- The trial in Bartz v. Anthropic begins on December 1, 2025, 7:30 AM sharp.
- Judge Alsup has already ruled what I called a Solomonic decision: (1) Anthropic’s training of its AI model with copies of books was fair use (with “further” uses of copies derived from pirated datasets or Anthropic’s own in-house book scanning of books similar to Google Book search’s project). But Judge also ruled: (2) Anthropic initial copying from notorious shadow libraries online containing many pirated books–files numbering 7 million–is a separate use for Anthropic’s own research library building and is NOT fair use.
- Because Bartz didn’t file a motion for summary judgment, Judge has all but ruled that Anthropic’s downloading of pirated books is infringement. Technically, Bartz will have to prove it at trial. Infringement is almost a forgone conclusion.
- The real issue at trial will be the jury’s calculation of statutory damages based on the number of copyrighted books/works in the class. The total number of works will be defined by Bartz’s submission of titles of works that fall within the criteria Judge Alsup selected: (1) works Anthropic downloaded from the shadow libraries LibGen and PiLiMi (total copies 7 million), (2) that have ISBN or ASIN and (3) copyright registration within 5 years of publication effective before Anthropic’s downloading or within 3 months of publication. These criteria ensure that all plaintiffs in the class are entitled to statutory damages under the Copyright Act.
- We won’t know the total number of works within the class until Sept. 1 when Bartz produces a list.
- But, even if the number is below 1 million, statutory damages can still be several billions of dollars. I lay out different possibilities here.
- And, if willful infringement is found by the jury, Anthropic’s chance of receiving business-ending liability escalates dramatically.
- For example, even if the class contains only 750,000 works, Anthropic faces a maximum damages amount of $112.5 billion dollars — more than the company’s current valuation.
September 1 and December 1, 2025. These dates may determine Anthropic’s business fate.
And, by business fate, I mean Anthropic’s continued existence.
Some readers may think I’m being hyperbolic. I wish I were. But the grim reality is that Anthropic now faces a class action (subject to the plaintiffs’ meeting the condition of preparing a list of works within the class by Sept. 1) that will leave in the hands of a jury in San Francisco (between 6 and 12 members) just how much in terms of statutory damages that Anthropic must pay.
I ran the numbers for damages with different calculations for possible amounts a jury could find. Most of them are bad for Anthropic. Many expose Anthropic to possible existential liability.
And, if the jury finds Anthropic’s downloading of pirated books from shadow library was willful (knowing it was illegal) infringement, the damages amount will be enormous. In my calculations, a best case scenario might be $300 million (10,000 works x $30,000). And the worst case scenario could be $900 billion (6 million works x $150,000). In between, there’s still a lot of scenarios where the damages amount will equal or surpass the current valuation of Anthropic. For example, with 1 million works in the class, and the jury picks $100,000 per work (close to what the jury picked in Cox Communications), the total award would be $100 billion!
an appeal might not save anthropic even if it’s right legally
Some readers have suggested that Anthropic will be saved by an appeal of Judge Alsup’s ruling on the downloading of pirated books (for library building) as a separate use from the training.
Not so fast. Typically, an appellant must file a supersedeas bond in the amount of the damages award to have it stayed (Fed. R. Civ. P. 62(a)). So, let’s say, the jury awards $100 billion in statutory damages in the “middle” range for willful infringement (see above), Anthropic would have to post a bond in that amount plus a modest interest.
How exactly would Anthropic come up with the funds to post the bond?
UPDATE: district courts do have the discretion to waive the bond requirement under Rule 62. One influential approach was set forth by the Seventh Circuit:
In Dillon v. City of Chicago, [866 F.2d 902, 904 (7th Cir. 1988)], the Seventh Circuit identified five factors that should be considered when evaluating a request to waive Rule 62(b)’s bond requirement: (1) the complexity of the collection process; (2) the amount of time required to obtain a judgment after it is affirmed on appeal; (3) the degree of confidence that the district court has in the availability of funds to pay the judgment; (4) whether the defendant’s ability to pay the judgment is so plain that the cost of a bond would be a waste of money; and (5) whether the defendant is in such a precarious financial situation that the requirement to post a bond would place other creditors of the defendant in an insecure position. 866 F.2d at 904.
Am. W. Bank Members v. Utah, No. 2:16-CV-326, 2024 WL 1658193, at *2 (D. Utah Apr. 17, 2024)
Not sure any of these factors except perhaps the fourth factor would be present in this case.
what about meta, openai, and Microsoft — do they face existential liability?
Other readers have asked me about Meta, OpenAI, and Microsoft — weren’t they using the same pirated books datasets?
Largely yes for Meta, but Judge Chhabria disagreed with Judge Alsup’s analysis of the downloading separate from the further purpose of training the AI model. Of course, the evidence related to Anthropic’s conduct and Meta’s conduct are different. So that may help to explain some of the differences in approach. So, at least right now, Meta doesn’t face the business-ending liability scenarios of statutory damages even in theory.
One irony is that Anthropic later decided to turn to buying books and scanning digital copies of them, no doubt concerned about potential liability from using pirated books from shadow libraries. (I haven’t seen much discussion of what Anthropic’s lawyers may have done, but, in the Meta case, it’s apparent lawyers got involved and advised the company on whether it was permissible to use copies from the shadow libraries under existing law. Such legal advice would be relevant to negating willfulness of infringement if the company had a good faith basis to believe their conduct was permissible. And, as I stated in testimony in a Senate hearing, I agree with Judge Chhabria’s fact-specific approach on this issue and think Judge Alsup’s categorical approach is wrong.) Judge Alsup said the digital scanned copies of purchased books was fair use, but statutory damages won’t be decreased by the millions of dollars spent to buy the books.
However, OpenAI and Microsoft face lawsuits in New York before Judge Stein, at least 13 lawsuits involving book author proposed class actions and a collection of news media, including the New York Times. If Judge Stein follows Judge Alsup’s approach to pirated books and if Judge certifies a class action of book authors similar to Judge Alsup, the risk exposure OpenAI faces will be far greater than Anthropic’s.
As I calculated in my law review article, OpenAI’s exposure to statutory damages from alleged copying of New York Times articles (meaning just 1 plaintiff) is something below $1.5 trillion if willful infringement is found and the maximum amount of $150,000 per infringed work used. (The number should be well below that outer limit because a compilation of news articles published in the same edition will be treated as a compilation and just 1 work for calculating statutory damages).
But, if you combine all the books and news articles that OpenAI allegedly downloaded, the total number of works copied by OpenAI in these lawsuits has to far surpass the outer limit in Anthropic, which only involves the books datasets.
It’s too early to tell how Judge Stein will view the downloading (separate from or a part of the training?). But OpenAI’s business fate may hinge on Judge Stein’s answer.
OpenAI’s current valuation is $300 billion. Under many scenarios of statutory damages if willful infringement is found, OpenAI would face business-ending liability as well.
Given Microsoft’s size, its standing presumably is more stable. But let’s not forget the book authors led by Kai Bird recently filed a class action lawsuit against Microsoft for its own training with pirated books.
As things stand now, Meta and Google seem least exposed to potential liability, while Anthropic the most exposed. And, depending on how Judge Stein rules on downloading, OpenAI could face a similar predicament as Anthropic.
What might have seemed far-fetched to some commentators–AI business-ending liability through statutory damages–is no longer if it ever was.
Related stories
One response to “Will a jury in SF decide Anthropic’s business fate? Most likely yes.”
We can only hope so…