Industries / Media, arts and recreation / Media and entertainment
Music
Music workers record, produce and publish songs, and collect the royalties owed each time a recording or a song is sold, streamed or played.
7
Milestones achieved, from 1860 to 2021.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1860
Sound written down
On 9 April 1860, Édouard-Léon Scott de Martinville's phonautograph traced a voice singing "Au Clair de la Lune" as a wavy line. He did not intend his tracings to be played back. In 2008 the research group First Sounds played them using a "virtual stylus" method from Lawrence Berkeley National Laboratory, and First Sounds calls the song "the earliest clearly recognizable record of the human voice yet recovered."
LessonA recording made only to be looked at could still be heard 148 years later, once scanning tools existed.
Credit
Édouard-Léon Scott de Martinville; First Sounds; Lawrence Berkeley National Laboratory
1877
Recorded and played back
Thomas Edison's team built a phonograph that recorded on tinfoil wrapped around a turning cylinder, made by the machinist John Kruesi, and showed it to Scientific American on 7 December 1877. The US National Park Service calls it "the first machine to both record sounds and play them back."
Credit
Thomas Edison, Charles Batchelor and John Kruesi
1925
Electrical recording
Until the mid-1920s the phonograph was "an acoustic/mechanical device," in the words of an IEEE history. A Western Electric team led by Joseph Maxfield built a system for electrical recording, AT&T licensed it to the Victor Talking Machine Company, and Victor began issuing electrically recorded "Orthophonic" records in 1925. Leopold Stokowski and the Philadelphia Orchestra were among the first to record that way.
Credit
Joseph Maxfield and colleagues, Western Electric; Victor Talking Machine Company
1957
One part at a time
From about 1949 the guitarist Les Paul layered parts by re-recording over his own tapes, which he called sound on sound, but each pass added noise and one bad pass ruined the take. Ampex's Ross Snyder conceived Sel-Sync, which let the recording head play back earlier tracks in time while a new part went onto another track, and Ampex sold the first eight-track machine to Paul. Recording each part separately, called overdubbing, became normal studio practice; four-track machines were the studio standard for most of the later 1960s. Sources differ on the year of the sale: the museum's multitrack history gives 1957, and a reprinted Mix magazine account gives 1955.
LessonSnyder expected a market of "a few dozen at the absolute most," and Ampex never patented the idea.
museumofmagneticsoundrecording.orgmuseumofmagneticsoundrecording.org
Credit
Les Paul; Ross Snyder and Mort Fujii, Ampex
1983
Instruments that share notes
MIDI (Musical Instrument Digital Interface) is a shared language that lets keyboards, drum machines and computers from different makers send notes and controls to each other. Dave Smith of Sequential Circuits presented the idea as the "Universal Synthesizer Interface" at the 1981 Audio Engineering Society convention, and MIDI was demonstrated at the 1983 NAMM trade show. By the MIDI Association's account, five companies did the early work: Sequential Circuits, Kawai, Korg, Roland and Yamaha, and Roland's Ikutaro Kakehashi was "the driving force behind MIDI."
Credit
Dave Smith; Ikutaro Kakehashi; Sequential Circuits, Kawai, Korg, Roland and Yamaha
2001
Napster held liable
Napster let people swap music files with each other over the internet. On 12 February 2001 the US Court of Appeals for the Ninth Circuit held that Napster was likely liable for its users' copyright infringement, because it knew of specific infringing files and could have blocked them but did not. The court also found the lower court's injunction too broad and required record companies to name the specific works before Napster had to block them.
LessonThe court held that a system able to carry legal uses was not liable just for existing; liability came from knowing about specific files and failing to act.
Credit
A&M Records and other record companies; US Court of Appeals for the Ninth Circuit
2021
A collective for streaming royalties
The US Music Modernization Act of 2018 created a blanket license for streaming services, run by a new body, The Mechanical Licensing Collective (The MLC). In February 2021, 20 streaming services handed The MLC $424,384,787 in past royalties they had not matched to the songwriters owed. In April 2021 The MLC made its first monthly payment, $24 million, and said it had matched about 80% of reported royalties to registered songs.
LessonRoyalties go unpaid when nobody can link a recording to the song behind it; the law gives The MLC two years to match past money before paying the rest out by market share.
Credit
US Congress; The MLC
Watch for
Licensed AI music services
Universal Music Group (UMG) and the AI song generator Udio settled UMG's copyright lawsuit and announced a subscription platform, to launch in 2026, that "will be trained on authorized and licensed music," with new revenue for UMG artists and songwriters.
Achieved whenThe platform is open to the public and an outside party, such as an auditor, a court filing or a news report with figures, confirms that the artists whose music trained it are being paid.
Claimed by
Universal Music Group and Udio, October 2025
Spotting fully AI-made tracks
Deezer, a music streaming service, says fully AI-generated tracks passed half of its daily uploads on peak days in June 2026, nearly 90,000 a day, while drawing 1% to 3% of streams. It says its detector is 99.8% accurate, that up to 85% of streams on such tracks in 2025 were fraudulent, and that Billboard uses the tool to find AI tracks on its charts. An October 2026 academic benchmark, MUSEBENCH, found that a detection method Deezer researchers published, which the authors say Deezer deployed, scored 97.6% balanced accuracy on its test but missed 93.6% of AI tracks once their pitch or tempo was changed. It did not test Deezer's production tool.
Achieved whenAn outside party publishes a test of the production detector on tracks Deezer did not choose.
Claimed by
Deezer, July 2026
Open
01
Compare the song-ownership databases
A public measurement of how often the US databases of who owns a song disagree. Take a fixed list of popular songs, look each one up in Songview, the joint ASCAP and BMI database, and in The MLC's public database, and record each writer's and publisher's share. Flag songs where the shares differ between databases or add up to more than 100%.
The briefCredit · unclaimed02
Keep the detector test current
A public test of AI-music detectors, rerun within a month of each new release of a major song generator. Each round makes 500 new tracks with the new version, under the generator's terms of service, pairs them with 500 human-made tracks under Creative Commons licenses, adds copies that are compressed to MP3 or shifted in pitch and tempo, and runs every detector that accepts test files. It publishes each detector's misses and false alarms.
The briefCredit · unclaimed
Open 01 · the brief
Compare the song-ownership databases
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public measurement of how often the US databases of who owns a song disagree. Take a fixed list of popular songs, look each one up in Songview, the joint ASCAP and BMI database, and in The MLC's public database, and record each writer's and publisher's share. Flag songs where the shares differ between databases or add up to more than 100%.
Why it is still open
ASCAP and BMI launched Songview in December 2020 and mark a song with a green check only when both agree on its owners, but Songview covers only those two organizations' repertoire. The MLC's own policy says it "does not judge or resolve" ownership conflicts such as "overclaims, where Claims to a Work exceed 100% of Shares," and holds the royalties until the claimants settle. This research found no published study comparing the two sets of records..pdf)
Built with what exists
The public Songview search on the ASCAP and BMI websites; The MLC's public work search; ISWC codes, the standard ID numbers for songs, to join records; a web agent for sites with no programming interface, run slowly and within each site's terms of use (see the Note "Deleting 80% to start over"); Python and a shared spreadsheet.
First step
Look up 50 songs from one year-end chart by hand in both databases, to learn the fields and see how often they differ.
Done looks like
1,000 songs, for example ten years of Billboard's year-end Hot 100, looked up in both sources, with the share of songs that disagree and the share that add up to more than 100% published along with every record; a hand check of 100 songs finds the tool's readings wrong in fewer than 2%.
Open 02 · the brief
Keep the detector test current
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI-music detectors, rerun within a month of each new release of a major song generator. Each round makes 500 new tracks with the new version, under the generator's terms of service, pairs them with 500 human-made tracks under Creative Commons licenses, adds copies that are compressed to MP3 or shifted in pitch and tempo, and runs every detector that accepts test files. It publishes each detector's misses and false alarms.
Why it is still open
Deezer's 99.8% figure is its own, and Billboard relies on its tool. MUSEBENCH, the largest outside test found (October 2026), checked 11 detectors but reused Suno and Udio tracks from SONICS, a test set published in 2024, so it says nothing about later versions. ArtifactBench, a benchmark of 22 generators on Hugging Face, is published by Intrect, a company that sells a detector, which ranks its own detector first; its test set is frozen in numbered versions rather than rerun when a generator changes. This research found no neutral test that is rerun as generators change.
Built with what exists
The MUSEBENCH paper's method and the open SONICS detector models; Creative Commons tracks from the Free Music Archive; paid accounts on current song generators; ffmpeg to compress files; an audio library such as librosa to shift pitch and tempo; Python.
First step
Run two open detectors on 100 tracks made with the current version of one generator and 100 Creative Commons tracks, and compare the results with MUSEBENCH's figures for the older version.
Done looks like
For one year, a round published within 30 days of each major generator release, each with at least 500 AI-made and 500 human tracks, and each detector's miss rate and false-alarm rate given with 95% confidence intervals.