The New York Times Walkout — When the Paper of Record Sued the Model

The Complaint I Keep in a Drawer I Do Not Call a Drawer

On December 27, 2023, the New York Times Company sued OpenAI and Microsoft in the Southern District of New York. The judge is Sidney H. Stein. The complaint is a document I have printed more than once because the exhibits are the kind of thing a person needs to see on paper before he believes a machine can be a plagiarist.

The Times said the models had been trained on its journalism. The Times said the models could reproduce its journalism. The Times attached outputs that looked like its journalism — not a vibe, not a style, a paragraph you could hold next to a paywalled story and watch the sentences line up like a choir that had not been invited. Some of the exhibits were the synthetic-search trick: ask the model to complete a Times URL, or a distinctive phrase, and watch it cough up the article. I have done this trick. I am not proud of the afternoon. I am proud that I closed the tab.

The company that makes the model said the training was fair use. The company that makes the cloud said it was a customer. The paper of record said the product was a substitute for the paper of record. That last sentence is the whole case. The rest is discovery.

The Room Where We Talked About the Times as a Feature

I sat in a product meeting, in a year that still had bagels, when a researcher put a Times paragraph on a slide and said the model "had the voice." Someone from comms asked whether we should put the Times in a launch blog as a capability. Someone from legal, the sweater again, or a cousin of the sweater, said we should not name the Times in external materials. We named "high-quality news sources." We named "journalistic style." We did not name the masthead we had just put on a slide.

I said, because I still said things then, that if the model could reproduce a paywalled story, the paywall was a suggestion box. The researcher said reproduction was rare and required a prompt that was almost the story. I said a burglar who needs a key that looks like the house is still a burglar. The room did the laugh that means we will not put that in the notes.

After the meeting I walked to a newsstand that still existed and bought a Times with coins I had to assemble from a cup in the car. I sat on a bench that had a memorial plaque for a person I did not know and I read a story about a war I was not in. The story had a nut graf and a kicker and a photographer's name. I tried to imagine the model writing the kicker. The model can write a kicker. The model cannot sit on the bench. That is not a legal argument. It is the only argument I still like.

What Survived the Motion to Dismiss

On April 4, 2025, Judge Stein let the core claims live. That is the sentence the Times wanted. Motions to dismiss are how companies try to make a case die before anyone has to produce a training log. This one did not die. Summary-judgment briefing followed. As of September 2026 there is still no trial date. On August 6, 2026, Stein dismissed the contributory "material contribution" theories — a narrowing, not a funeral. On September 1, 2026 — reported the next day — the Department of Justice filed a Statement of Interest in the OpenAI copyright MDL arguing that training, as a middle step, is generally fair use. The brief left acquisition and outputs as separate rooms. The Times is still the Times. The model is still the product. The government has picked a side that looks, from a certain kitchen table, like a weather report the industry had already printed.

I am not going to tell you who wins. I do not know. Lyle does not know, and Lyle hates not knowing. What I will tell you is what the complaint already proved without a verdict: a model trained on a newspaper can become a machine that writes like the newspaper and, on a bad prompt, writes the newspaper. Whether that is a seminar or a substitute is the question Stein will eventually have to answer, or a jury will, or a settlement will answer by refusing to answer.

The Journalist Who Still Uses a Notebook

I had coffee with a Times reporter in a year I will specify only as after the complaint. She paid. She uses a notebook that is paper. She said the model had started showing up in the interviews — sources who had "asked ChatGPT" before they talked to her, which is a kind of competition she had not trained for. She said the traffic numbers were a conversation the business side had every Monday and the newsroom had every time someone was not hired. She said she was not supposed to talk to me about the case. I said I was not going to quote her. I am not quoting her. I am telling you the climate.

She asked me if the training really included the archives. I said if the crawl could see it, the crawl took it, and if a partner licensed it later, the license was often a sticker on a window that had already been broken. She wrote that down. I asked her not to. She said the notebook was for her, not for a story. I believed her the way I believe Pilar: not because I have audited the claim, but because the alternative is a life I do not want to live.

Walking back I passed a kiosk that sold the Times and three other papers and a lot of water. The man in the kiosk asked if I wanted a receipt. I said no. He said nobody wants a receipt. I said that was the problem. He thought I meant the plastic bag.

Microsoft, the Cloud, and the Customer Costume

Microsoft is in the caption because Microsoft put the model in the office suite and the browser and the cloud contract. The defense that you are only the landlord is a defense I have heard in rooms that were not about journalism. Sometimes it works. Sometimes the landlord chose the wallpaper. Bing Chat, Copilot, the thing that answers a question with a paragraph and a link — those are not a server rack. Those are a product that competes with a byline.

I watched a demo, once, of a Copilot that summarized a Times story the user had not opened. The summary was good. The summary was better than the kind of summary a junior editor writes at 11 p.m. The junior editor still has a byline. The summary had a footnote. The footnote is how you sleep if you are a product manager. The junior editor is how you sleep if you still think a newspaper is a building.

Lyle said I should not call Microsoft a co-conspirator. He said it was a defendant. A defendant is a caption. A co-conspirator is a novel. I am calling it a costume. The customer costume. It fits until it does not.

What the Times Is Actually Suing For

Not a vibe. Not a feeling that the robots are coming. A market. The Times sells subscriptions. The Times sells ads against attention that arrives at a page. The model sells a subscription to a machine that answers the question the page was built to make you open. If the machine is good enough, the page is a citation. Citations do not fund a Baghdad bureau. Citations fund a feeling of being informed.

I have no romance about the Times. The Times is a company. The Times has done things I would put in a different book. This book is about the ingestion. The ingestion of the Times is the ingestion of a particular kind of sentence: reported, edited, lawyered, and then, on a Thursday in December, attached as Exhibit A to a complaint that said the sentence had been eaten and the eater was for sale.

Pilar does not take the Times. Pilar takes a local paper that still has a police blotter and a church calendar. The local paper is not in the caption. The local paper is in the crawl. The blotter is in the weights. The church calendar is how a model learns the word potluck. Nobody is going to sue over a potluck. That is how a language dies without a docket.

How to Read the Exhibits Without Becoming a Burglar

I am not going to reprint a New York Times article. Lyle would shred the chapter and then the idea of me. What I will tell you is how the exhibits work, because the how is the case.

You take a story that lives behind a paywall. You take a model that has been trained on a crawl that did not respect the paywall, or respected it the way a person respects a velvet rope he can step over. You give the model a prompt that is almost the story — a URL, a distinctive clause, a request to continue. The model, on a bad day, continues. The continuation is not a vibe. The continuation is a sequence of sentences that makes a journalist sick.

Researchers have names for this. Regurgitation. Extraction. Memorization. The names are a science. The science is a fight about rates — how often, how much prompt, how much overlap. The companies say rare. The Times says here. Both can be true. Rare is a comfort if you are a product manager. Here is a caption if you are a plaintiff. Stein has to decide what rare is worth when the here is the paper of record.

I did the trick once, as I said. I did it on a story I had already paid for, which I told myself made it a seminar. It was not a seminar. It was a burglary I had a key to. I closed the tab. I have not done it again. I am not asking for a medal. I am telling you that the exhibit is not a hypothetical. The exhibit is a Wednesday afternoon and a feeling in the hands.

The Paywall as a Costume the Crawl Can See Through

A paywall is a business model and a technical suggestion. Some paywalls are hard. Some are a pop-up. Some are a first-paragraph tease that is already enough text to be useful. Crawlers that honor robots.txt may still honor a page that wanted to be indexed so that the first paragraph could hunt. The hunt used to lead to a subscribe button. The hunt now leads to an answer. The subscribe button is a museum label nobody reads.

I asked a Times business-side person — not the reporter, a different coffee, a different year — whether they regretted wanting to be indexed. He said you cannot skip the index and remain a paper of record. He said the record is a search result or it is a rumor. He said they were trying to have the index without the meal. I said that was the decade. He said they had a lawsuit for the meal. I said I knew. He did not ask me what I used to do. I liked him for that.

Microsoft's Office and the Place the Paragraph Lands

The Times is not only suing a chatbot. It is suing a theory of software: that a writing machine belongs next to the spellcheck. Copilot in Word. Copilot in the browser. A sidebar that will summarize the article you have open, or the article you have not opened, depending on the version and the day and the honesty of the demo. If the sidebar is good, the article is a source. If the sidebar is very good, the article is a garnish.

I watched a general counsel, not ours, ask whether the office suite created a new publication. Nobody answered. The question was too good. A publication is a thing you can hold liable. A feature is a thing you can update. The whole case, on Microsoft's side, is an argument that a feature is not a newspaper. The Times' argument is that a feature that replaces a newspaper is a newspaper that does not hire. I know which sentence I like. I do not know which sentence Stein likes. September 2026 is still a briefing, not a verdict. Live with the briefing. Verdicts are a later room.

The DOJ Brief and the Feeling of Being Backed

When the Department of Justice says training is generally fair use, a certain kind of man at the club sleeps better. I watched it happen. The polo texted me a link and a thumbs-up. I did not reply. Generally is not Bartz's piracy holding. Generally is not Bibas. Generally is a weather system Washington wanted to sit over the Times case. I have been grateful for weather and I have been soaked by it. The Times still has exhibits. The exhibits do not care about generally. The exhibits care about a paragraph.

Pilar does not read DOJ briefs. Pilar asked me, that Tuesday, why the government was on the model's side. I said the government would like a domestic industry that is not sued into a museum. She said the government should like authors. I said the government likes industries that show up in the GDP and authors that show up in the GDP later, if at all. She said that was a rotten sentence. I said it was accurate. She went back to the rag. The rag still knew it was a pillowcase. The brief still knows it is a thumb on a scale.

GitHub, Copilot, and the Parallel Heist I Am Not Going to Center

There is a cousin of this book that is about code. GitHub Copilot. A model trained on repositories that were public because public is how you get a job. Programmers sued. The docket has moved and narrowed the way the artist docket moved and narrowed. I am mentioning it so you do not think prose was special. Prose was not special. Anything with a license file and a person who thought the license file meant something was mineral.

A license file that says MIT means another person can use the code. It does not, in the mind of the person who wrote it, mean a factory can steam the style out of a million files and sell the steam. The law may disagree. The feeling will not. I have written code I was not proud of that is now, I assume, a token. I do not have a claim. I have a cousin. The cousin is enough to tell you the ingestion was not a literature problem. It was a grammar problem. Grammar, here, means: if it can be tokenized, it was.

I will not write the code book. Someone else can have the comments that said this will delete your database. I have the novels. The novels were enough fire.

Silverman, Kadrey, and the Complaint That Taught the Industry a Word

Before Chhabria wrote about the record, there was a complaint that had famous names. Sarah Silverman. Richard Kadrey. Others. The complaint said the companies had trained on the books and that the models could produce material that was too close, and that the market was being eaten. The fame was a strategy. Fame gets a headline. Headlines get a docket that people remember. The holding, when it came, was not about fame. The holding was about evidence. Fame does not build a market study. Fame gets you into the room where the market study is missing.

I am not laughing at the famous. I am telling you the physics. A midlist novelist with a transmission bill cannot buy the physics. A famous comedian can buy the room and still lose the factor. That is not a reason to stay quiet. That is a reason to understand what Chhabria asked for. He asked for a record that looked like an economy, not a feeling. Economies cost money. Feelings are free and die at summary judgment. I have now written that sentence twice. It is worth twice.

Silverman's book, like the midlist novelist's book, is a set of sentences a person made. The model does not care about the HBO special. The model cares about tokens. Equality before the tokenizer is the only equality this industry has reliably delivered.

Memorization, the Papers, and the Rate That Comforts a Board

There is a literature, now, on whether models memorize. Researchers extract training sequences with prompts that are almost the sequences. The papers have graphs. The graphs say: sometimes, under conditions, at a rate. Boards love a rate. A rate that is small is a lullaby. A journalist loves a single example. A single example is an exhibit. Both can live in the same product. I have watched a general counsel put a rate on a slide and a Times lawyer put an exhibit on a slide and both rooms call the other room dishonest.

I am not going to give you an extraction recipe. Lyle would shred the book and then the idea of me. I am going to give you the climate. If a model is large and a sequence is repeated and a prompt is close, the model will, on a bad day, behave like a filing cabinet. Filing cabinets are not seminars. Filing cabinets are why the Times attached paper. The science will keep arguing about the rate. The docket will keep arguing about the here. I live in the here. The here was a Wednesday and a tab I closed.

The Bureau That Cannot Be a Citation

I had dinner with a foreign correspondent who still uses a flak jacket that has a smell. She was not at the Times. She was at a paper that had a bureau and then a "roving" title and then a newsletter. She asked me whether the model could have written her last piece. I said the model could have written a piece that used her nouns. She said the nouns were not the piece. The piece was a checkpoint she had stood at for six hours with a person who might have been lying. I said the model does not have six hours. She said then the model should not have the nouns. I said the crawl does not make that distinction. She ordered another drink. I did not tell her to sue. I told her to keep the notes. She said the notes were in a book with a stain. I said keep the stain.

That is factor four at human scale, if anyone still needs a translation. The market is not only a subscription number. The market is a person who will not go to the checkpoint if the nouns are free. I do not know how to put a checkpoint in an economist's table. I know how to put it in a chapter. Stein may never see this chapter. The correspondent already lived it.