Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Prologue — A decade of "I pick a song" becoming "the AI picks for me"

Spring 2026, a cafe in Gangnam, Seoul. A designer in her 30s opens Spotify on her morning commute at 7:50. At the top of the screen sits Daylist; today's caption reads "monday morning lo-fi indie folk wake-up". She looks at that one line and thinks: that is exactly my mood. Inside, 10 of the 30 tracks are songs she has never heard before.

At 3 p.m. the same person opens Spotify again. The Daylist caption has shifted to "tuesday afternoon chill house focus instrumental". Half the tracks have refreshed. She taps the heart on this one too.

At 8 p.m. she hits the Spotify DJ button. A bass-heavy voice says in English: "I'll lay down an evening set in the jazz-funk mood you've been on all week," then half an hour of radio rolls. She is less surprised by the music than by the question of how something knows her taste this well.

At the same hour a student in Shibuya, Tokyo plays Apple Music's Discovery Station, an office worker in Busan plays Melon's "this moment" playlist, and a composer in New York types "rainy berlin techno after a breakup" into Amazon Music Maestro to assemble a playlist. Music consumption in 2026 has almost fully shifted from "I choose a song" to "the AI suggests a song and I decide whether to skip".

This piece walks the full landscape. From Spotify · Apple Music · Amazon Music · YouTube Music · Tidal · SoundCloud through Melon · Genie · Bugs · FLO to AWA · LINE Music — what AI each platform uses, what ML architecture is running underneath, who is paid for which song, and how AI-generated music is shaking the whole market.

1. The 2026 music streaming map — five regions

A single map of the market shows five overlapping regions.

┌────────────────────────────────────────────────────────────────────┐

│ Global Big 5 │

│ Spotify · Apple Music · Amazon Music · YouTube Music · Tidal │

│ ~800M paying subs combined, the front line of AI discovery │

├────────────────────────────────────────────────────────────────────┤

│ Hi-Fi / audiophile lane │

│ Tidal HiFi+ · Qobuz · mora qualitas · Apple Music Lossless │

│ "sound quality" is the axis, not AI │

│ K-pop / Korea region │

│ Melon (Kakao) · Genie (KT) · Bugs (NHN) · FLO (SKT) · VIBE │

│ K-pop discovery, domestic charts, telco bundles │

│ Japan region │

│ AWA (CA+Avex) · LINE Music · dHits · Recochoku · mora │

│ + Spotify JP · Apple Music JP · Amazon Music JP │

│ Creator / community / Web3 │

│ SoundCloud · Bandcamp · Audius · Royal · Sound.xyz │

│ Underground, direct-to-fan, tokenised royalties │

└────────────────────────────────────────────────────────────────────┘

These five regions are separate but they all meet in one person's ears. You hear a K-pop song on Melon, see it on Spotify's global chart, watch the video on YouTube Music, and read the lyrics on Apple Music. This piece walks the five in turn.

2. Spotify — the company that made discovery a standard

Founded in Stockholm in 2008, Spotify in spring 2026 has roughly 700 million monthly active users and around 270 million paying subscribers (estimated). It is the single largest streaming platform in the world and the company that defined "AI discovery" for the rest of the industry. Key features:

**Discover Weekly** — launched 2015. A weekly 30-track playlist of "songs you'll probably love" delivered every Monday. A hybrid of Collaborative Filtering, NLP-derived song embeddings, and raw audio features. When it shipped, the meme "Spotify knows me better than my friends" defined a whole industry.

**Release Radar** — every Friday, new songs from artists the user follows or has been listening to. The standard channel for new-music discovery.

**Daily Mix** — clusters a user's listening into roughly six buckets and produces an infinite mix per bucket.

**Daylist** — launched September 2023. A dynamic playlist that combines time of day, mood, and genre. Even for the same user the morning, afternoon, and evening copies differ. The auto-generated copy is itself shared as a meme.

**Spotify DJ** — launched February 2023. An AI voice ("X") introduces the next track with "this is what's coming, because…". Started with OpenAI for text and ElevenLabs-style synthesis, later moved in-house. English first, then Spanish, Portuguese, French, German, Italian.

**Spotify AI Playlist** — launched April 2024. A single prompt like "rainy sunday afternoon lo-fi" produces a 30–50 track playlist. Beta first in the UK and Australia, then global. Direct competitor to Amazon Music Maestro.

**Smart Shuffle** — injects songs the algorithm thinks fit your playlist into the shuffle so the playlist effectively extends forever.

**Niche Mixes** — narrow context/genre mixes such as "nu-disco running" either built by the user or recommended by Spotify.

**Spotify Wrapped** — early December every year. A year's listening data packaged into shareable infographics. Since 2024 AI-generated commentary appears in places, turning a yearly review into a social event.

3. Spotify's recommendation ML stack — how it all actually runs

Underneath one word "recommendation" several models run at once.

**1) Collaborative Filtering (CF)** — Spotify's original foundation. The signal is "users like you listened to this". Matrix factorisation, ALS (Alternating Least Squares), evolving to learned embeddings. Roughly half of Discover Weekly is this signal.

**2) Content-based Filtering (CBF)** — features extracted from the audio itself. Mel spectrograms into a CNN to produce song embeddings, then nearest neighbours by audio similarity. Crucial for cold-start of new songs.

**3) NLP-derived song / artist embeddings** — learned from blog text, reviews, playlist titles, tags. Captures "what moods is this song described with".

**4) Two-tower models** — a user tower and an item tower produce embeddings that meet at a dot product. This has become the standard retrieval architecture across recommender systems. Spotify publishes multiple papers at RecSys describing variants.

**5) Reinforcement Learning (RL)** — used to decide track sequence in radios and in the DJ. "What should this user hear in the next 30 minutes to maximise session satisfaction" framed as a reward function.

**6) LLM-based natural-language interface** — for AI Playlist, the prompt is parsed and decomposed into mood / genre / BPM / year / region axes, which then feed back into CF / CBF / two-tower retrieval.

| Feature | Dominant ML signal |

| --- | --- |

| Discover Weekly | CF + song embeddings + audio features |

| Release Radar | Follow graph + CF + audio similarity |

| Daily Mix | Listening clusters + CF |

| Daylist | Time-of-day context + CF + mood labels |

| DJ | RL radio sequencing + LLM commentary |

| AI Playlist | LLM prompt decomposition + two-tower retrieval |

| Smart Shuffle | CF + song-embedding similarity |

4. Apple Music — the opposite philosophy to Spotify

Launched June 2015; as of 2026 around 100 million paid subscribers (estimated). Apple Music's philosophy is the opposite of Spotify's.

**Editor + algorithm hybrid** — from day one Apple has emphasised human editorial curation. The A-List, New Music Daily, Today's Hits and similar playlists are still maintained by human editors; the algorithm sits beside them as a helper.

**Discovery Station + Stations** — endless radio seeded by song, artist or genre. In 2024 Discovery Station added a mode that intentionally raises the share of songs new to you.

**Apple Music Sing** — December 2022. Lyric emphasis, vocal-volume control, and a karaoke experience. Available on iPhone, Apple TV and iPad. Particularly popular for K-pop and J-pop.

**Spatial Audio + Dolby Atmos** — Apple's most aggressive differentiator. Coupled with head-tracking on AirPods Pro/Max to standardise a sense of "audio with space".

**Apple Music Classical** — launched March 2023 on the back of the Primephonic acquisition. A dedicated app that handles classical metadata (composer, conductor, ensemble, movement) properly — the search problem that regular Apple Music could not solve.

**Replay** — Apple's answer to Wrapped. Not just end of year but updated continuously throughout the year.

**Time-synced lyrics** — one of the most visible differences from Spotify. Lyrics scroll line by line as the song plays, identical on iOS, Mac and Apple TV.

**Live Radio — Apple Music 1 / Hits / Country** — live radio channels with human DJs, interviews and first plays of new songs. A broadcast-like experience Spotify does not match.

5. Amazon Music — Maestro and Alexa fused

Amazon Music has tiers from a free Prime tier through Unlimited. As of 2026 Unlimited subscribers are estimated at roughly 100 million. Differentiators: "deep integration with Echo and Alexa" and "audiobooks via Audible".

**Amazon Music Maestro** — announced as a beta in April 2024. Accepts a prompt like "give me a playlist for a Sunday afternoon drive through the desert" and builds the set. Direct competitor to Spotify AI Playlist.

**Amazon Music HD / Ultra HD / Spatial** — launching HD in 2019 with "lossless at no extra charge" shook the market; the others eventually followed.

**Audible integration** — bundles Amazon Music Unlimited with Audible so music and audiobooks live in the same app.

**Alexa voice control** — natural-language music search is most fluent here: "Alexa, play me something like Phoebe Bridgers". Tightly integrated with Echo, Fire TV and cars.

**SongID + X-Ray for Music** — surfaces song info, lyrics and trivia in real time. More "metadata first" than Apple's lyric approach.

**Podcasts + Wondery integration** — Amazon's 2020 Wondery acquisition strengthened podcasts; in some markets a direct competitor to Spotify Podcasts.

6. YouTube Music — the video catalogue as a super card

YouTube Music as of 2026 has roughly 120 million paid subscribers (estimated). The single differentiator: every music video on YouTube lives inside the catalogue.

**Album & Song Mix (algorithmic radio)** — start one song, get infinite radio. Signals include YouTube watch history and search history. Hits the taste of a "person who watches YouTube" with unusual accuracy.

**YouTube Music Recap** — the annual recap started in 2021. Counter to Spotify Wrapped.

**Listening Modes** — activity / time-of-day / mood-based modes. Tap workout, commute, focus, relax and get an automatic curation.

**AI-generated Radio — Conversational Radio** (beta) — opened to a slice of users in 2024. A natural-language prompted radio such as "play me upbeat 90s hip-hop, no skits"; you can re-tune the result inline.

**Live concerts + music videos** — YouTube's strongest weapon. K-pop live, Japanese idol shows and indie live sets all live in one app.

**Background play + ad removal for Premium subscribers** — bundled with YouTube Premium, one of the strongest value bundles on per-user-per-month price.

7. Tidal — sound quality and artist payouts

Tidal launched in 2014, was acquired by Jay-Z in 2015, and acquired by Square (Block) in 2021. As of 2026 paid subscribers number in the millions — much smaller than the global big four — but its differentiation is clear.

**Tidal HiFi Plus** — Master Quality (moved from MQA to in-house FLAC-based) + Dolby Atmos + Sony 360 Reality Audio. A first pick for audiophiles.

**Direct-to-Artist payments** — policy introduced in 2022. A share of subscription fees is paid directly to the artists each user listens to the most. Tidal stresses transparency in artist payouts.

**Editorial tone** — strongest in hip-hop, R&B and electronic. The grain of the songs surfaced is visibly different from Spotify.

**Possible integration with Block (ex-Square)** — Cash App and Square merchants make a vision plausible: merch, tickets and tips for artists in one app. Some of that vision has materialised in 2026, some has not.

8. SoundCloud — the last redoubt of underground music

Founded in Berlin in 2007, SoundCloud remains the creator-first platform. Open uploads, unreleased tracks, mixes, beats, underground hip-hop and techno circulate here faster than anywhere else.

**SoundCloud Next Pro** — paid creator plan: unlimited uploads, analytics and monetisation.

**Fan-Powered Royalties** — introduced 2021, ahead of Tidal's similar policy. Pays artists from the listeners who actually played them, not from a single pooled bucket.

**AI Curated Playlists + GoNext** — AI-curation features launched between 2024 and 2025. Designed to expose smaller artists through algorithmic placement.

**Repost Service / Distribution** — self-distribution from SoundCloud out to other DSPs (Spotify, Apple Music, YouTube Music).

**Cultural position** — the home of 2010s SoundCloud rap; Lil Peep, XXXTentacion and Juice WRLD were discovered here. In 2026 it remains irreplaceable in indie, experimental and mix culture.

9. Other Western platforms — Deezer · Pandora · Qobuz

**Deezer** — French, founded 2007. Signature feature **Flow** — an endless personalised radio that one button starts. Closer in tone to a fusion of Spotify Daily Mix and Daylist. Strong in France, Brazil and Latin America.

**Pandora** — founded 1999 in the US, built on the Music Genome Project. Roughly 450 musical attributes are hand-tagged per song to power radio. Acquired by SiriusXM in 2019. Mostly invisible outside the US but the origin of "attribute-tagged" recommendation.

**Qobuz** — French. Audiophile-only. Hi-Res 24-bit downloads and streaming. A deeper lossless catalogue than Tidal. The value is "editors plus sound quality", not AI discovery.

**Bandcamp** — founded 2007. Direct-to-artist for indie. Artists sell songs, albums and merch themselves. Acquired by Epic Games, then sold to Songtradr in 2023. The pride of indie funding.

| Platform | HQ | Differentiator |

| --- | --- | --- |

| Deezer | Paris | Flow personalised radio |

| Pandora | Oakland | Music Genome radio |

| Qobuz | Paris | Audiophile Hi-Res |

| Bandcamp | Oakland | Direct-to-artist indie |

10. Korea — the five-way race between Melon · Genie · Bugs · FLO · VIBE

Korean music streaming runs by its own logic, separate from the global big five. Charts and telco bundles are tightly entwined.

**Melon** — operated by Kakao Entertainment. Number one in Korea. The "Melon chart" is the de-facto first-tier indicator for the K-pop industry. Linked to Kakao Talk and Kakao Pay, bundled into KakaoBank billing. The combination of editorial curation and chart authority is its strength. AI shows up as "For You Mix" and "Melon DJ" features.

**Genie** — KT's subsidiary Genie Music. Strong share through KT mobile bundles. Music sits alongside concerts and merch as services. Often slightly easier to access than Melon for global K-pop users because it depends less on telco-locked accounts.

**Bugs** — NHN Bugs. The Korean service most serious about sound quality. Tried FLAC 24-bit downloads and streaming early; its indie, jazz and classical curation is relatively deep.

**FLO** — Dreamus Company under SK Telecom. Bundled with SKT T Membership and T Universe; emphasises AI recommendation. Mood-based recommendation features like "today's mood" and "your taste analysis" arrived early.

**VIBE / NAVER VIBE** — Naver, launched 2018. Once famous as the first platform to try per-play royalty payments; later restructured alongside NAVER NOW. As of 2026 it lives inside Naver's music and audio ecosystem alongside NOW., keeping its own colour for K-pop discovery.

| Korean platform | Operator | Differentiator |

| --- | --- | --- |

| Melon | Kakao Entertainment | #1, chart standard, Kakao bundle |

| Genie | KT Genie Music | KT bundle, concert/merch packs |

| Bugs | NHN | High-quality audio, indie/jazz/classical |

| FLO | SKT Dreamus | SKT bundle, mood recommendation |

| VIBE | NAVER | NAVER ecosystem integration |

K-pop is global, but if you read K-pop "through the chart" you are almost always reading the Melon · Genie · Bugs · FLO four-to-five-way race.

11. Japan — AWA · LINE Music · dHits · Recochoku · mora, and the global incumbents

Japan adopted streaming a beat later than the rest of the world, so it ended up the most diverse market.

**AWA** — launched 2015. A joint venture between CyberAgent and Avex. Got mood and time-of-day based curation in early, plus LOUNGE (real-time listening together) as a social feature.

**LINE Music** — tied to LINE accounts. Profile BGM in LINE messenger, even call ringtones — built the position of "music inside the messenger". Strong in Japan, Taiwan and Thailand.

**Spotify Japan** — entered Japan in September 2016 with a free-ad-supported plus paid-Premium model. Daily Mix, Discover Weekly, Daylist and AI Playlist arrive identical to the global build.

**Apple Music Japan** — lyric sync, Spatial Audio and Replay identical to the global version.

**Amazon Music Japan** — the free Prime tier penetrated heavily in Japan. Echo penetration is relatively high here.

**dHits (NTT DOCOMO)** — a service for DOCOMO mobile subscribers. Embedded deep in Japan's mobile market.

**Recochoku** — historically the consortium of major Japanese labels; handles both downloads and streaming.

**mora** — Sony Music. The Japanese standard for Hi-Res downloads. The streaming arm **mora qualitas** focuses on lossless.

**KKBOX** — primarily Taiwan and Hong Kong, but a non-trivial Japanese user base.

Japan retained a culture of "downloading and owning songs" longer than most markets, so mora, iTunes Store and Recochoku remain working channels in 2026. The global big five have grown quickly, but local players still hold their place firmly.

12. Music recommendation ML — the five approaches

Step away from platform names and the big picture of music recommendation looks like this.

**1) Collaborative Filtering (CF)** — "listeners like you played this". The oldest workhorse. Weak at cold-start of new songs.

**2) Content-based Filtering (CBF)** — similarity computed on the song's own audio (BPM, key, energy, danceability, timbre). Strong at cold-start and at the long tail.

**3) NLP-based embeddings** — learned from blog text, reviews, playlist titles and tags. "What words is this song described with" becomes a signal.

**4) Two-tower / embedding retrieval** — user and song embedded into a shared vector space, then ANN (Approximate Nearest Neighbour) search via Faiss, ScaNN or Vespa.

**5) Reinforcement-learning radio sequencing** — "what should the next song be so the user keeps listening" framed as a Markov decision process. Drives the order of tracks in Spotify DJ and Apple Music Stations.

| Approach | Strength | Weakness |

| --- | --- | --- |

| CF | Very strong on big user bases | Weak at new songs and long tail |

| CBF | Handles cold-start | Weak signal on "why a person would love this" |

| NLP embeddings | Captures cultural context | Vulnerable to noise |

| Two-tower | Fast at large scale | Expensive to train |

| RL sequencing | Optimises session satisfaction | Reward design is hard |

The big platforms use all five at once. One model produces N candidates, another shrinks to K, a third decides final ordering (the classic "candidate generation -> re-ranking" two-stage).

13. Audio fingerprinting — Shazam · AudD · ACRCloud · MusicBrainz

The other axis of music recognition is fingerprinting — "what song is this".

**Shazam** — UK, founded 2002, acquired by Apple in 2018. A 5-to-10-second clip from the microphone is fingerprinted and matched against the catalogue. As of 2026 cumulative recognitions are estimated at 70 billion+. Standard on iOS and Android.

**AudD** — a developer-friendly API. Song recognition plus lyrics plus metadata in one call. Used by small apps and live-stream monitoring.

**ACRCloud** — strong in China and Asia. Music recognition plus broadcast monitoring plus live-stream monitoring. Used widely by radio and TV.

**MusicBrainz** — an open music-metadata wiki plus Picard (a client that auto-tags local files). Non-profit. Provides the canonical music identifier (MBID).

**SoundHound + Hound** — distinct from Shazam in that it recognises humming and singing too. Extended into a voice assistant (Hound) and car integration.

This space looks separate from streaming but it is part of discovery: fingerprinting is the bridge that moves a song heard on the street into the catalogue in one second.

14. AI-generated music — Suno · Udio · Stable Audio and the industry collision

Between 2023 and 2024, AI-generated music moved from "toy" to "industry threat".

**Suno** — launched 2023, through v3 and v4 producing full songs (vocal + lyrics + melody + accompaniment) from a single text prompt. Integration into Microsoft Copilot pushed general-user adoption sharply.

**Udio** — founded by ex-Google DeepMind staff, launched 2024. Direct competitor to Suno. Often described as cleaner with stronger vocal synthesis.

**Stable Audio (Stability AI)** — instrumental-focused. Used for video, game and ad backing tracks.

**Riffusion · MusicGen (Meta) · Lyria (Google DeepMind)** — academic and open lineage. Less polished than Suno or Udio but catching up quickly.

**RIAA v Suno/Udio lawsuit (June 2024)** — the US Recording Industry Association sued both companies for copyright infringement, alleging training data included unlicensed recordings. The music equivalent of NYT v OpenAI.

**Spotify's policy against AI clones** — strengthened between 2024 and 2025. Songs cloning a specific artist's voice with AI are removed when discovered. AI-generated songs that imitate only a genre still sit in a grey zone.

**Heart on My Sleeve (April 2023)** — an AI-cloned vocal track impersonating Drake and The Weeknd exploded on TikTok. Removed from every DSP within days, but it left the industry with the realisation that "anyone can now do this".

The industry's response is split. Some labels are negotiating licensing; others are suing across the board. As of 2026 the picture is dominated by un-settled grey zone.

15. Web3 music — Royal · Audius · Sound.xyz

Blockchain-meets-music experiments emerged alongside the 2021–2022 NFT boom.

**Royal** — founded 2021, with figures like JD Davis. Sells a slice of a song's future royalties as a token so fans become co-owners. Drew big attention; faded in the 2024–2025 market downturn.

**Audius** — founded 2018. Decentralised music streaming. Songs are stored on IPFS, the AUDIO token is the incentive. Built on Cosmos.

**Sound.xyz** — a music-NFT marketplace. Each release prints a limited NFT for early fans. Meaningful activity in indie between 2022 and 2023, then polarisation.

**Catalog Records · Glass Protocol · Nina Protocol** — similar indie experiments.

Web3 music's promises were large but it never settled at industry scale through the 2024–2025 downturn. The idea of tokenising a song survives in adjacent forms — concert tickets, merch and fan memberships — inside a small group of companies.

16. Royalties and distribution — DistroKid · CD Baby · TuneCore · Believe

In the streaming era artist income is set by per-play payouts, but distributors sit between artists and DSPs.

**DistroKid** — founded 2013. The most popular indie distributor. Flat annual fee from 19.99 USD and 100% royalty retention.

**CD Baby** — the original indie distributor. Per-song and per-album pricing. Can bundle publishing administration.

**TuneCore** — annual pricing with 100% royalty retention. Extends into publishing and sync licensing.

**Believe Music** — French, founded 2005, IPO 2021. A heavyweight in indie distribution. TuneCore sits under it. M&A discussion was a major story in 2024.

**United Masters · Amuse · LANDR · iMusician** — challengers in the same space as DistroKid and CD Baby.

**Spotify for Artists · Apple Music for Artists · YouTube Studio** — artists see their own listening data. Which cities, time zones and playlists their songs travel through becomes visible. The first input for marketing and touring.

| Distributor | Pricing | Strength |

| --- | --- | --- |

| DistroKid | from 19.99 USD/year | indie default, simple |

| CD Baby | per song / per album | publishing administration |

| TuneCore | yearly + free tier | sync, publishing |

| Believe | contract-based | global indie heavyweight |

| United Masters | free + revenue share | brand partnerships |

17. Per-play revenue and the split — the real numbers

The most frequent question in the streaming era is "how much does an artist make from one play". Estimates frequently cited between 2024 and 2026:

- Spotify: roughly 0.003 to 0.005 USD per play (varies by licensing, country, label).

- Apple Music: roughly 0.007 to 0.01 USD per play.

- YouTube Music: around 0.002 USD per play.

- Amazon Music: around 0.004 USD per play.

- Tidal HiFi+: around 0.012 to 0.013 USD per play.

- Melon: around 6 KRW per song (a commonly cited streaming-plus-download mix average).

These numbers are simplifications. Actual payouts pass through labels, publishers and distributors and split further. It is common for one million plays to leave only hundreds-to-thousands of dollars in an artist's hands. That is why the industry debate around "streaming is not enough" does not end.

18. Listener scenarios — who should pick which stack

**Scenario A: a Korean office worker, K-pop plus global pop**

- Primary: Melon or Genie (charts + Korean fan community + concert/merch bundle).

- Secondary: Spotify (global Discover Weekly + Daylist + DJ).

- Side: Apple Music (Sing + Spatial Audio + lyrics).

**Scenario B: a Japanese student, J-pop plus anime**

- Primary: AWA or LINE Music (listen-with-friends + messenger integration).

- Secondary: Spotify Japan (global discovery).

- Side: mora (high-quality downloads + limited editions).

**Scenario C: a US office worker, commute plus driving**

- Primary: Spotify Premium (Daylist + DJ + AI Playlist).

- Secondary: Apple Music (Spatial Audio + Sing).

- Side: YouTube Music (live video catalogue).

**Scenario D: a European audiophile plus indie fan**

- Primary: Tidal HiFi+ or Qobuz (lossless + Hi-Res).

- Secondary: Deezer Flow (personalised radio).

- Side: Bandcamp (direct-to-artist indie).

**Scenario E: an indie artist themselves**

- Distribution: DistroKid or TuneCore to start; Believe as you scale.

- Analytics: Spotify for Artists + Apple Music for Artists + YouTube Studio.

- Direct: Bandcamp + SoundCloud.

- Fan direct: Patreon · Buy Me a Coffee · Sound.xyz.

**Scenario F: a classical specialist**

- Primary: Apple Music Classical (correct metadata).

- Secondary: Qobuz (Hi-Res masters).

- Side: IDAGIO (classical-only streaming).

19. Where the ethics are hardest — a 2026 snapshot

Music AI is fighting on five ethical fronts at once.

**1) Training data — unlicensed recordings** — the core of the RIAA case against Suno and Udio. Ultimately a question of how licensing agreements will be struck.

**2) AI-cloned vocals** — after the Drake / Weeknd case nearly every DSP added "AI clone of a specific artist" to its takedown policy. AI songs imitating only a genre remain a grey zone.

**3) Royalty pool dilution** — when small AI-generated songs flood the catalogue, the per-play pool shrinks for real artists. Deezer announced in 2023 that it tracks AI tracks separately.

**4) Fake streams / bot listening** — combinations of AI tracks plus bot plays designed to siphon royalties have multiplied. The first US criminal indictment happened in 2024.

**5) Algorithmic bias and homogenisation** — the stronger the recommender, the narrower a listener's diet. The old critique that "Spotify buries indie" came back in the AI discovery era.

No single company can solve these five. They require negotiation between labels, DSPs, governments and artist organisations. As of 2026 none of them is cleanly resolved.

20. How "humans pick songs" survives

The stronger AI discovery becomes, the clearer the place for human selection. Two forms keep working.

**1) Editorial curation — Apple Music · Melon · NTS**

- Apple Music keeps editorial playlists to the end.

- Melon's chart and editorial curation are the K-pop industry's primary indicators.

- NTS Radio in London became the global standard for indie and electronic listeners on human DJ-led radio.

**2) Friends and community curation — Last.fm · Discord · friends' playlists**

- Last.fm, launched 2002, scrobbles listening data from any platform into a personal music diary.

- Music Discord servers, music Twitter and K-pop fandom communities remain alternative discovery channels.

- A playlist made by a friend is still the strongest recommendation that exists.

AI discovery grows the volume; human curation grows the context. Both survive.

21. Where music and AI go after 2026

Three large currents are visible.

**1) "AI makes the song + AI picks the song" fused** — Spotify and Amazon experimenting with quietly mixing AI-generated songs into AI discovery. Labelling "who made this song" becomes mandatory.

**2) The natural-language interface for listening** — Amazon Music Maestro, Spotify AI Playlist and YouTube Music Conversational Radio standardise natural language together. "Search for a song" becomes "describe a song".

**3) Reconvergence of live and streaming** — Tidal plus Square, Spotify's concert-ticket experiments, YouTube's live performance catalogue all come back. The industry consensus that "per-play revenue is not enough" pushes the answer toward live.

The "subscription + per-play payout" frame of music streaming will stay, but on top of it the layers of AI discovery, AI generation, live, and direct fan support will keep getting thicker. 2026 is the middle of that change.

22. References

- Spotify Newsroom — DJ launch announcement: https://newsroom.spotify.com/2023-02-22/spotify-debuts-a-new-ai-dj-right-in-your-pocket/

- Spotify Newsroom — AI Playlist beta: https://newsroom.spotify.com/2024-04-08/ai-playlist-beta-launch-uk-australia/

- Spotify Engineering blog: https://engineering.atspotify.com/

- Spotify RecSys research: https://research.atspotify.com/

- Apple Music: https://www.apple.com/apple-music/

- Apple Music Classical: https://www.apple.com/apple-music-classical/

- Amazon Music Maestro announcement: https://www.aboutamazon.com/news/entertainment/amazon-music-maestro-ai-playlist

- YouTube Music Help: https://support.google.com/youtubemusic/

- Tidal HiFi Plus: https://tidal.com/

- SoundCloud Newsroom: https://newsroom.soundcloud.com/

- Deezer Flow: https://www.deezer.com/

- Pandora Music Genome Project: https://www.pandora.com/about/mgp

- Qobuz: https://www.qobuz.com/

- Bandcamp: https://bandcamp.com/

- Melon: https://www.melon.com/

- Genie: https://www.genie.co.kr/

- Bugs: https://music.bugs.co.kr/

- FLO: https://www.music-flo.com/

- AWA: https://awa.fm/

- LINE Music: https://music.line.me/

- mora: https://mora.jp/

- Shazam: https://www.shazam.com/

- MusicBrainz: https://musicbrainz.org/

- ACRCloud: https://www.acrcloud.com/

- AudD: https://audd.io/

- Suno AI: https://suno.com/

- Udio: https://www.udio.com/

- Stable Audio: https://stability.ai/stable-audio

- RIAA v Suno/Udio coverage (Reuters): https://www.reuters.com/legal/major-music-labels-sue-ai-firms-suno-udio-2024-06-24/

- Audius: https://audius.co/

- Royal: https://royal.io/

- Sound.xyz: https://www.sound.xyz/

- DistroKid: https://distrokid.com/

- CD Baby: https://cdbaby.com/

- TuneCore: https://www.tunecore.com/

- Believe Music: https://www.believe.com/

- Spotify for Artists: https://artists.spotify.com/

- Apple Music for Artists: https://artists.apple.com/

- YouTube Studio for music: https://studio.youtube.com/