The price of AI training data: whether court-supervised settlement becomes a recurring licence or a one-off

Six alternative branches. These are subjective probabilities from the seed, not observed frequencies or investment recommendations.

Why this trend

A number now exists where previously there was only argument, and it is being used as the anchor in every subsequent negotiation and complaint. The structure is deliberately written to work whether this sector stays merged or is split: the propositions and indicators name book publishing, news publishing and music separately, because the corpus's own finding is that book publishing has already been paid a realised cash sum while music holds an unpriced option with no disclosed terms in any of six announced licences.

Load-bearing assumption

The $1.5bn Bartz settlement priced the piracy of acquisition rather than the act of training, so its per-work benchmark of roughly $3,000 survives only if no court holds that training on a lawfully acquired corpus is fair use.

Preconditions

Not recorded.

base · 36% · band 28-44%

Licensed coexistence. The OpenAI MDL produces a mixed ruling or settles; bilateral music licences extend to Sony and to independents; generative audio settles as a professional creation tool rather than a consumer substitute. Music grows mid-single-digits on price and emerging-market volume, books grow low-single-digits with digital audio outperforming, and local news continues to consolidate and shrink. AI licensing becomes a real but modest revenue line, in low single-digit percentages of major-label revenue.

Mechanism

The industry's institutional reflex is licensing, not enforcement: Napster was shut down and changed nothing, iTunes and Spotify licensed the behaviour and changed everything, and the people running the current negotiations lived through that. A mixed ruling preserves enough leverage for bilateral deals to continue without producing a clearing price, so terms stay confidential and revenue stays small and unnamed. Warner's $16m of settlement revenue on $1,864m - 0.9% - is the scale this branch implies.

Preconditions, early indicators and assumptions

Calibration Basis

analogue

Early Indicators

Indicator
A major label disclosing AI licensing revenue as a named line item above 2% of quarterly revenue
Source
S-24-07
Would Be Visible By
2028-03
Indicator
Sony settling with or licensing to Suno or Udio, or 1:26-cv-06120 terminating
Source
S-24-06
Would Be Visible By
2027-12
Indicator
AAP StatShot annual US publishing revenue growth between 1% and 5%
Source
S-24-03
Would Be Visible By
2027-09

Affected Industries

What Businesses Should Do

Rightsholders: register copyrights on time. The Authors Guild survey found 93% of the 864 books with compliance data showed publisher failure to register, and registration is what makes statutory damages - and therefore any claim - economically viable. Developers: provenance of the corpus is now the dominant legal risk, not the act of training.

Precedence Note

Distinguished from downside by the content of the ruling, not by its timing. A ruling that denies summary judgment on any category is base or better; a clean fair-use grant for lawfully acquired corpora is downside.

Would Change Our Mind

Two rightsholders disclosing recurring AI licensing revenue in consecutive quarters, which would move mass to upside.

Load Bearing Assumption

Text: The MDL produces a mixed or partial ruling rather than a clean holding either way.
Confidence: medium
Load Bearing: true
If Wrong: A clean holding in either direction empties this branch: for developers it moves mass to downside, for rightsholders to upside.

upside · 12% · band 6-20%

AI licensing matures into a third revenue pillar alongside streaming and sync. Terms become semi-public through filings, per-work and per-catalogue rates emerge, the publishing side converts its settlement precedent into recurring licences rather than one-off damages, and detection and provenance become standard infrastructure sold to collecting societies globally.

Mechanism

The constraint that releases is confidentiality. Bilateral secrecy persists only while rightsholders believe individual negotiation beats a published rate; once a rate is disclosed in a filing or a litigation exhibit, the benchmark becomes public and smaller rightsholders and collecting societies can transact against it. Deezer's licensing of its AI detector to Hungarian and Dutch collecting societies is the early form of the infrastructure leg.

Preconditions, early indicators and assumptions

Calibration Basis

judgement_only

Early Indicators

Indicator
Two or more rightsholders disclosing recurring - not settlement - AI licensing revenue in consecutive quarters
Source
S-24-07
Would Be Visible By
2029-06
Indicator
A collecting society publishing an AI training tariff with rates attached
Source
S-24-20
Would Be Visible By
2029-12
Indicator
Financial terms of any major-label AI licence surfacing in a filing or court exhibit
Source
S-24-06
Would Be Visible By
2028-12

Affected Industries

What Businesses Should Do

Build the rights-metadata and registration infrastructure now. In this branch the binding constraint on being paid is not the licence, it is whether you can prove which works you control.

Precedence Note

Requires recurring revenue, disclosed. Settlement proceeds, however large, belong in base - the corpus is explicit that Warner's $16m is a settlement and not a licence.

Would Change Our Mind

IFPI and the majors continuing to oppose any published tariff for another two years.

Load Bearing Assumption

Text: Rightsholders come to prefer a public benchmark to bilateral leverage.
Confidence: low
Load Bearing: true
If Wrong: If bilateral leverage keeps paying the majors more, terms stay confidential, no benchmark forms, and this mass returns to base.

downside · 24% · band 16-32%

Fair use wins and the leverage evaporates. Judge Stein grants summary judgment to OpenAI and Microsoft on fair use for lawfully acquired corpora. Bartz is re-read as a piracy-specific outcome with no licensing implication, the bilateral music deals are renegotiated downward or abandoned, catalogue re-rates lower on expected long-tail dilution, and the compensation thesis on which publishers are currently planning collapses.

Mechanism

Every current negotiating position rests on the unresolved question of whether training is fair use. A holding that it is, for lawfully acquired material, removes the threat that makes bilateral licences worth signing - at which point the rational developer strategy is to buy content once, legally, and train freely. The corpus records the specific inversion this would cause: the same trade bodies now opposing a statutory remuneration right would immediately demand one, because bilateral leverage would be gone.

Preconditions, early indicators and assumptions

Calibration Basis

judgement_only

Early Indicators

Indicator
A summary-judgment ruling in MDL 25-md-3143 granting the fair-use defence across asserted categories
Source
S-24-06
Would Be Visible By
2027-12
Indicator
IFPI or a major label publicly reversing its opposition to statutory remuneration or collective management
Source
S-24-20
Would Be Visible By
2028-12
Indicator
A music catalogue transaction printing below 13x, against Reservoir's disclosed 15.5x
Source
S-24-14
Would Be Visible By
2028-12

Affected Industries

What Businesses Should Do

Do not capitalise expected AI licensing revenue in a catalogue valuation. It is an unpriced option on an unresolved legal question, and the corpus's own finding is that no music licence anywhere has disclosed a rate.

Precedence Note

Requires the ruling itself. Slow licensing progress without a ruling is base.

Would Change Our Mind

A ruling that treats lawful acquisition as insufficient on its own, which would move mass back to base and upside.

Load Bearing Assumption

Text: A fair-use holding for lawfully acquired corpora would be read as controlling across both the book and music halves of this sector.
Confidence: medium
Load Bearing: true
If Wrong: If the halves diverge - fair use for text, not for voice and recordings - music keeps its leverage and half this mass returns to base.

disruption · 10% · band 4-16%

Discovery moves inside the model. The referral collapse Pew measured for news - clicks falling from 15% to 8% when an AI summary appears, source clicks inside the summary at 1%, sessions ending on 26% of visits against 16% - extends to music and books. People ask an assistant what to listen to or read and never reach a catalogue surface, and whoever controls the assistant controls discovery.

Mechanism

The substitute is the assistant as the discovery layer, and the cost delta that lets it win is the user's time rather than money. The incumbent asset that loses value is the catalogue surface itself - Spotify's browse and search, the retailer's storefront, the publisher's homepage. The corpus's own generalisation is that this sector absorbs supply-side shocks and does not absorb demand-side ones: home taping, file-sharing and self-publishing all multiplied supply without destroying the industry, while iTunes unbundling the album and AI search removing the click changed distribution and did damage.

Preconditions, early indicators and assumptions

Calibration Basis

analogue

Early Indicators

Indicator
A DSP reporting a decline in search-and-browse-originated listening share
Source
S-24-08
Would Be Visible By
2029-06
Indicator
A major retailer reporting assistant-originated book sales as a distinct channel
Source
S-24-03
Would Be Visible By
2029-12
Indicator
News publishers disclosing referral-traffic declines in SEC filings rather than in advocacy statements - USA TODAY Co's Q1 2026 10-Q disclosed nothing at all about AI search
Source
S-24-16
Would Be Visible By
2028-12

Affected Industries

What Businesses Should Do

Build direct relationships and first-party data now, because in this branch the intermediary you are optimising for disappears. Local news has the least ability to do this and is where the damage concentrates.

Precedence Note

Requires a disclosed channel-share decline, not commentary. Publisher complaints without a series stay out of this branch.

Would Change Our Mind

Two years of DSP disclosure showing search-and-browse share stable.

Load Bearing Assumption

Text: Assistant-mediated discovery scales to music and books the way it has to news search.
Confidence: low
Load Bearing: true
If Wrong: If music and book discovery are habitual rather than search-mediated, this branch does not generalise and its mass returns to base.

regulatory · 10% · band 4-18%

The EU legislates against the incumbents' wishes. Despite IFPI's explicit September 2026 opposition, the Commission proposes a statutory remuneration right or extended collective licensing for AI training, and the UK finally concludes with a text-and-data-mining exception plus rights reservation. Independent artists and mid-list authors get paid, major-label bilateral leverage is diluted, and compliance costs rise for every developer.

Mechanism

The instrument would be a Commission legislative proposal - not a consultation and not a code of practice. The current settlement is transparency-by-disclosure under AI Act Article 53(1)(d) rather than a remuneration right, and the US Copyright Office examined compulsory and collective licensing in its Part 3 pre-publication report and recommended neither. The direction is two-sided: a TDM exception with rights reservation would liberalise rather than restrict, and both are regulatory outcomes.

Preconditions, early indicators and assumptions

Calibration Basis

base_rate

Early Indicators

Indicator
A European Commission legislative proposal naming remuneration or collective management for AI training
Source
S-24-24
Would Be Visible By
2029-06
Indicator
A UK government response concluding the copyright-and-AI consultation that closed 2025-02-25
Source
S-24-25
Would Be Visible By
2028-12
Indicator
Publication of the EU GPAI Code of Practice copyright-chapter signatory list
Source
S-24-24
Would Be Visible By
2027-12

Affected Industries

What Businesses Should Do

Track the instrument type, not the announcement. Consultations, codes and pre-publication reports have produced nothing binding in this sector for over a year; only a legislative proposal changes the economics.

Precedence Note

Regulatory outranks base when a legislative proposal issues, in either direction. A code of practice or a consultation does not qualify.

Would Change Our Mind

A Commission work programme that omits AI and copyright entirely.

Load Bearing Assumption

Text: The EU legislates on this file against the stated preference of its largest rightsholders.
Confidence: low
Load Bearing: true
If Wrong: If the Commission defers to IFPI's position, this branch is near-remote and its mass returns to base.

failure · 8% · band 3-15%

The fraud economy wins. Detection does not keep pace, AI catalogue floods every platform, royalty pools are diluted faster than fraud can be demonetised, identity verification fails, and the metadata layer becomes unreliable enough that rights cannot be matched to payments at all. In this world the upload flood matters far more than the streaming-share statistic allows.

Mechanism

Deezer receives roughly 90,000 fully AI-generated tracks a day, exceeded 50% of daily uploads at peak in June 2026, and identified up to 85% of streams on those tracks as fraudulent against platform-wide fraud of 8%. The thesis fails not because AI music wins listeners but because the cost of manufacturing plausible catalogue collapses faster than the cost of policing it rises. Whoever is left holding the asset is the collecting society with an unmatched royalty pool, and unmatched pools historically take years to reconcile.

Preconditions, early indicators and assumptions

Calibration Basis

judgement_only

Early Indicators

Indicator
A second platform publishing AI-upload and fraud statistics materially worse than Deezer's
Source
S-24-10
Would Be Visible By
2028-12
Indicator
A collecting society reporting a rising unmatched or black-box share attributable to synthetic catalogue
Source
S-24-20
Would Be Visible By
2029-06
Indicator
Deezer's reported fraudulent share of AI-track streams rising above 90% from the 85% recorded for 2025
Source
S-24-10
Would Be Visible By
2028-06

Affected Industries

What Businesses Should Do

Treat catalogue integrity and metadata as an operating cost line, not a compliance afterthought. In this branch the platforms that cannot prove what they are paying for lose the ability to pay anyone correctly.

Precedence Note

Failure outranks all other branches on occurrence. It is distinguished from disruption by where the damage lands: disruption removes discovery, failure removes the ability to match rights to payments.

Would Change Our Mind

Two platforms publishing declining AI-fraud shares after deploying detection at scale.

Load Bearing Assumption

Text: Detection cost rises more slowly than the cost of manufacturing synthetic catalogue falls.
Confidence: low
Load Bearing: true
If Wrong: If detection scales cheaply - Deezer claims 99.8% accuracy and is already licensing the detector - this branch is contained and its mass returns to base.

Additional scenario notes

Probabilities Sum

What Must Be True To Grow

Capability: settled on both sides - models train, and detection runs at a claimed 99.8% accuracy at one platform.
Economics: asymmetric. Book publishing has been paid $1.5bn, roughly $3,000 per work; music has been paid nothing disclosed in any of six announced licences.
Supply: inverted - the constraint is excess supply, at roughly 90,000 AI tracks a day on one platform.
Demand: fully AI-generated music is 1-3% of streams on the only platform publishing the series, and up to 85% of those streams were fraudulent.
Permission: the failing gate. No court has ruled on training-as-fair-use for a lawfully acquired corpus; the US Copyright Office recommended no legislation; the UK has produced nothing in over a year.
Capital: the counterparties can pay - OpenAI and Anthropic took 43% of H1 2026 global VC - which is why settlement is available as a resolution mechanism at all.

What Could Stop It

Regulatory reversal in the form of a fair-use holding, which would remove the basis for the entire compensation thesis
Measurement revision - every AI-music figure in the corpus traces to one platform that sells the detection product its numbers justify
Demand deflation - the compensation thesis depends on five companies remaining solvent and willing to settle, in an environment the FOMC has flagged for AI-firm valuations and leveraged financing
Attention withdrawal - a copyright formality, not a technology, decided who got paid, and 93% of surveyed books failed it

Uncertain Assumptions

That the Bartz per-work figure is a benchmark at all, when it priced piracy of acquisition rather than training
That the Authors Guild's 93% registration-failure finding generalises - it is a self-selected advocacy-body sample with no independent replication, though no publisher-side data exists either
That Warner is representative of major-label economics, when it is the only one of three majors publishing auditable quarterly segment figures
That litigation facts are accurately characterised here at all: the corpus records that CourtListener and RECAP were robots-blocked, so no complaint, order or settlement agreement in this sector was read in full

Authored

2026-09-15