Authors v OpenAI: how to read a lawsuit without turning allegations into facts

On September 19, 2023, the Authors Guild and 17 writers filed a proposed class action against several OpenAI entities in the US District Court for the Southern District of New York. The initial complaint, case 1:23-cv-08292, alleged that their books had been copied without permission to train language models. On that date, no judge had decided whether those assertions were true or whether the use infringed copyright.

That last sentence prevents a lawsuit from being distorted. A complaint gives one side’s theory and requests relief; it is neither a judgment nor a neutral investigation. Accurate reporting separates what is alleged, the support offered, the opposing party’s answer, the legal question and the court’s decision.

Who sued and what they requested

The filing named the Authors Guild and David Baldacci, Mary Bly, Michael Connelly, Sylvia Day, Jonathan Franzen, John Grisham, Elin Hilderbrand, Christina Baker Kline, Maya Shanbhag Lang, Victor LaValle, George R. R. Martin, Jodi Picoult, Douglas Preston, Roxana Robinson, George Saunders, Scott Turow and Rachel Vail. That was 17 authors, not “17 others” in addition to Martin and Grisham.

The case was filed on behalf of a proposed class of professional fiction writers. “Class action” did not mean the class had already been certified. The court would later have to decide whether procedural requirements were met. Meanwhile, the named writers and association were the parties advancing the case.

The plaintiffs asserted direct copyright infringement and requested damages, profits, injunctions, costs and other relief. Readers should inspect the final pages of every complaint: they distinguish a broad account from the legal conduct the court is asked to declare and remedy.

The central allegations

The complaint alleged that OpenAI copied complete registered works and included them in datasets used for training. It also said collections sourced from pirate book repositories were likely involved. “Alleged” and “argued” must remain attached to those statements until admissible evidence or a judicial finding exists.

The data claim did not come from a list published by OpenAI. Indeed, the GPT-4 technical report said it withheld further details about architecture, hardware, training compute, dataset construction and training method. That lack of transparency made title-by-title verification difficult; it did not itself prove a particular book had been copied.

The complaint used ChatGPT answers and summaries as indications of familiarity with the works and described outputs employing authors’ characters or worlds. A detailed summary can guide an investigation, but does not directly reveal the data route. It might derive from the book, reviews, authorized databases or other sources. Attribution requires logs, dataset versions, dates and technical evidence obtained through the case.

Training and output are different acts

The litigation implicated two places where copying might occur. During data preparation, a work can be downloaded, stored and transformed into training examples. During use, a system may produce an output that reproduces protected expression or is substantially similar. Each act has its own participants, evidence and legal analysis.

A model that answers questions about a novel does not necessarily contain a readable library of copies, but that technical description does not resolve whether copies were made during training. Asking only whether the model “memorizes” confuses earlier conduct with the final state. Investigation must follow the full cycle: acquisition, storage, filtering, training, retention and generation.

Artistic identity adds another distinction. A general style does not automatically receive the same protection as a specific sentence, character or expressive sequence. A fraudulent book sold under another person’s name may also raise trademark, attribution or unfair-competition issues. Not every legitimate concern belongs to one cause of action.

What fair use means

The US Copyright Office explains that fair use is assessed case by case through four factors: purpose and character of the use, nature of the copyrighted work, amount and substantiality used, and effect on the potential market. There is no automatic rule that the labels “research” or “transformative” make every copy lawful.

Nor is it enough that a work could be found online. Accessible does not mean uncopyrighted or authorized for every use. Conversely, copying a whole work does not decide the result on its own: courts weigh all factors and applicable precedents. Treating either formula as a closed answer would substitute for the court.

As of September 19, OpenAI had not filed an answer in that docket setting out its defenses. Assigning a specific argument about these books to the company would have been premature. Balanced reporting does not invent the reply; it says the reply is absent and adds it when it enters the record.

The evidence matrix

For each work, a useful table would have six columns: copyright registration, particular copy, provenance, acquisition time, use in a model version and challenged output. The first may be documented through a certificate. The middle columns usually require discovery: inventories, logs, communications and testimony. The last requires preserving the prompt, date, model and full response.

Then label the support type: public document, party statement, expert inference or admitted fact. A character match can support a question without resolving its source. An identical line may be more informative, but needs context and comparison excluding unprotectable material.

Reproducibility matters. One answer can change with model, filters or randomness. Anyone using it as evidence should preserve a screenshot, export, available parameters and repeated attempts, without soliciting or publishing long protected passages. The goal is to demonstrate the relevant behavior with the least material necessary.

What a court may decide

Before reaching the merits, the case could face standing, pleading, class-definition and discovery questions. Surviving an early motion would not mean winning; it would mean a claim could proceed under the applicable procedural standard. A settlement would not necessarily create a rule for the whole industry.

A liability decision would need to identify specific acts and apply law to that record. Remedies could differ: damages for works, limits on certain practices, information or technical measures. “Ban AI” was not a description of the requested relief.

Industry effects would depend on scope. A rule concerning a pirate dataset would not be identical to one involving licensed data. A finding about substantially similar output would not automatically resolve training. The exact words of a decision would matter more than the winner named in a headline.

How to follow the case

Begin with docket number, court, date and document. Then record each claim’s status: alleged, answered, dismissed, in discovery or decided. Always link the order, not a party’s summary. If a complaint is amended, preserve both versions and do not attribute the new allegation to the earlier date.

The Authors Guild explained its position the following day and published the author list. It is a primary source for what the party wants and why it is acting; it is not an independent source for whether OpenAI infringed. Identifying that function makes it useful without adopting its conclusion.

The transferable skill is reading litigation as a state map: allegation, evidence, defense, rule, decision and remedy. In September 2023 there was a serious complaint and difficult technical questions; there was no verdict. Preserving that boundary protects the reader and makes it visible when a court crosses it.