bartz v. anthropic: settlement outline and implications

first, for all who i’ve argued with over the last three years:

told you so!

fair use is a real thing that deserves protection, and copyright is often unfairly weaponized, but this was never going to end any other way. you cannot pirate millions of works for a commercial activity without consequences.

will this be the end of ai? no, of course not. even anthropic and openai, despite their financial “independence,” can raise enough private capital to cover any settlement. meta and google were never at risk from a solvency perspective anyway.

it will, hopefully, be the beginning of a more fair future.

the “pirate” who settled

detailed analysis of why anthropic settled bartz et al. v. anthropic pbc just days before damaging discovery rulings, avoiding a december 2025 trial that could have reshaped ai training law. the case centered on anthropic’s admitted use of “pirated” book datasets to train claude, with potential damages reaching $22.5 billion.

the original complaint pulled no punches: “anthropic has built a multibillion-dollar business by stealing hundreds of thousands of copyrighted books. rather than obtaining permission and paying a fair price for the creations it exploits, anthropic pirated them.”

on august 26, 2025, anthropic settled rather than proceed to a december 1 trial that would have exposed exactly how claude was trained on admittedly pirated materials. the settlement came hours before a noon deadline to produce documents that a special master ruled weren’t privileged.

case basics

citation: bartz et al. v. anthropic pbc, no. 3:24-cv-05417-wha (n.d. cal. filed aug. 19, 2024)
judge: william h. alsup (who can actually code)
plaintiffs: andrea bartz (the lost night), charles graeber (the good nurse, the breakthrough), kirk wallace johnson (to be a friend is fatal)
core allegation: “anthropic has built a multibillion-dollar business by stealing hundreds of thousands of copyrighted books”
trial date avoided: december 1, 2025

why anthropic settled: the final week

five days of adverse rulings

august 22, 2025: special master orders production of 35 documents anthropic claimed were privileged, including one where the legal department was “intentionally excluded”

august 23, 2025: three adverse orders:

anthropic must provide missing copyright agreements (doc 341)
expert witness battle schedule set - damages could reach billions (doc 342)
special master compels answers about “pirated books” list (doc 343)

august 25, 2025: discovery disputes escalate:

8am: sarah rodriguez fails as 30(b)(6) witness
evening: three emergency motions expose anthropic’s obstruction
missing slack channels about datasets never produced
mysterious spreadsheet about deletions clawed back

august 26, 2025:

noon: deadline to produce damaging documents
afternoon: settlement announced to avoid production

the evidence anthropic couldn’t explain

“pirated books” terminology: special master’s order used quotes, suggesting anthropic internally called the datasets “pirated”
excluded legal department: document antpriv_0000642 showed “the legal department was intentionally excluded”
missing slack channels: [redacted] channel with “obvious relevance given its title” never produced
deletion spreadsheet: five copies of spreadsheet about [redacted] produced then all clawed back claiming privilege
co-founder involvement: document questioning about jared kaplan (co-founder) immediately triggered clawback

what the december trial would have looked like

evidence the jury would have seen

if the case had proceeded to trial, the jury would have reviewed:

the books3 smoking gun:

anthropic admitted using the pile dataset, which included books3
books3 = 196,640 pirated books from bibliotik (“a notorious pirated collection”)
created by shawn presser who tweeted: “now you do. now everyone does. presenting ‘books3’, aka ‘all of bibliotik’”
eleutherai’s own paper admitted: “we included bibliotik because books are invaluable for long-range context modeling”

anthropic’s admissions:

july 2024: company spokesperson confirmed use of the pile to proof news
december 2021 paper: described dataset as “32% internet books” (code word for pirated copies)
internal references to “pirated books” (per special master’s order)
inability to explain retention or deletion practices

damages calculations:

approximately 150,000 copyrighted works identified
statutory damages range: $750 to $150,000 per work
potential exposure: $112 million to $22.5 billion
lost licensing revenue claims from author class

why fair use wouldn’t save anthropic

the complaint demolished anthropic’s potential fair use defense:

commercial purpose: anthropic projected $850 million revenue in 2024, valued at $18+ billion after raising $7.6 billion from amazon and google

wholesale copying: “anthropic downloaded known pirated versions of plaintiffs’ works, made copies of them, and fed these pirated copies into its models”

market harm:

claude marketed to “draft everything from a text message or email to a screenplay or a novel”
tim boucher “wrote” 97 books using claude in less than a year, selling them for $1.99-$ 5.99
each book took only “six to eight hours” to generate
ai-generated copycats flooding amazon when real authors release books

judge alsup’s technical background would have enabled examination of:

how training data is processed and stored
whether copies persist after training
technical necessity of using complete works
alternatives to copyrighted material

witness testimony disasters

ben mann (anthropic’s principal technical witness):

former openai vp who led gpt-2 and gpt-3 development
designated as 30(b)(6) witness on dataset handling
repeatedly claimed inability to answer basic questions
10,000+ documents produced afternoon before his deposition
anthropic clawed back key document mid-deposition when plaintiffs tried to use it

sarah rodriguez (30(b)(6) discovery witness):

unable to describe document retention systems
could not identify repository structures
no knowledge of deletion protocols
deposed august 25, one day before settlement

these 30(b)(6) failures created litigation risk - corporate representatives unable to answer basic questions about company operations indicate either inadequate preparation or problematic underlying facts.

the settlement terms (what we know)

what’s public (or probable)

binding term sheet signed august 25, 2025
class-wide settlement covering all authors whose works may have been used
permanent resolution of all related claims involving the parties
mediation by hon. layn phillips (renowned tech mediator)

what’s (currently) sealed

dollar amount (likely substantial given liability exposure)
any injunctive relief (probably none)
individual allocations (how much each author gets)
cy pres provisions (where unclaimed funds go)

i think i am eligible to join the class; will confirm when the process is public.

reasonable estimates

based on the complaint’s allegations and liability exposure:

maximum statutory damages: 150,000+ works × $150,000 = $22.5 billion

minimum statutory damages: 150,000+ works × $750 = $112.5 million

likely settlement range: $50-500 million (typical for tech settlements avoiding precedent)

per-author recovery: varies by number of works infringed

future impact: no injunction likely, anthropic continues using the data

earlier, eddie lee predicted “business-ending liability” for anthropic, calculating that even 100,000 works could result in $1-3 billion in damages at standard rates, with willful infringement potentially reaching $15 billion.

parallels to openai’s trajectory

the bartz case mirrors tensions in openai’s evolution:

the mission vs. money conflict

like openai’s 2019 capped-profit pivot, anthropic faces the fundamental tension between:

original mission: ai safety and benefit to humanity
commercial reality: need for capital and competitive pressure
legal exposure: copyright liability from training methods

the transparency retreat

anthropic’s discovery obstruction parallels openai’s transparency decline:

openai 2018: gpt-1 fully documented
openai 2019: gpt-2 withheld for “safety”
openai 2023: gpt-4 details secret
anthropic 2025: can’t/won’t explain training data sources

settlement implications

why anthropic settled

the timing reveals anthropic’s calculus:

discovery disaster: 30(b)(6) witnesses couldn’t answer basic questions
document production issues: late dumps and clawback disputes exposed
adverse rulings imminent: special master likely to rule against them
december 1 trial looming: risk of precedent-setting verdict

what the settlement means

for anthropic

prevents precedent that could threaten their other cases
potentially undermines openai, who now has to explain what was revealed in deposition…
protects secrets about actual training data practices
pays for peace

for the ai industry

first major settlement in ai training litigation
establishes value for copyright claims
encourages more lawsuits from content creators
no legal precedent on fair use question

for authors

compensation for class members (amount unknown)
validation that their claims have merit
faster resolution than years of litigation
but: no injunction against future use

predictions and precedent

forward-looking predictions

based on this settlement and typical patterns in tech litigation:

near term (3-6 months)

copycat lawsuits: the class certification precedent will trigger similar cases
settlement amounts leak: financial press will likely uncover the sealed terms
statutory damage fear: other ai companies will rush to settle after seeing anthropic fold
licensing deals emerge: preventive agreements become standard practice

medium term (6-18 months)

legislative action: congress intervenes to prevent “business-ending” scenarios
damage cap proposals: tech lobby pushes for statutory damage reform
technical adaptations: shift to synthetic data becomes existential necessity
insurance crisis: liability insurance for ai training becomes prohibitively expensive

long term (2+ years)

the great relicensing: ai companies retroactively license training data?
market bifurcation:
- us models constrained by copyright liability
- international models trained on “everything”
judicial precedent: supreme court finally addresses ai fair use
new business models: authors become data vendors, not just content creators

of course, there is always discussion of the white house’s opinion, but the fact is that this has very little to do with them. no executive order has the power to udno the actions of courts in such clear statutory situations. there are practically zero regs and very little agency guidance involved here.

the exaggeration factor

while predictions of “business-ending” damages grab headlines, anthropic’s quick settlement suggests the real calculus is more nuanced. the company likely paid far less than the theoretical maximum to make the case disappear - probably in the hundreds of millions, not billions. the threat of massive statutory damages serves as a powerful negotiating tool, but actual business destruction remains unlikely for well-funded ai companies willing to settle.

lessons for ai companies

discovery preparedness

the 30(b)(6) failures reveal critical governance gaps:

document retention policies must be clear and followed
corporate representatives need actual knowledge
technical details can’t be hidden forever
courts expect transparency about data sources

the settlement calculation

anthropic’s decision matrix likely included:

trial risk: potential for massive statutory damages
precedent risk: adverse ruling affecting entire industry
discovery exposure: damaging internal communications
competitive impact: distraction from product development

strategic positioning

the case demonstrates that ai companies must:

document data provenance defensibly
prepare for litigation from day one
balance growth with legal risk
consider preventive licensing

connection to broader ai developments

the china factor

deepseek’s $6 million training cost changes the equation:

if models can be trained cheaply, copyright damages matter less
but u.s. companies face stricter legal environment
creates competitive disadvantage for rule-following companies

technical details revealed

despite redactions, technical insights emerged:

data pipeline architecture

acquisition: bulk downloads from specific sources
storage: distributed across multiple repositories
processing: [redacted] transformation steps
integration: merged into training pipeline
retention: unclear policies on raw data

organizational knowledge

engineers knew about copyright concerns
leadership involved in dataset decisions
deletion policies changed over time
documentation exists but heavily protected

conclusion

if it concludes, the bartz settlement is a big moment. even if it doesn’t, it will still stand as the first time the cards began to fall.

ultimately, bartz v. anthropic demonstrates that the era of “ask forgiveness, not permission” is finally ending. it’s time to actually do things like adults who live in a political system descended from the magna carta.

i can’t wait.

sources and references

primary documents

document 341-355: complete case file analysis
openai timeline: industry context and parallels

key filings analyzed

motion to approve class notice (doc 350)
letter motion re discovery failures (doc 348, 353)
settlement notices (doc 354, 355)
expert evidence orders (doc 346)
privilege disputes (doc 344, 345)