Seattle Times and Newsday filed separate suits against OpenAI and Microsoft, claiming the companies used their articles to train large language models without permission. The complaints seek damages and court orders to stop further use of the publishers' content.
The filings were first reported on Grok AI News.
What the Complaints Allege
Both papers state that OpenAI scraped and ingested full articles to build GPT models. They claim this constitutes copyright infringement and unfair competition. The suits ask for monetary compensation and an injunction blocking continued training on the same material.
Microsoft faces secondary liability claims because it invested in and hosts OpenAI infrastructure.
Pattern of Publisher Actions
These cases join earlier suits from The New York Times, The Intercept, and several European outlets. Most allege the same core issue: mass ingestion of paywalled or copyrighted text for model training without licenses.
Early 2024 filings focused on direct copying of articles. Later suits, including these, add claims about lost licensing revenue and dilution of subscription value.
Technical Context of Training Data Use
Large language models require billions of tokens during pre-training. Public web crawls supply the bulk of that data. When publishers remove robots.txt allowances or demand opt-outs, developers must either filter datasets or negotiate licenses.
Current open datasets such as Common Crawl already contain exclusions for major news domains. Further restrictions would increase the cost of curating compliant training corpora.
Legal and Practical Risks
Developers relying on unfiltered web data face rising exposure. A successful injunction could require retraining or filtering existing models. Insurance costs for AI startups have already climbed in response to similar litigation.
Companies that license data directly, such as those using Scale AI or News Corp partnerships, reduce this risk but pay higher per-token fees.
| Approach | Cost per million tokens | Infringement risk | Speed to scale |
|---|---|---|---|
| Unfiltered web crawl | Near zero | High | Fast |
| Licensed news corpus | $5–15 | Low | Medium |
| Synthetic data only | $2–8 | Low | Slow |
Who This Affects Most
Teams building general-purpose models or retrieval systems that index news archives should review their data pipelines. Niche vertical models trained on technical or non-news domains face lower immediate exposure.
Publishers weighing their own licensing deals can use these suits as precedent when negotiating rates.
Outlook for Training Practices
Courts have not yet ruled on fair-use defenses in these specific cases. Outcomes will likely hinge on whether training constitutes transformative use and whether market harm can be quantified.
Developers should track licensing marketplaces and dataset provenance tools that emerged after the New York Times suit.
Bottom line: The new lawsuits raise the compliance bar for any team ingesting news content at scale.
Continued litigation will push the industry toward paid data agreements or narrower training scopes rather than blanket web scraping.
Top comments (0)