Episode Summary
Executive Summary: This podcast features host Jason Calacanis and legal expert Chechlia Zinetti analyzing the landmark New York Times copyright lawsuit against OpenAI. They dissect the case's core allegations—unauthorized use of NYT content for training AI and verbatim output generation—through the lens of copyright's four fair use factors. The discussion explores historical parallels (Napster-to-iTunes), potential outcomes (settlement, market licensing, or a high court ruling), and the broader implications for startups and the AI industry.
Main Topics: Core Allegations & Legal Theories (Priority: 5/5): Explaining the NYT's two-pronged claim: 1) OpenAI used ~1-2% of the NYT archive for training without consent, and 2) ChatGPT can reproduce NYT articles verbatim (Exhibit J shows 100 instances). Four Fair Use Factors Applied to AI (Priority: 5/5): Detailed walkthrough of purpose/character (transformative vs. commercial), nature of work (creative journalism), amount/substantiality (verbatim vs. derivative), and market effect (direct competition via AI). Historical Precedents & Analogies (Priority: 4/5): Comparison to Napster->iTunes, Google vs. Oracle (declaring code), Perfect 10 vs. Google (thumbnails), and 2 Live Crew vs. Roy Orbison (parody), illustrating how tech cases often lead to market-based solutions. Potential Outcomes & Industry Impact (Priority: 5/5): Three scenarios: settlement (marketplace licensing), court ruling (narrow fair use win for AI), or Congressional action. Emphasizes that no outcome will 'end AI'. Discovery Risks & Internal Communications (Priority: 3/5): Discussion of the explosive discovery phase, citing the Oracle/Google 'Lindholm email' as a cautionary tale. OpenAI's internal discussions about content use could be highly damaging. Subscription/Attribution as a Solution (Priority: 3/5): Mock proposal for a future where ChatGPT requires user authentication with NYT or Disney+ to access copyrighted material, fostering a new revenue-sharing ecosystem.
Key Arguments: NYT argues that OpenAI's training and verbatim outputs harm its subscription business and the value of its content. OpenAI likely claims transformative use: training an LLM is different from reproducing articles, akin to Google's thumbnail case. The open-source community poses a harder enforcement challenge than corporate OpenAI, as there is no single defendant to sue. Market-based licensing (like music industry mechanicals) is the most probable long-term solution, not a complete ban on AI training. The case will likely settle or result in a narrow ruling, given ChatGPT's broad public utility and the difficulty of an injunction. Discovery risk is high: internal emails could reveal OpenAI's intent and knowledge about using copyrighted content. Copyright law is 'squishy'—judges balance four non-dispositive factors, making precedent and perceived social benefit key. Subscription wall and authentication could enable a future where users pay both for AI and content access.
Data Points: NYT archive age: 1851 - The New York Times has been publishing since 1851, giving it one of the oldest and largest content libraries of any news organization. NYT content percentage in training data: 1-2% - The NYT complaint alleges that a measurable 1-2% of OpenAI's training data came from New York Times articles, a non-trivial portion. Verbatim output examples: 100 instances - Exhibit J in the lawsuit shows 100 user prompts where ChatGPT reproduced NYT articles verbatim, occasionally with minor word changes. ChatGPT user adoption: 100 million - ChatGPT reached 100 million users faster than any previous consumer internet application. OpenAI annualized revenue: $1.6 billion - OpenAI's annualized revenue is estimated at $1.6 billion, showing the economic stakes involved. Settlement prediction (Jason): Nine figures minimum - Jason predicts any settlement will be at least $100 million, plus a licensing structure for ongoing use. Injunction likelihood (Chechlia): Less than 10% - Chechlia estimates the chance of a court issuing an injunction to stop OpenAI's operations as very low due to societal benefit.
Pivotal Quotes: "There's no universe where this case would end AI. And so the result is: do we end up with a licensing scheme? Like, is this napster to iTunes, right?" — Chechlia Zinetti: She argues that the case will not kill AI but will reshape how content is licensed, similar to the music industry's evolution. "If you can create this incredible AI that's able to make images, you should be able to figure out what was the source of those images." — Jason Calacanis: He pushes back on the idea that attribution is too technically difficult for AI companies to implement. "What we've actually been asked to do by Larry and Serra is to investigate what technical alternatives exist to Java for Android and Chrome. We've been over a bunch of these and think they all suck. We conclude that we need to negotiate a license for Java." — Tim Lindholm (Google engineer): Chechlia reads the infamous 'Lindholm email' from the Oracle vs. Google case, warning that internal communications can become explosive evidence in discovery.
Implications: Founders in AI must plan for content attribution and licensing from day one, regardless of fair use arguments. Expect a market-based solution (revenue sharing) to emerge, similar to music licensing. The case will not kill AI, but it will set ground rules for training data provenance.
About This Week in Startups
Jason Calacanis covers startups, tech, markets, media, and all the hottest topics in business and technology. He also interviews the world’s greatest founders, operators, investors, and innovators.