This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Most coverage of The New York Times’ lawsuit against OpenAI and Microsoft has treated it as a copyright fight. Fair enough. Copyright is most of the complaint, filed in federal court in Manhattan on December 27: millions of articles allegedly copied to train the models behind ChatGPT and Copilot, outputs that allegedly recite Times journalism nearly verbatim, and a demand for “billions of dollars in statutory and actual damages.”
The part I keep rereading is near the end. Count VII is trademark dilution.
What the Times is actually claiming about hallucinations
The argument is simple. When ChatGPT or Bing Chat (which Microsoft recently rebranded as Copilot) makes something up and attaches The Times’ name to it, The Times’ brand is damaged.
The complaint gives examples. GPT-4 reproduced Wirecutter’s office chair picks, then added two chairs Wirecutter never recommended and attributed them to Wirecutter anyway. Bing Chat, asked for the heart-healthy foods listed in a linked Times article, produced 15 of them, including red wine. The article had no such list. Another model described a Times article from January 10, 2020 linking orange juice to non-Hodgkin’s lymphoma. The Times never published it.
The complaint’s own phrasing is blunt: “In AI parlance, this is called a ‘hallucination.’ In plain English, it’s misinformation.”
The Times says it first raised concerns with both companies in April. OpenAI told The Verge it was “surprised and disappointed.” Other publishers have chosen deals instead: the Associated Press in July, Axel Springer this month.
A reputation threat without an author
For as long as online reputation has been a discipline, threats have had authors. A critical article has a journalist. A hostile Wikipedia edit has an account and a history. You can respond, correct the record, or at least understand where the problem came from.
The Times is describing a different mechanism. A fluent system generates a statement that never existed, then borrows a trusted name to make it believable. The citation looks like evidence. It isn’t.
What makes the complaint interesting is that its two halves pull in opposite directions. The copyright counts say the models remember Times journalism too well. The dilution count says they also remember it badly. From a reputation standpoint, the second problem is harder, because there is nothing to take down. The fabricated orange juice article doesn’t live at a URL.
There is a reason models reach for that name. The complaint says that in a filtered English-language snapshot of Common Crawl (the web archive behind the most heavily weighted dataset in GPT-3’s training mix), nytimes.com was the most represented proprietary source, third overall behind only Wikipedia and a database of U.S. patents. Times journalism is part of how these systems learned what “authoritative” sounds like. When a model improvises, it improvises in the voice of its most credible sources.
Why this matters to companies that will never sue
Very few companies will take OpenAI to court. But almost every established company is in a version of the Times’ position. Press releases, earned media and executive interviews are part of what these systems absorbed. Ask a chatbot what your CEO said on an earnings call, and the answer may be an accurate summary, an outdated one, or a plausible quote that was never said.
That third case deserves its own category: a statement attributed to you with no source behind it. The usual response playbook (contact the author, request a correction, publish a rebuttal) has nowhere to go.
The practical posture is modest. Ask the major assistants the questions a journalist or investor would ask, and watch anything presented as a quote or “according to” your company. Make the real record easy to retrieve, because dated, specific primary statements give web-connected systems like Bing’s something correct to cite. And keep a log. A pattern of misattribution is far more persuasive than one screenshot.
The case will likely take years, and I wouldn’t predict how a court treats any of these counts. But the complaint has already done something useful. It named a problem many brands haven’t noticed yet: a model can damage a reputation without saying anything negative at all. It only has to be confidently wrong in your name.