The Model Is a Baked Cake. NYT v. OpenAI Will Likely Cut a Middle Path

Nathan Labenz introduces Cecilia Ziniti as founder of GC AI, an assistant for in-house counsel, previously general counsel at Replit and lead counsel on Amazon’s Alexa team, and the author of widely shared Twitter threads on generative AI and the law. He is frank that, aside from a couple of provisional patent filings — one written with GPT-4 — this is not a field he has studied. The conversation is built as a primer: why intellectual property exists, what branches it has, and how those branches now collide with large language models.

The spark is the New York Times lawsuit against OpenAI, which Labenz calls a potential precedent-setter that could be referenced in history books. Legislators from the Founders, who wrote IP into the Constitution, through later Congresses that extended copyright, did not anticipate generative AI. For now the courts have to apply an old legal paradigm to a new technical one. Among the questions Labenz brought: what rights do content owners have, does the Times have a good case, and could a court really order GPT-4 destroyed, as the Times demands.

He tells Ziniti that AI builders spend time on upside visions and on catastrophic risk, and far less on the practical collision with law. Lawsuits have been piling up; this one, he says, has counterparties big enough to make a real case. She agrees to start from first principles — why IP exists, its main sections, the live issues — before walking into how AI changes the picture. The education that follows is aimed at enthusiasts who have not spent much time thinking like lawyers.

Ziniti starts with a fact most technologists miss: the big tech companies are not big holders of physical property. What they hold is software, marketing, and brand. Look at the Fortune 20 in 1980, she says, and there is not a single tech company; today four of the five most valuable names, the Magnificent Seven neighborhood, are built on intangibles. The law recognized that long ago. Article I, Section 8 of the Constitution tells Congress, for a limited time, to grant monopolies to authors and inventors in order to promote the progress of science and the useful arts. That clause, she says, gave rise to the patent office and the trademark office.

Her line is blunt: the United States’ strong intellectual-property regime is what gave rise to Hollywood and what gave rise to Silicon Valley. Labenz is surprised it sits that high in the constitutional text. The four usual buckets are patents, trademarks, copyrights, and trade secrets. The classic trade-secret example is the Coke formula: two people know it, and they are not supposed to fly together. For software and LLMs, she says, the first three buckets do most of the work, each created and protected in a slightly different way.

Patents are a formal process. You apply for a right to exclude, the document contains claims that lay out steps, and the Patent Office scrutinizes the filing through office actions. You do not get one automatically. Copyright attaches on creation. Write a poem about your kid in a fixed medium — typed or written out — and congratulations, you have a copyright. Recite Taylor Swift lyrics from memory and you do not, because she already owns that expression. Rewrite the song about potato chips and you own the new one. Inputs and outputs of AI systems, she hints, will keep running into that distinction.

Labenz is struck that Google invented the Transformer and everyone else is using it. Could they have patented it and chosen not to? Software patents were reined in under the *Alice* line of cases, Ziniti says, but you can still patent a software process with a specific result, embodied in a particular way. She checked: the Transformer model was patented by Google. Enforcing it is a losing game. Google is the platform; it wants everybody developing. Its patent strategy is defensive. Little-known fact, she adds: every Android implementation sends about ten dollars to Microsoft, because Microsoft’s operating-system patents run so deep that the settlement of the smartphone wars put that fee into the bill of materials.

Google is one of the larger patent filers, but it uses the portfolio as a shield. Sonos sued and spent seven or eight years in a war of attrition; Google had plenty to assert back, and a great result would likely have been a cross-license. Labenz wonders whether ChatGPT is the one moment that could break the détente: OpenAI and Microsoft got ahead in a way that threatens search. Ziniti’s answer is mostly public relations. After ChatGPT, Google called a “code red” and rushed Bard and Gemini. If it then sued, the headline writes itself: old-line Google, in its last throes, tries a legal mechanic against OpenAI. On the technical side, you can design around a patent. Skip a claimed step and you may not infringe — but proving that takes a court case of its own.

Her first legal job was Yahoo in the early 2000s, when Yahoo stock was climbing two dollars a month and she paid for law school with it. Yahoo had bought Overture, which patented sponsored search — “literally the entire Google business model.” The claim said placement was “determined using bid amounts.” Google’s defense was that it also looked at relevance and brand, a plurality of factors, so it was not practicing the patent. The case settled. Patent litigation has a whole ritual called claim construction in which the parties fight over the definition of each word. Like fair use, she says, it is expensive and unpredictable. When a system is gray and turns on the meaning of specific words, litigation becomes a risk, not a clean answer.

Labenz asks whether a patent monopoly means you can simply refuse to deal. There are statutory wrinkles for generic drugs, and big firms have open-sourced patents, but the Constitution says exclusive rights. It is closer to real property: you need not turn the lot next to the county fair into a parking lot. Eminent domain exists and is in tension with that original text; a fun fact she offers is that “life, liberty, and the pursuit of happiness” replaced an earlier pairing with property. How Google commercializes the Transformer is a strategy question. Seeing OpenAI turn it into a breakout hit, she shares Labenz’s curiosity about what other gems Google is sitting on — and notes that discovery would drag secrets into public view, which is why in-house lawyers teach the New York Times test: would this Slack message survive the front page.

Copyright is exclusive too, but it is also treated as a personal right; in France it is a *droit moral*, an ethical claim by the creator to exploit the work. Law school teaches a metaphor Ziniti likes: an IP right is a bundle of sticks. A movie can be licensed separately for the plane, for streaming, for cable, for overnight, for Italy. Netflix keeps a huge licensing department writing checks; Hollywood’s business-affairs teams exist to split those sticks, plus merch and everything downstream. The important limit on that bundle is fair use.

She corrects a common picture. Fair use is not a defense in the sense that you infringed and then excuse it, even though that is how it plays out in practice. Fair use is beyond the reach of the copyright: the exclusive right never extended to that use in the first place. The test has four factors, they are not exhaustive — the statute says a court *may* consider them and can consider others — and it is a balancing test. Copyright is known as gray in a way the tax code is not, which makes sense because it governs creative expression. Factor one is the purpose and character of the use: for-profit or not, transformative or not. In technology cases, that is where the most play is, and it is the factor OpenAI will fight on.

Factor two is the nature of the copyrighted work. The Times would say of course it is creative: articles back to 1851, reporting that leads to criminal prosecutions, 50 to 100 million readers a week. OpenAI would likely admit this is not a phone book. Factor three is the amount and substantiality of the portion used — whole work versus clips, the sampling and critique cases. Factor four is the money: the effect on the market and value of the original. A one-to-one substitute, people canceling Times subscriptions because ChatGPT will spit the article, is much less likely to be fair than a complementary use. The Library of Congress, she notes, keeps a well-organized fair-use index with summaries that you can actually look up.

Big technology shifts, she says, come with lawsuits because money is moving. When she was at Yahoo, car advertisers were still in print and on TV; the share of wallet shifted, losers sued, winners got sued. Google’s keyword-advertising trademark cases all went Google’s way, and she adds a political-economy observation: judges really like using Google, and when judges like a technology, the market finds a way to rule for you. Scraping sits next door. LinkedIn’s terms of service bar automated scraping, but facts themselves are not copyrightable. *Feist v. Rural Telephone* went to the Supreme Court; the copier won, because copyright promotes creativity, not the labor of assembling a phone book. A person’s name and job title are facts. The Times’ paywall, by contrast, is a property right: like opening the carnival lot only on Sundays, the owner decides where and how people see the work.

Labenz splits the Times’ possible complaints into three layers, and Ziniti says he separated them perfectly — it is how the law and OpenAI have been thinking about the case. One: you took our stuff, mixed it with other stuff, and trained. Two: ChatGPT can output near-verbatim articles, what OpenAI has called regurgitation. Three: browsing features now fetch live pages, raising Google- and Facebook-like questions about showing the whole article, a summary, or a headline and a link. The regurgitation issue is why she thinks this case is stronger than others. The Times hired someone to elicit verbatim pieces from a deep library. Under the factors that looks like using the whole work, offering a possible substitute, taking control from the copyright owner, and not being transformative. OpenAI has called it a singular bug they are fixing. She is not an engineer, but “that seems a little weird.” She expects discovery to test the isolation story.

Reading the complaint, regurgitation is not really the core. The core, and the reason the case has caught so much attention, is training. Over the weekend she talked with a relatively famous creator who had three or four emails saying, in effect, I trained my own GPT on all your work and it has been so helpful. How is she supposed to feel? Even if it is not verbatim, even if the phrases are reworded, if it is her business advice, should there be compensation? Copyright is not a perfect fit. Music is the better analogy. Weird Al licenses the songs he parodies because the system is clear and paying is cheaper than a fight. iTunes, she notes, arrived after Napster, which Jason Calacanis told her *felt like stealing*. Generative AI does not feel like Napster, and it does not yet feel like iTunes. Something in between may emerge — a product that lets you “GPT yourself” and pay the person whose voice you wanted.

The fair-use fight will turn on whether there is a new purpose, meaning, or message. A Roger Ebert critique that quotes a few lines is the classic different purpose. OpenAI will push that GPT is a different way to use the work. The Times’ opening is that the product still has to implement measures so it will not be a substitute. OpenAI’s blog response says it is fixing the bug and has content measures that stop users from asking for copies — the Kinko’s analog, where you click that you have rights to copy. Terms of use plus technical controls both matter. Napster never hosted the files, but it knew copyrighted music was most of what moved across a 1999 university server. That knowledge is how contributory infringement attaches to a technology provider.

There is no LLM equivalent of the DMCA yet. The Digital Millennium Copyright Act gave photographers a notice-and-takedown path against GeoCities and Shutterfly. Ziniti expects something like that to emerge if OpenAI is to say, credibly, that infringement is not the intent: a way for rights holders to pull content out or to get paid, as musicians do. In the meantime the defense she predicts is almost a product slogan: it is called the general-purpose Transformer, and the intended use of this device is like a VCR for your own personal use. The names, she laughs, matter. She has told engineering teams not to put “Eye of Sauron” in the code; “Transformer” is a nicer fact for the brief.

If the parties will not deal, is there anything between OpenAI winning on fair use and the Times winning destruction of the models? The law, Ziniti says, is more adaptable than the headlines. Counts of a lawsuit are typically bifurcated; one issue can go up on appeal and come back on remand for facts. A middle ground is very possible: a rule that regurgitation is infringement unless the provider takes reasonable measures to prevent it, with “reasonable” then defined by an industry standard. They could settle. Damages could run backward — GPT-3.5 was worse on regurgitation than GPT-4, so perhaps a royalty on that era. The bundle-of-sticks frame is useful: how many sticks does the Times keep, which does it rent. What is hard is the scale. If BCG is right that generative AI is a trillion dollars of productivity, any deal struck now risks “selling the baby.” She disagrees with commentators who say this is the end of OpenAI. That would be very unlikely to draw a full injunction. The cat is out of the bag.

Injunctions themselves have four factors, and one is harm that cannot be solved with money. Very few things cannot. Copyright does have statutory damages. After Napster, the record companies went after individuals. A Minnesota woman, Jamie Thomas, had on the order of 80 songs; statutory damages were around $15,000 a work, and she ended up owing about $400,000. OpenAI has said it will indemnify users, which Napster did not, so this looks more like a clash of titans than a campaign against ordinary people. Labenz’s wife, a lawyer, has trained him to ask: are there really damages? It is hard to imagine a parade of subscribers who canceled the Times because they now have ChatGPT. The products are different, and almost nobody is using the model as a reprint machine.

He reads OpenAI as wanting a precedent that training is allowed, and the Times as wanting Congress to build a new regime. Ziniti thinks a congressional solution is very possible and that litigating with that in mind is a classic move. The Times has also lost at the Supreme Court on copyright before, so they know the court path is risky. Principle is in the case, and the law is actually decent at that question: will future journalists keep creating if they know the work will train LLMs without pay? Courts at the highest levels weigh the public good. On damages, a Times lawyer can find examples. She points to Andrew Chen’s viral thread about using GPT only for news. When behavior changes in a tech shift, someone is left holding the bag.

Her favorite copyright quote came from a guest speaker in law school, counsel to a major rapper — she cannot remember if it was Diddy or Beyoncé — who deadpanned that copyright is not about the money. “It’s about all of the money.” If the Times can get an ongoing royalty forever, of course they will try. A judge cannot simply order perpetual pay absent a statute or a deal. You can invent math: the complaint had a graphic saying the Times is X percent of Common Crawl, Common Crawl is X percent of the training data, divide a $100 billion valuation by that percentage. You can come up with a number. It will not be clean.

Modalities multiply the mess: four kinds of IP, then text, images, video, music, brand. Ziniti expects Congress or the courts to have to draw clearer lines, or the field will get a canonical case of each type. Nintendo and Sony are ferocious rights holders; a billion people have played Mario, and she has read that the character lives in children’s dreams. Disney’s purchase of Marvel was an IP acquisition, and Disney is an exploiting machine. The Times case is cleaner for being text. Images are more viscerally compelling — she has generated plenty of Marios herself — but the idea–expression dichotomy still applies. “Italian plumber” is not monopolizable. A stylish plumber in perfect suits would be more Italian, she jokes, than Mario. Graphic fair-use cases ask whether the impression is so clearly the original that nobody could have come up with it independently. A SpongeBob-inspired restaurant that kept the crab looking the same and renamed the Krusty Krab something like the Rusty Crab still lost. She expects a brand registry: ask DALL-E for Mario and you get an error; ask for Italian plumber and it ships. A self-serve UI for rights holders to claim keywords would be the product version of that, though scale and clarity are the open questions. DMCA was a congressional solution; this may be too.

OpenAI has already been more aggressive on the image side, striking a deal with Shutterstock, while on text it offers an opt-out. In August it began respecting `robots.txt`, so a site can instruct OpenAI’s crawler not to scrape, including images. That does not reach models already trained. GPT-3.5 and GPT-4 are, in her metaphor, a cake: you baked it with flour, you are not extracting the flour and the vanilla. Nutmeg might be two little shakes, and you can still taste it. That would be the Times’ argument: two little shakes, but we made the model this much more fluent. Labenz’s version is that GPT-4 was trained some 18 months earlier; you are not going to retrain it to honor a new robots file.

Commentators on Twitter have said that if they are going to figure out AGI, they can figure out how not to infringe. Ziniti’s own tests of DALL-E match a rumor that a stronger model is watching a weaker one: you start to get an Italian plumber or a princess-saving scene, then the render catches itself and throws a copyright-policy error. She believes they are pursuing that strategy in various ways. Labenz describes OpenAI’s weak-to-strong generalization work from the superalignment team — GPT-2 as a weak teacher, GPT-4 as a strong student — and his unease that one way to raise the delta is to make the student more willing to override the teacher. Ziniti takes the pedagogy: in her prompting class for lawyers she tells them the models are “fast, smart, overeager interns who have read the entire internet.” Teaching a model not to infringe is the same unsolved problem as teaching it anything else you care about.

Labenz puts the public-good question directly: the AI doctor of global humanitarian dreams is too good a promise to lose because Disney wants to protect the Marvel Cinematic Universe. Does the law already balance that way? Fair use has been a public-good balancing test since the 1800s, roughly as old as the Times itself. OpenAI’s response blog did not only claim fair use; it claimed to be good for the world, and the example it cited was a dog whose life was saved with ChatGPT. Two details interest her as a lawyer. They picked a dog, not a human, because a human anecdote would look like practicing medicine. Fitbit can infer pregnancy from extra blood volume, and every blog post about that ends with “and then I talked to my doctor,” because Fitbit is a personal fitness device, not a medical one. The law will find a way. The cat is out of the bag, the genie is out of the bottle, and unlike blockchain this is a technology she cannot stop inventing use cases for.

What surprised Labenz about 2023 is how little pushback came from licensed professions. He expected an AMA versus OpenAI over unlicensed medical advice; instead there have been tragic-comic stories of a lawyer filing ChatGPT-invented cases. Ziniti is a little surprised too, but not empty-handed. The California bar issued guidance on using ChatGPT. Chief Justice John Roberts, in his year-end report on the judiciary, said legal research will soon be “unimaginable” without AI. Her three explanations: the tools are extremely useful — on GPT-4 she can draft a demand letter that took ten, fifteen, even sixty minutes in about a minute, and she sees jaws drop in class; practicing lawyers lived through Google and are digital enough to get it; and expertise has been unbundling for years, the Ben Thompson story of a guy on the internet commanding an audience. People do worry about the job. Garbage answers to legal queries are still common enough that she is bullish on a professional in the loop. Doctors still have to see the body. “Dr. Google is not great. Maybe Dr. ChatGPT is better.”

Labenz cites a Google DeepMind paper he files under buried leads: a diagnostic chatbot versus humans, and versus humans plus AI, in a chat-only setting. Accuracy was on the order of 60 to 30, AI over human, the biggest delta he has seen in medicine — not a narrow edge you can caveat away. At two-to-one, he says, the licensing regime gets hard to defend. Ziniti notes that the California bar has a duty of efficiency: if a tool would let you represent the client better, you are more or less obligated to use it. The same logic would apply to a doctor whose goal is patient health. Somebody still has to train the systems. She has been in an online moms’ group for 18 years and has written volumes on childhood ailments; models have almost certainly ingested that folk knowledge. The intern metaphor is the practical rule she teaches: let the model draft, you sign, because responsibility does not outsource.

Open source changes the practical picture — if GPT-4 were ordered destroyed, Mixtral and other weights would still be out there, sometimes with no clean defendant — but Ziniti answers like a lawyer: there is always somebody you can sue, a host, a compute provider, a downstream company that “got it off the internet” and knew it was tainted. Open source is another way to exploit rights, freedom not free beer. Downstream, a legitimate company will have a hard time pulling a Napster-innocent-user face. Open-source compliance is already a business; startups now check dependency graphs with AI. As for what the law *should* be, she predicts a commercial “GPT yourself” product for the influencer whose voice people are already wrapping. Nightshade-style image poisoning “feels kind of illegal” to her; more likely a big lab buys a watermarking startup and turns it into the legit marketplace. Kindle, she recalls, was a lot of dealmaking by Bezos with publishers. She would rule the Times case the way Labenz framed it: on regurgitation, reasonable measures, almost a tort duty; on training, transformative, plus individual rights — register, test, get a report on how often people ask about you. Google’s transparency report, in her recollection, ran on the order of seven billion takedowns in a year. Anthropic’s influence functions are the closest research to a Spotify-like split, and they are not close: small models look like keyword matching, larger ones show semantic influence, and scoring every ChatGPT query is probably overwhelming. Labenz adds that artists already get thin per-play rates on Spotify, OpenAI is still on the order of a billion dollars of revenue, and a rev-share spread across the corpus might not feed anyone — while Taylor Swift made a billion on tour and once brought iTunes to its knees over trial-period pay.

GC AI, at getgc.ai, is in private beta. It sits on GPT in a chat modality, but she has stripped the “I am not a lawyer, consult a lawyer” hedge because her users *are* lawyers, and she has prompted it to speak like a general counsel: bottom line up front, not a golden retriever. The Wikipedia joke she cites is that GPT is a golden retriever, and golden retrievers are not good lawyers. The rest is vertical workflow: Slack, where the legal questions actually arrive, and Google Docs, aimed at in-house counsel at tech companies. She hopes to launch in a month or two. Until a DMCA-for-models or a real compensation rail exists, her usable advice from the hour is narrower: check `robots.txt`, do not outsource the signature, and watch for a market that lets a creator productize her own GPT rather than rage or celebrate into a vacuum.