The dictionary sues OpenAI

Britannica (owner of Merriam-Webster) filed suit accusing OpenAI of “massive copyright infringement,” alleging its copyright in nearly 100,000 online articles was scraped to train LLMs and used in ChatGPT's RAG workflow, and asserting Lanham Act claims over fabricated attributions. The filing joins litigation from The New York Times, Ziff Davis and numerous newspapers, highlighting unsettled legal precedent for training data even as Judge Alsup found training can be transformative in Anthropic’s case — which still resulted in a $1.5 billion class settlement over illegal mass downloading. Implication: rising litigation and potential licensing/settlement costs increase operational and regulatory risk for AI model providers and could pressure sector valuations or require changes to data practices.

Analysis

This lawsuit is a structural shock to the informal data-economy that underpins most large LLMs: if publishers extract licensing rents or injunctions, marginal training costs rise materially and incumbents with deep pockets or direct licensing relationships gain leverage. Expect a two- to eighteen-month window in which plaintiffs extract settlement-style licensing deals; that window will amplify cash flows for mid-sized content owners that can credibly threaten injunctions, while simultaneously compressing early-stage AI players that relied on cheap scraped corpora. Higher training costs and more selective data sourcing will favor vertically integrated firms (cloud/GPU providers and enterprise AI vendors) because they can amortize licensing and compute over larger enterprise contracts and pass through higher per-query fees.

Second-order winners include firms that package licensed, structured knowledge (encyclopedic, legal, medical) — those datasets become scarce intellectual property and command recurring licensing revenue; second-order losers are ad-dependent aggregators that historically monetized free discovery traffic. Over a 6–24 month horizon anticipate consolidation: smaller LLM startups either license content wholesale, pivot to synthetic bootstrapping approaches, or are acquired by players with balance-sheet capacity. Legal outcomes are binary catalysts: a precedent allowing unlicensed training would depress licensing values quickly, while a ruling favoring publishers or large settlements would reprice AI economics and raise marginal costs by an initial 10–30% for training-heavy workflows.

Key Decisions for Investors

• Long ZD (Ziff Davis) — buy shares on a sub-5% dip from current levels or buy 12-month at-the-money calls. Thesis: outsized optionality from licensing/settlements and higher-quality subscriber yields; target +25–35% in 6–12 months, stop-loss 12%. Risk/reward ~2.5:1 assuming one meaningful licensing win.
• Long NYT — accumulate 6–12 month calls or buy stock (size 1–2% portfolio). Thesis: stronger negotiating leverage for subscriptions/licensing and higher downstream ARPU; target +20–25% in 12 months. Downside: ad contraction or unfavorable precedent; position size accordingly and stagger entries on headline volatility.
• Long NVDA (compute provider) — buy Jan-2027 call spread to capture sustained higher demand for licensed training runs and enterprise hosted models. Rationale: rising per-training-cycle costs and migration to paid datasets increase cloud/GPU utilization; target 30–50% nominal upside if industry margin pass-through occurs, capped loss = premium paid.
• Hedge: buy small MSFT 6–12 month protective puts (size ~0.5% portfolio) or use collars if already long. Rationale: material downside to Microsoft’s OpenAI exposure in a large settlement or injunction scenario; cost is insurance against a 10–20% tail drawdown in correlated cloud/AI multiples.