Retrospective field notes / 100 notes
Tool releases, licensing schemes, platform policies, detection research and the older precedents behind AI music tools, each read for the audio task and the rights status the documents state.
Historical event dates and source dates are separate from the preparation date of this local edition. Every note is a retrospective draft prepared 16 September 2026; none was published on its historical date.
100 notes
A museum record of Bell Labs' 1961 speech-synthesis demo shows what it proved, decades before voice-consent debates began.
WIPO's own treaty text gives performers moral and making-available rights, decades before synthetic voices existed.
The Copyright Office's own summary shows what the DMCA's safe harbors cover, and what they never say about AI.
Grand Upright v. Warner Bros. shows a court granting an injunction and referring the case for prosecution.
Bridgeport v. Dimension Films set a bright-line clearance rule that other circuits have not all adopted.
A Supreme Court opinion and ASCAP's own page describe why songwriters organized to collect performance royalties.
Yamaha's own design account shows how the DX7 declared a digital era through hardware, not an algorithm alone.
Fraunhofer and ISO's own records date mp3's standardization, naming and licensing history, and its audible trade-off.
Pandora's own pages describe musicologists tagging songs by hand, a method distinct from generative AI tools.
A federal judge held MP3.com's full-album copying infringed, even though subscribers needed to own the CD.
Waves' product page describes Clarity Vx as neural-network vocal noise removal, dated 7 March 2022, without addressing rights to the recording.
sonible's page frames smart:comp's AI analysis as generating a range of settings a user still auditions and edits by hand.
TONEX's page claims highly accurate amp captures but publishes no benchmark, and its terms put upload rights on the user.
Antares' current page bundles AI-labelled vocal tools into a subscription and does not state what happens to sessions after a lapse.
Steinberg's own new-features page for Cubase 15 lists AI-powered stem separation as a Pro-tier addition, undated on the page itself.
Avid's living what's-new page documents AI-labelled features like Speech-to-Text without branding Pro Tools around AI.
Image-Line's FL Studio 2026 page ties AI stem separation to paid editions and does not grant a licence to redistribute results.
BandLab's help article states a user does not own SongStarter's generated track but can use it without fees or royalties.
Spotify's engineering blog documents Basic Pitch as an open-source, instrument-agnostic audio-to-MIDI model, released June 2022.
Arturia's product pages describe Augmented instruments as sampled performances blended with synthesis, not an AI system.
Dreamtonics' own product page says every Synthesizer V voice is a licensed, consented recording, not a cloned or scraped one.
ACE Studio's own site labels its 160-plus AI voices royalty-free but does not spell out restrictions on the pricing page.
Voicemod's terms condition its voice-creation tools on having the source speaker's consent, separate from its AI voice-actor library.
Respeecher's own site frames a signed consent agreement with each voice donor as core to its voice-conversion service.
The companies' own 2021 releases describe a PE-backed alliance with unified leadership, not one firm buying the other outright.
Musiio's own announcement says SoundCloud acquired its AI tagging and search technology to improve discovery, not payouts.
Shutterstock's 2020 release says it acquired certain Amper Music assets for search, discovery, and editing tools, price undisclosed.
Reuters reported Meta was close to buying Accusonus in 2022, but Accusonus's own site never named an acquirer before sunsetting ERA.
Positive Grid's Spark amp pages describe Smart Jam as a real-time practice tool, without addressing rights to the generated backing.
RX 12's Music Rebalance and Dialogue Isolate estimate mix elements for restoration, and iZotope frames the result as an editing aid, not a recovered multitrack.
XO clusters a producer's one-shots by sonic similarity, but XLN Audio's terms leave any imported sample's rights with the customer.
In 2022 Serato added real-time vocal, melody, bass, and drum isolation to DJ Pro, without changing rights in the recording being separated.
In 2020, djay's Neural Mix paired Algoriddim's real-time engine with AudioShake's separation technology, running on-device via Apple's chip.
AudioShake's own site frames stem separation as a licensed B2B service for labels, publishers, and platforms, not a direct-to-consumer tool.
Loopcloud's Jamahook-powered similarity search surfaces matching samples, and Loopmasters' 2019 licence lets them join a composition but bars standalone reuse.
Output's FAQ states a track made in Arcade while subscribed stays royalty-free after cancellation, while the sound library itself stays licensed, not owned.
LANDR's terms grant a perpetual licence to a downloaded Sample once incorporated into a Recording, a separate product from its mastering service.
EZdrummer 3's Tap2Find searches a MIDI library for the closest match to a tapped rhythm, and Toontrack's EULA treats the result as licensed library content.
Brusfri is a single-purpose machine-learning noise reducer for field recordings, and Klevgrand's terms confirm a user's own finished project stays releasable.
CISAC's July 2023 statement asks governments to require licensing, credit and transparency for AI systems trained on members' works.
PRS for Music's principles say authorisation must stay with rightsholders and AI content must be labelled as such.
GESAC, of which SACEM is a member, asked EU legislators for AI record-keeping and transparency it called necessary for trustworthy AI.
ASCAP's own AI resource sets six principles and says fully AI-generated works cannot be registered with the society today.
BMI's AI guidance defines four categories of AI involvement in a song and says only two remain eligible for registration and pay.
Sweden's STIM launched a collective AI licence covering only consenting members' works, ahead of any EU-wide AI music rule.
The Recording Academy's 2023 rule update says only human creators can be nominated, even when a submission uses AI content.
The UK's Ivors Academy campaigns for AI consent and transparency, and gives songwriters a template letter refusing AI training use.
A coalition including the Featured Artists Coalition sets three non-negotiables for AI music deals: consent, control and pay.
NMPA's 2023 Copyright Office submission argues AI training on song catalogues fails every fair-use factor and needs a licence.
Warner's announcement shows KLAY signed distinct deals with WMG, UMG and Sony entities, not one shared industry licence.
YouTube's own posts show named artists opted into specific AI experiments individually, not into a blanket catalogue licence.
AFM's own 2023 statements argue replacement remuneration cannot substitute for musicians' jobs disappearing to AI.
Rightsify's own pages show Hydra's consumer generator was discontinued in 2024, replaced by the Gramosynth training-data pipeline.
SME's own opt-out terms and Senate testimony state a licensing-first policy but do not describe any specific AI agreement's scope.
UMG's own 2024 and 2026 announcements restate the same responsible-AI, human-artistry and shared-value language for two different partners.
Tracklib's own pages state a cleared sample requires registering the release on its Licenses page, not just holding a subscription.
SAG-AFTRA's own statements require union bargaining and consent before a performer's voice can be used in AI simulation.
MMF's own materials describe an advisory guide and template letter for managers, not a settled industry AI contract standard.
Spotify's newsroom post explains what its AI Persona badge covers, how it is applied, and what it leaves undecided.
Beatport's published stance rejects wholly or substantially AI-made music while permitting assistive AI processing.
Apple's artist support pages cover metadata and distributor choice but do not mention AI-made or AI-assisted music.
TikTok's synthetic media policy and its 2023 tool require creators to disclose realistic AI-made audio, not just video.
DistroKid's own article sets four upload conditions and warns stores can still reject an accepted release.
SoundCloud's terms state uploads will not train voice- or music-replicating AI models without an affirmative opt-in.
Meta detects AI images automatically but asks users to disclose realistic AI-made audio and video themselves.
Amazon's general terms restrict feeding catalogue content into AI, but no artist-facing AI-upload policy was found.
YouTube's altered-content toggle covers AI music explicitly and can lead to a label, removal, or suspension.
One CD Baby help page bars AI-generated content outright while a Downtown Music policy allows some licensed AI use.
A peer-reviewed cross-over experiment found identical piano performances rated less likeable and expressive when attributed to an AI.
An unreviewed SSRN paper reports listeners called AI-labelled music less expressive, though equally enjoyable, than identical human-labelled audio.
A Frontiers in Education study of 600 Chinese music students found AI features rated useful yet unable to replace human teaching.
A CHI 2025 study of the Amuse songwriting assistant found ten songwriters felt more guided and no less creative with AI-suggested chords.
An unreviewed 2026 preprint reports Suno collapses genre distinctions while Lyria 3 narrows diversity within a genre.
An AI & Society ethnography of Melbourne music technologists found many prize mistakes and uncertainty that generative tools are designed to avoid.
Deezer's newsroom reports 97% of an eight-market survey could not tell AI-generated music from human-made tracks in a listening test.
Deezer's own reporting describes a patent-pending detector that flagged 13.4 million AI-generated tracks in 2025 by artifacts in the audio.
Essentia, maintained by Universitat Pompeu Fabra's Music Technology Group, is dual-licensed audio-analysis software used across MIR research.
The Free Music Archive dataset packages over 100,000 Creative-Commons-licensed tracks as a rights-cleared alternative to scraped training audio.
Waves, iZotope, and Native Instruments define commercial use and AI training differently in their own licence text.
Waves, Antares, and iZotope's own terms show what each licence model does when payment or verification stops.
SOUNDRAW, Loudly, and LANDR each define commercial use differently across their own licence pages.
Algoriddim's Neural Mix is marketed as on-device, but its privacy policy does not extend that promise to audio.
Waves, Antares, and Celemony each govern one stage of a vocal chain under separate, differently dated terms.
LANDR, Output, and Loopcloud each forbid standalone reuse of their sounds, but differ on what survives cancellation.
IK Multimedia, sonible, and XLN Audio's own terms show what a plugin's paperwork does and does not say about AI.
LANDR names a specific AI-training purpose and a 24-month account clock; Image-Line states only a general principle.
Tracklib licenses a known recording; AIVA's tiered terms assign copyright the Copyright Office has not resolved.
Spotify names -14 LUFS directly; LANDR and eMastered treat that number as a reference, not a rule.
Serato, AudioShake, and iZotope describe their stem separators for different audiences, and only one names a rights caveat outright.
sonible, LANDR, and eMastered price their AI-assisted tools as a one-time purchase, a subscription, and a flat per-plan fee.
DistroKid's own upload rules and the Copyright Office's disclosure guidance both reward a producer who already knows which tool touched what.
The Copyright Office's disclosure duty and Fairly Trained's certified-models list both give a producer language for asking a collaborator what was used.
MusicGen's own weights carry a noncommercial licence, while Google's Gemini API bills per song and treats free and paid use differently.
MusicGen and Google's Lyria 3.5 both publish model cards naming training data and limits; many commercial plugins publish neither.
The Copyright Office's disclosure duty and iZotope's restoration workflow both assume a producer can still find the paperwork years later.
Stability AI's licence page and Meta's MusicGen model card describe their models' data very differently, and a producer weighs how much that matters.
Native Instruments' and Avid's own published terms show how differently vendors address licence type and AI training on their content.
DistroKid conditions uploads on rights ownership while TuneCore conditions them on whether a GenAI tool's training data was licensed.
WavTool's browser DAW marketed generated audio as unlicensed and royalty-free before the whole product went offline in 2026.
Downtown's 2025 content rules bar unlicensed and mimicking AI content, while CD Baby's older FAQ excludes AI content outright.
Try another category or clear your search.
