Legal-Technical Litigation Dossier

Seattle Times & Newsday v. OpenAI & Microsoft

Auditing the prayer for Algorithmic Disgorgement (model destruction), training set contamination, and eBay v. MercExchange permanent injunction hurdles across foundation model architectures.

Disgorgement Injunction Likelihood
24.2%
Court ordered total destruction of weights
Retrain Compute Write-Off
$18.5M
~1.2M H100 GPU Hours
Statutory Damages Ceiling
$126.3B
Up to $150k / registered willful work
Expected Settlement Val.
$42.1M
Annualized multi-year archive licensing

eBay v. MercExchange 4-Factor Injunction Test 547 U.S. 388 (2006)

1. Irreparable Harm (Market Substitution & Direct Competition) 85%
LLMs quoting, summarizing, and synthetic search answering bypass publisher paywalls and ad impressions.
2. Inadequacy of Monetary Damages (Ongoing Inability to Calculate Dilution) 68%
Persistent embedded representations inside billions of weights make precise royalty partition challenging.
3. Balance of Hardships (Retrain Write-off vs. Journalistic Insolvency) 28%
Courts hesitate to issue structural death-sentences destroying commercial models worth hundreds of millions.
4. Public Interest (AI Utility vs. Free Press Viability) 55%
Public interest heavily balances preserving independent investigative journalism against AI innovation.

Technical Remedy: Algorithmic Unlearning vs. Model Disgorgement MACHINE UNLEARNING FEASIBILITY

Remedy Mechanism Technical Method Compute Cost Collateral Degradation Litigation Acceptance
Exact Algorithmic Disgorgement Weight zeroing / Full scratch re-train without Seattle Times / Newsday corpus $18.5M Zero (clean slate, but massive write-off) Favored by Plaintiffs (FTC Precedents: Everalbum, Cambridge Analytica)
Influence-Function Pruning Hessian-vector inverse weight modification targeting plaintiff tokens $1.4M ~3.2% general knowledge perplexity degradation Experimental; lacks definitive federal evidentiary validation
Negative Preference Tuning (DPO / RLHF) Targeted unlearning via negative log-likelihood on Seattle Times text $240K Susceptible to jailbreak regurgitation; latent weights preserved Likely rejected as insufficient structural disgorgement
Mandatory Compulsory License Continued deployment paired with court-ordered archive royalty pool Royalty Stream None (weights intact) Common equitable compromise under eBay factor 2

Litigation Claims Profile CAUSE OF ACTION BREAKDOWN

Plaintiffs: The Seattle Times Co., Newsday LLC Defendants: OpenAI Inc., Microsoft Corp. Forum: S.D.N.Y. (Related to NYT v. OpenAI) Prayer for Relief: Destruction of GPT models containing Plaintiffs' works

The lawsuits assert that millions of Pulitzer-prize winning investigative stories were copied without permission to train commercial models (GPT-4, Copilot). Crucially, the complaints allege DMCA §1202(b) violations for stripping Copyright Management Information (author bylines, copyright notices, terms of service) from training corpora to avoid automated infringement tracking.