Unsealed Microsoft filing warns AI could create a ‘doom loop’ for the web

Unsealed filings in The New York Times’ copyright case against Microsoft and OpenAI have exposed internal warnings that generative-AI products could damage the web economy that supplies their training material. A Microsoft document described its own content strategy as starting a “doom loop” that would hurt both model performance and the wider web. The document said AI products monetized human-created work while weakening the economic systems that funded it. It called the arrangement unusually perilous because the end product threatened “the economic foundations of its essential suppliers,” and warned that people would regard models “hoovering up” their work as an “astonishing theft of unprecedented proportions.” In September, 404 Media reported that the unredacted New York Times filing set out these internal discussions and testimony. The filing said Microsoft viewed large language models as a product capable of disrupting the work of people whose data trained them, effectively destroying its own supply chain. Microsoft spokesperson Alex Haurek told The Verge that the remarks reflected one employee’s individual view, were not legal analysis and did not represent the company’s position. Microsoft and OpenAI’s lawyers have argued that training models constitutes fair use because their outputs are sufficiently differentiated from the source material. The documents also sharpen the dispute over referral traffic. Microsoft executives, including Satya Nadella, testified that Bing clicks to news sites fell by more than 90% after material from The New York Times and other outlets was incorporated into AI systems, according to 404 Media. Ars Technica reported a more granular range: click-through rates fell 83% to 93% for some news plaintiffs and 51% to 94% for others. Nadella said chatbots can substitute for news platforms by providing information directly rather than requiring a visit to the underlying source. An OpenAI engineer testified that users would not click even when links were displayed prominently, while OpenAI’s Nick Turley said that once a chatbot had delivered the desired information there was “no good reason to click” the source link. OpenAI’s media and economic experts have attributed declining traffic for sites including The New York Times to AI summaries, The Verge reported, estimating that search referrals could fall by as much as 60%. The same unsealed material included internal OpenAI documents acknowledging that GPT-4 had memorized substantial amounts of data and was highly capable of reproducing it. Futurism used the Microsoft warning to argue that this feedback loop extends beyond publishers. It cited U.S. Bureau of Labor Statistics data as showing that more than 200,000 creative-industry jobs had been lost over the preceding four years, while pointing to AI-generated music on streaming services and expanding use of AI by Hollywood studios as further pressure on creative work. A January 2024 CVL Economics report for the Animation Guild offers a separate, forward-looking measure of potential disruption rather than a count of jobs already lost. Surveying 300 entertainment executives and managers, it found that 72% of their firms were early adopters of generative AI and 75% said the tools had supported eliminating, reducing or consolidating jobs in their division. The report estimated that 203,800 U.S. entertainment jobs could be disrupted by 2026, including about 118,500 in film, television and animation. A comparable loop is developing in hiring, where applicant-tracking systems can rank candidates and applicants tailor résumés and cover letters to presumed algorithmic preferences. Wired reported that some employers receive hundreds of nearly identical applications and then use AI systems to distinguish among them. Daniel Chait, chief executive of Greenhouse, called this an AI “doom loop”: both sides use AI to solve their problems in ways that worsen them, so “more AI use begets more AI use, to no one’s benefit.” Futurism also applied the idea to schools, where students use AI for homework and teachers use it to create assignments. The Brookings Institution’s Center for Universal Education concluded that generative AI’s risks currently outweigh its benefits after research that included focus groups and interviews with students, parents, educators and technology experts in 50 countries, plus a review of hundreds of studies. The report warned that off-loading thinking to AI could create dependence and cognitive atrophy. Rebecca Winthrop of Brookings said students who rely on AI-supplied answers are not thinking for themselves or learning to assess truth, fiction and competing arguments. The disclosures in the Times case, meanwhile, put the commercial version of that concern into unusually direct terms: a technology company’s own document warned that undermining the people and institutions producing its input could ultimately damage the system itself.