured without excessive speculation
The only public data series (2025-2026) explicitly describing the share of 'fully AI-generated' tracks in upload flows comes from Deezer:
January 2025: >10% of tracks delivered daily detected as AI (~10,000 tracks/day).
April 2025: 18% (~20,000 tracks/day).
November 2025: 34% (~50,000 tracks/day).
January 2026: ~39% (~60,000 tracks/day).
These tracks are announced as excluded from algorithmic recommendations (and, depending on the period, labeled) to limit dilution and fraudulent monetization.
The 'upload vs listening' contrast is central: mid-2025, Deezer indicated that detected AI music accounted for ~0.5% of streams, but that a majority of those streams was fraudulent (up to ~70%). In January 2026, Deezer reported that in 2025 up to ~85% of fraudulent activity tied to these detected AI tracks was detected/removed, and that a massive volume of AI uploads persists.
Reference chart on 'fully AI' uploads (Deezer)
The points plotted above include only dated publications (Jan 2025, Apr 2025, Nov 2025, Jan 2026).
Annual table 2020-2026: what is observed vs estimated
The table below answers the request for 'annual 2020-2026', but it must be bluntly clear: before 2025, there is no consolidated public measurement of '% AI music' on DSPs comparable to Deezer 2025-2026. The 2020-2024 bounds are therefore orders of magnitude (ceilings/floors) and not established statistics.
Year | % observed 'fully AI' in uploads (Deezer proxy) | Plausible global estimate 'fully AI' in DSP uploads (range) | Plausible global estimate 'fully AI' in streams (range) | Calculation basis / rationale |
|---|---|---|---|---|
2020 | n/a | ~0-0.1% | ~0-0.01% | AI music generation still mostly experimental; absence of DSP disclosures; research-grade models exist (e.g., Jukebox). |
2021 | n/a | ~0-0.5% | ~0-0.05% | Emergence of public 'AI-assisted' creation apps, but no public DSP series; Boomy claims high-volume production from its public launch (self-reported). |
2022 | n/a | ~0-1% | ~0-0.1% | Pre-Suno/Udio consumer period; distribution mainly via niches (mood music, noise, 'functional' content). (Low public observability.) |
2023 | n/a | ~0.1-1% | ~0.01-0.2% | Acceleration of visibility via voice deepfakes and product launches; viral examples (cloned voices) but no indication of a majority of uploads. |
2024 | n/a | ~0.5-5% | ~0.05-0.4% | Democratization of text-to-music generators (quality and accessibility) + major legal tensions; still no annual public DSP reporting. |
2025 | ~24% (annual average approximated from Deezer points) | ~10-30% | ~0.3-1% | Deezer series (10%->18%->34%); strong fraud, weak conversion to listening ('0.5% of streams' on Deezer mid-2025). |
2026 (early) | ~39% (Jan 2026) | ~15-40% | ~0.5-2% | Deezer point 39% + anti-fraud industrialization; strong uncertainty on actual consumption excluding fraud. |
Correct (non-fallacious) reading of these figures: '% of uploads' != '% of listening' != '% of revenues'. The 2025-2026 dynamic mainly shows upload industrialization (part of which is opportunistic/fraudulent) and a DSP reaction (demotion, tagging, removal, disclosure standards).
Concrete use cases and case studies since 2020
Lyrics generation and textual ideation
The most common case in 'professional' production is not a '100% AI album' but the use of AI as an ideation accelerator: theme variants, verse/chorus structures, rhymes, hooks, style prompts. On text-to-music platforms, lyrics generation is often integrated in the flow (e.g., tools combining text generation + audio generation), mechanically increasing the volume of formally 'complete' songs produced by non-musicians.
Case study (human/AI hybrid): BBL Drizzy relies on initial vocal material generated via an AI tool; the human element (concept, lyrics, cultural timing) remains decisive, but AI serves as a fast 'rendering engine.'
Instrumental composition and 'end-to-end' audio generation
Two uses coexist:
B2C: instant generation (prompt) of 'radio-like' songs for TikTok/YouTube/ambient use, feeding a volume economy. Industrialization is read less in the charts than in upload metrics (see Deezer).
B2B: generation of 'mood music' for soundtracks (commerce, video creators, podcasts) where the main pressure is cost/speed rather than artistic identity. A study commissioned by the International Confederation of Societies of Authors and Composers projects very high penetration of the 'music libraries' segment (order of magnitude ~60% of library revenues in 2028) and ~20% of streaming revenue by 2028 for AI-generated music. These are projections, not measurements, but they frame the economic logic (substitution on utilitarian uses).
Voice synthesis, voice cloning and 'digital replicas'
Three usage models are observable:
Unauthorized imitation (maximum risk): the episode of the track Heart on My Sleeve (voice imitating Drake and The Weeknd) becomes a 2023 textbook case: strong virality, takedowns at rights holders' request, and an immediate debate on voice rights and training.
'Demo' imitation by artists (gray zone): David Guetta publicly showed using a synthesized rapper voice (Eminem) in a non-commercial / performance context, illustrating the ease of cloning and its ambiguities (confused audience, absent consent).
Voluntary voice licensing (contractual model): Grimes announced profit sharing for creations using her voice via a licensing/royalty arrangement ('opt-in voice model' logic). Similarly, Holly Herndon has built a 'vocal instrument' approach (Holly+) oriented toward consent and experimentation.
Mixing, mastering and post-production
Here, AI mainly acts as assistive technology: setting suggestions, dynamics correction, audio restoration, stem separation.
LANDR markets an automated mastering engine trained on sound engineers' workflows and years of data; the promise is speed + consistency, with a 'assistant' rather than 'replacement' repositioning.
iZotope explicitly structures its suites around ML-based assistants (Master Assistant, Mix Assistant) and modules that process stems (e.g., per-stem EQ). They are typically used by humans to speed up iterations.
On separation, the open-source tool Spleeter (from Deezer) illustrates the spread of 'industry' ML tools to consumers (remix, archives, analysis). But this is not 'generation' but 'decomposition' (with derivative-rights implications).
Recommendation, curation and 'AI in consumption'
Recommendation is the oldest and most systemic AI domain: personalization, playlists, ranking.
Spotify launched an 'AI DJ' (2023) combining personalization and generation of voiced commentary, and later added interactions (voice queries).
In 2025, Spotify also announced support for an AI disclosure standard in credits via DDEX, a sign that 'governance' is moving toward the metadata supply chain, not just end-of-chain moderation.
Ecosystem of actors
The AI music value chain is now segmented into five layers: (1) audio generators (text-to-music), (2) voice synthesis, (3) studio tools (mix/master, stems), (4) DSPs & distribution, (5) standards/rights collection.
The table below is intentionally comparative and 'pragmatic' (commercial model + license), because that is where the real risks for artists/labels lie (who owns what, and on what terms).
Actor | Country (HQ/main activity) | Commercial model | Dominant AI features | Public clients/partners | Created (per consulted sources) | License / use conditions (summary) |
|---|---|---|---|---|---|---|
Deezer | France | Streaming + tech licensing | 'Fully AI-generated' detection, tagging, anti-fraud; exclusion from recos | Tech licensing deal with SACEM | n/a (not documented here) | Internal detection; announced B2B opening (licensing). |
Bandcamp | United States | Music marketplace | AI content moderation/ban | n/a | n/a | Announced ban on music/audio generated 'wholly or substantially' by AI; ban also on impersonations. |
Suno | United States | Subscription (freemium); partnerships | Text-to-song (vocals + instru), fast generation | 'Groundbreaking' partnership with Warner Music Group | 2022 (secondary source) | Plan-differentiated terms: commercial use granted to paying subscribers; ownership/monetization clauses subject to local law. |
Udio | United States | Subscription + 'walled garden' approach (post-settlement) | Text-to-music; remix/mashup (evolution toward licensed platform) | Strategic deal with Universal Music Group; deal with WMG | 2023 (secondary source) | Shift toward licensed service; download restrictions mentioned after deals; commercial use/ownership depends on terms and versions (high instability). |
Boomy | United States | Freemium + monetization (distribution) | Track generation + distribution | n/a | 2019 (Boomy source) | Platform emphasizes creation and publishing; self-reported volume statements (treat with skepticism). |
LANDR | Canada | SaaS | AI mastering, DAW plugins | n/a | 2014 (secondary source) | Standard software license; audio outputs owned by the user, but no upstream rights warranty. |
iZotope | United States | Software/bundle sales | ML assistants for mix/master; restoration | n/a | n/a | Software license; 'assist' tools (not necessarily generation). |
Meta | United States | Open source + R&D | Audio watermarking (AudioSeal) | n/a | n/a | MIT-licensed code on AudioSeal repo; proactive watermarking logic for synthetic-voice detection. |
Google DeepMind | UK/US (Google group) | R&D | Watermarking (SynthID) applied to audio generated via music model | n/a | n/a | 'SynthID' watermarking announced for audio generated via music model; traceability tool, not a right. |
Warner Music Group | United States | Major (rights + licensing) | Licensing for 'licensed' AI creation | Deals announced with Suno and Udio | n/a | Opt-in/consent contractual frameworks (announcement), litigation settlement. |
Universal Music Group | EU | Major | AI licensing + platform | Strategic deal with Udio | n/a | Positioning as 'licensed AI music creation platform'. |
Recording Industry Association of America | United States | Trade group | Litigation / enforcement | Complaints against generators | n/a | Filings (e.g., Udio) and litigation strategy. |
Rigor note: on many AI platforms, 'licenses' change quickly (ToS versions), and the terms 'ownership', 'commercial use' and 'copyrightability' are not equivalent. Typical example: a platform may 'grant commercial use' while declaring that it cannot guarantee a copyright 'vests' legally in the output.
Comparative legal framework and copyright
A useful comparison must cover: (i) output ownership/protection, (ii) legality of training (inputs), (iii) voice/replicas (neighboring rights, personality rights), and (iv) post-2020 case law.
Comparative table by jurisdiction
Jurisdiction | Outputs: ownership / protection | Inputs (training): key rules | Voice / deepfakes | Decisions / milestones since 2020 |
|---|---|---|---|---|
European Union (incl. France) | In practice, protection relies on originality and a human contribution; the AI Act tightens model-side transparency obligations (by category). | DSM Directive: TDM exception with opt-out (Art. 4); litigation and debate on 'machine-readable' opt-out form. | AI Act imposes transparency obligations for some deepfakes; layered regulation (data protection, personality). | German decision (Hamburg) applying the TDM exception (LAION case): one of the first European decisions often cited on the TDM exception. |
United Kingdom | Specificity: 'computer-generated works' (CDPA s.9(3)) attributing authorship to the person who made the 'arrangements necessary'. | Limited TDM exception (non-commercial research, lawful access); UKIPO consultations on AI adaptation (creative industries vs AI tension). | No federal 'right of publicity' as in the US; typical remedies: passing off, breach of confidence, etc. (high operational uncertainty). | Consultation and public debate 2024-2025; pressure from artists/industries (institutional sources cited by USCO). |
United States | USCO: material wholly generated by AI is not copyrightable; case-by-case if sufficient human contribution. | 'Fair use' at the center of litigation; USCO Part 3 analyzes the issue and discusses licensing options. | USCO Part 1 recommends a federal law on 'digital replicas' (protection + 1A exceptions, safe harbor). + state patchwork (e.g., Tennessee 'ELVIS Act'). | Appeal (2025) confirming the human-author requirement in an 'AI-only' case. + major-label complaints against generators (2024). |
Canada | Cautious approach: human author required in practice; government consultations on modernization. | Uncertainty on TDM/fair dealing applied to training; dedicated federal consultation (ISED). | Personality rights/protection against impersonation via common law and related texts; no AI music framework as explicit as a 'digital replica act'. | 'Copyright in age of generative AI' consultation (paper) and 2024-2025 debates. |
China | Case law trend to recognize protection if the output reflects personalized expression through intellectual investment (at least in analogous cases). | Strong administrative regulation on 'deep synthesis' and generative AI services (labeling/management obligations) (framework evolving). | 'Platform governance' regulatory approach: obligations to signal/label synthetic content. | Beijing Internet Court (2023) ruling: the court concludes there is a human author (prompt/parameters/selection) and recommends AI labeling in good faith / public's right to know. |
India | Legal definition of the author of a 'computer-generated' work: 'person who causes the work to be created' (interpretation tension). | Still fragmentary regime on AI training; debates focus on reform/clarification. | Strong judicial activation of 'personality rights' against deepfakes (actions involving platforms; injunctions). | Rise of 'personality rights' litigation (e.g., high-profile 2025 actions against Google/YouTube for deepfakes). |
Critical point (often ignored): most copyrights protect expression, not 'style'. Yet voice synthesis and 'style imitation' make the boundary between 'expression' and 'identity attribute' indexable/monetizable. USCO explicitly recommends covering realistic 'digital replicas', but does not recommend (at this stage) including 'style' within the scope of new federal law.
Risks, stakes and recommendations
Risks and stakes
Ethical (consent, deception, attribution)
Voice synthesis makes credible performances possible without consent, with a risk of deception (audience, media, platforms) and harm to reputation. 'Opt-in' models (voluntary voice licensing) reduce risk but shift the question toward governance (who controls the model, what usage limits, revocation).
Economic (royalties, fraud, dilution)
The measurable short-term threat is fraud and dilution: Deezer describes a majority of detected AI streams as fraudulent (mid-2025: up to 70%) and announces an intensification in 2025 (up to ~85% of fraudulent activity tied to detected AI tracks, per its 2026 communication). Spotify acknowledges massive 'spam' pressure (removal of >75M tracks over 12 months) and deploys anti-impersonation measures + filtering + disclosure.
Artistic (standardization, cultural noise)
The flood of 'functional' content (mood music, noise, cloned tracks) can shift incentives: manufacturing volume optimized for the algorithm rather than developing artistic worlds. DSPs respond with delisting/restriction.
Market (licensing negotiation, new intermediaries)
The 2025-2026 trajectory shows normalization via licensed deals between majors and AI players to create models trained on authorized catalogs and offer 'walled-garden' experiences. This can create an 'entry barrier' for smaller players while securing catalog monetization.
Technical (detection, watermarking, circumvention)
Technical trajectories converge toward: (i) signature detectors (Deezer approach), (ii) proactive watermarking ('AudioSeal' and 'SynthID' approaches). None of these mechanisms is a complete solution: watermarking assumes adoption at source; post-hoc detection is a cat-and-mouse game.
Practical recommendations
For artists and producers (individuals)
Reading the ToS is not 'optional': it is your real contract. Otherwise, you expose yourself to false security: 'I have commercial rights' does not imply 'I hold an enforceable copyright', and platforms sometimes say so explicitly. Concretely: document the human contribution (stems, sessions, scores, vocal tracks), keep proof of subscription/plan at generation time, and avoid any voice/persona imitation without written consent (release + license).
For labels / publishers / managers
The rational pivot is the supply chain: require AI disclosures in credits (alignment with DDEX), contractual clauses on training (opt-in/opt-out, remuneration, audit), and 'digital replica' clauses (authorized uses, duration, revocation, media/country scope). On fraud, investing in fingerprinting + bot detection + 'royalty hygiene' (demonetization of suspicious streams) is a revenue-protection imperative, not a 'moral' topic.
For policymakers / regulators
The most useful texts are not those that 'ban AI', but those that make the economy traceable: (i) clear rules on replicas (voice) and fast injunction procedures, (ii) information obligations on training and machine-readable rights reservations (opt-out), (iii) interoperable labeling standards (metadata). The USCO recommendation on a 'digital replica' law is an example of an architecture (right + exceptions + safe harbor). On the EU side, the operational clarity of the TDM opt-out (DSM Directive) is a concrete friction point that deserves harmonization.
Timeline 2020-2026
Milestones relating to the measurement of 'fully AI' uploads and DSP responses come from Deezer/Reuters publications (2025-2026) and Spotify communications (2025), while the legal milestones come from USCO reports and public litigation.
Main references in APA 7
CISAC & PMP Strategy. (2024). Study on the impact of Generative AI in Music & Audiovisual industries (Executive summary / slides).
Deezer. (2025, January 24). Deezer deploys cutting-edge AI detection tool for music streaming.
Deezer. (2025, April 16). Deezer reveals 18% of all new music uploaded to streaming is fully AI-generated.
Deezer. (2025, November 12). Deezer and Ipsos survey: AI fools 97% of listeners (incl. 34% uploads).
Reuters. (2026, January 29). Deezer licenses AI music detection tool to Sacem, plans wider rollout.
United States Copyright Office. (2024). Copyright and Artificial Intelligence, Part 1: Digital Replicas.
United States Copyright Office. (2025). Copyright and Artificial Intelligence, Part 2: Copyrightability.
United States Copyright Office. (2025). Copyright and Artificial Intelligence, Part 3: Generative AI Training (Pre-Publication Version).
Directive (EU) 2019/790 (DSM). (2019). Text and data mining exception and rights reservation (Article 4).
EUIPO. (2024). Hamburg District Court decision: LAION v. Kneschke (TDM exception context).
Warner Music Group. (2025, November 25). Warner Music Group and Suno forge groundbreaking partnership.
Universal Music Group. (2025, October 29). Strategic agreements with Udio for a new licensed AI music creation platform.
Spotify. (2025, September 25). Spotify strengthens AI protections for artists, songwriters and listeners.
Beijing Internet Court. (2023). Civil Judgment (AI-generated image authorship and disclosure).