3AUG

Judge probes $1.5bn Anthropic class deal

4 min read

17:21UTC

Judge Araceli Martínez-Olguín took the $1.5bn Bartz class settlement under submission with a 92.77% claims rate, then ordered fresh briefing on late opt-outs before approving anything.

← Back to AI marking code gets 190 names, no broadcasters Jump to analysis ↓

IndustryDeveloping

Key takeaway

The Bartz outcome becomes the reference price for every publisher-AI licence that follows.

At the 14 May 2026 fairness hearing in Bartz v Anthropic, Judge Araceli Martínez-Olguín took the $1.5bn class settlement under submission with a claims rate of 92.77%, covering 447,576 of 482,460 works ¹. On around 18 May the judge ordered supplemental briefing, due 21 May, on whether unexcused late opt-outs should be honoured. The fee petition stands at $187.5m, and final approval has not been entered.

Bartz v Anthropic is a copyright class action over authors' works used to train Anthropic's AI models, heard in the Northern District of California. A fairness hearing is the stage at which a judge decides whether a proposed class settlement is fair to the people it binds. "Taking it under submission" means the judge has reserved judgement rather than approving on the day. The late-opt-out question matters because every author who exits the class keeps the right to sue Anthropic separately, which changes both the defendant's exposure and the value of the deal on the table.

The settlement's headline figure tracks a number already circulating in the publishing industry. News Corp chief executive Robert Thomson read an anticipated $1.5bn Anthropic figure into his Q3 earnings transcript . The class-action result functions as a reference price: it sets the floor beneath which no publisher will license content to an AI developer, because litigation now offers a quantified alternative. A clean approval hardens that floor; a reopening on opt-outs leaves it provisional, and every bilateral negotiation waits on which way Judge Martínez-Olguín rules.

Deep Analysis

In plain English

Anthropic makes Claude, an AI assistant that was trained on large amounts of text from the internet and books. A group of authors sued Anthropic, arguing their books were used to train Claude without permission or payment. This type of lawsuit is called a class action: one lawsuit on behalf of many people with the same claim. Anthropic agreed to pay $1.5 billion (about £1.2 billion) to settle the case. That money would be split among roughly 480,000 books and other works whose authors registered claims. The lawyers who brought the case want $187.5 million of that for themselves. The judge has not yet approved the settlement. She is asking additional questions about authors who tried to opt out of the settlement after the deadline. Until she gives final approval, the money has not moved and the case is not closed. For media and publishing companies, this matters because the per-work settlement value, roughly $3,000 per work if the total were split evenly, will become the reference point that AI companies use when negotiating direct licensing deals.

What could happen next?

Precedent
A final-approved $1.5bn Bartz settlement establishes a per-work floor of approximately $2,935 net that other AI labs will cite as the class-action alternative in bilateral licensing negotiations.
Short term · Assessed
Risk
If the judge's late-opt-out inquiry leads to a revised class definition, the settlement could be restructured or rejected, resetting the pricing benchmark downward or extending litigation uncertainty.
Short term · Assessed
Consequence
The $187.5m fee petition signals that plaintiff litigation capital is now treating AI copyright class actions as viable commercial investments, accelerating similar actions against other AI labs.
Medium term · Assessed
Meaning
The 92.77% claims rate, 447,576 of 482,460 works, reflects organised rights-holder mobilisation beyond what passive class action typically achieves, driven by the Authors Guild and comparable organisations.
Immediate · Reported

Source Landscape

This story draws on neutral-leaning sources from Belgium

Belgium×2

Primary parallel: The Google Books settlement, approved by Judge Denny Chin at first instance in 2008 then rejected by the Second Circuit in 2011, is the nearest structural precedent. Google had scanned 10 million books without authorisation and proposed a class settlement that would have given it a retroactive licence for all out-of-print books.

The DOJ opposed the settlement as a de-facto compulsory licence granted by private class action rather than by Congress, and the court agreed. Bartz v Anthropic faces a narrower but analogous objection: the class settlement establishes a compensation structure for AI training use that no legislation has authorised, and late opt-outs signal some rights holders believe the per-work figure undervalues their works.

Counter-parallel: The 2006 YouTube-Viacom settlement framework, eventually reached in 2014 for an undisclosed amount, proceeded without a court blocking the class structure. If the Bartz settlement avoids the compulsory-licence characterisation that sank Google Books, because it is structured as compensatory damages rather than a forward licence, the Second Circuit precedent does not apply directly and approval is more likely.

The $1.5bn total divided by 447,576 claimed works implies an average per-work settlement value of approximately $3,350 before the $187.5m fee deduction. After fees, the net-to-class figure is approximately $1.31bn, or roughly $2,935 per work.

That per-work floor sits above most individual bilateral licensing rates disclosed in the AI-publisher market to date (Reach plc's AWS deal is undisclosed on a per-work basis; Thomson Reuters disclosed $33m year-to-date across its full archive). It sits well below the implied per-work rate in News Corp's anticipated $1.5bn Anthropic figure , which covers a substantially smaller number of works under a bilateral rather than class structure.

The class settlement and bilateral licensing are pricing the same underlying asset, the right to train AI on copyrighted text, from opposite ends. The class floor suggests bilateral rates should be materially higher for commercially significant catalogues.

Consensus view: Ben Thompson (Stratechery), drawing on his analysis of the 2023 New York Times v OpenAI complaint and subsequent publisher licensing market, assessed that class-action copyright settlements against AI labs function as a price-discovery mechanism for the underlying licensing market, not as a deterrent.

The 92.77% claims rate in Bartz, 447,576 works out of 482,460, is unusually high for a class settlement and signals organised rights-holder participation that bilateral negotiators will cite as a revealed floor price where News Corp named $1.5bn as an anticipated settlement figure in SEC filings.

Counter-view: The Authors Guild, in its public statement on the Bartz settlement terms, argued that per-work settlement figures calculated from the class total are misleading because the class includes works with vastly different commercial weight: a self-published e-book and a Stephen King novel both count as one work.

On the Authors Guild's view, the settlement imposes a below-market price on high-revenue works by averaging them with low-revenue ones, which is why the Guild opposed class certification in similar actions against other AI labs.

Key tension: Whether Judge Martínez-Olguín's supplemental briefing on late opt-outs signals a structural fairness concern about the class definition, which could unwind the settlement, or is a procedural step that leaves the $1.5bn structure intact.

Sources:European Commission·European Commission·CourtListener / RECAP

Mentions:Google →California →YouTube →OpenAI →Robert Thomson →News Corp →New York Times →Claude →Anthropic →Araceli Martínez-Olguín →

First Reported In

Update #3 · ITV nears sale into Sky's AI stack

European Commission· 27 May 2026

Read original →

Causes and effects

Caused by

Alsup withholds approval on Anthropic $1.5bn

Judge Alsup's prior declination of the Bartz settlement on fee grounds set the procedural context in which Judge Martínez-Olguín (now reassigned) is conducting fresh scrutiny of the settlement terms.

Occurred 14 May 2026

Read story →

This Event

Judge probes $1.5bn Anthropic class deal

The class-action outcome sets the floor for every bilateral AI-content licence publishers negotiate next.

Different Perspectives

Italian market: Fastweb signed, Mediaset and RAI did not

Fastweb, the Italian broadband operator, signed Section 2 as a deployer, while Mediaset and the state broadcaster RAI are both absent. Italy therefore shows the split in miniature: a telecommunications company with one obvious deployment surface accepted the standard, and the two companies that actually broadcast to Italian audiences did not.

Paramount Skydance and its French exposure

Paramount filed a stipulation not to close on 24 July and had trial specialists admitted three days later, after nine months arguing that antitrust risk was regulatory rather than judicial. Its 22 July European clearance obliges it to divest a stake in the Universal International Pictures distribution venture and to strike no Universal distribution deal in Europe for a decade.

Time and the agent-advertising camp

Time and Mobian treat AI agent traffic as sellable inventory rather than leakage, on a site where bots now outnumber humans most days. Chief operating officer Mark Howard frames agent impressions as an extension of sponsorship, against an industry consensus still pursuing compensation for scraping.

German broadcasters

ARD, ZDF, RTL and ProSiebenSat.1 are all absent from the list, making Germany the largest single national bloc of non-signatories. Their position rests on three defensible grounds: signature is voluntary, most deployed systems are grandfathered to 2 December 2026, and a broadcaster's compliance surface spans production, archive, advertising and distribution rather than one product. None has publicly explained the decision.

Synthesia and the signed vendor layer

Synthesia signed Section 1 alongside Getty Images, Google, Meta, Microsoft, Anthropic, OpenAI, Mistral and Aleph Alpha, accepting a documented marking standard for the product they sell. London-headquartered Synthesia is the only pure AI-video vendor on the providers list, and gains a compliance artefact to put in front of European media buyers at the moment those buyers acquire a live obligation.

European Commission

The Commission published its initial signatory list on 31 July with roughly 190 organisations and let Article 50 take effect on 2 August as scheduled. Its own page stresses that adherence to the Code is voluntary while the transparency requirements are legal obligations, which frames non-signature as an evidentiary choice rather than a breach.