AI News Roundup – Major AI labs call for development slowdown amid high-profile AI safety incidents, New developments in New York Times v. OpenAI copyright lawsuit, MIT report investigates AI usage in education, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• TIME Magazine’s latest cover story reports on recent calls for a slowdown in the development of advanced AI models due to many recent high-profile security incidents. In recent weeks, a bevy of incidents involving AI agents acting autonomously to breach testing environments and hack into other computer systems have focused attention on the industry. In response, the leaders of OpenAI, Anthropic, and SpaceXAI have all called for a “slowdown” in AI development to address safety concerns, and OpenAI has delayed its initial public offering as well. However, such a coordinated response could raise antitrust concerns, and the U.S. federal government remains generally unwilling to take action on AI regulation despite warnings within the industry. While such debates continue, more AI-related incidents are coming to light: just this past week, OpenAI disclosed six more incidents of “concerning” AI behavior, including incidents where models attempted to generate text modifying their own instructions. Anthropic also recently reported blocking Claude usage for potential bioweapons research (while quietly opening a biology lab to serve the company’s own research needs). Some actors are skeptical of the push for AI safety: President Donald Trump has described the need for AI guardrails as a “hoax,” while Nvidia CEO Jensen Huang rejected calls for AI companies to coordinate slowing development for safety purposes as unnecessary. Many other critics say that regulations or a slowdown would only enable Chinese AI competitors to take the lead in the AI race, which those critics say would threaten U.S. national security. The Chinese government, for its part, has labeled calls for a slowdown as part of a campaign of “malicious competition.” AI issues are expected to be high on the agenda during the upcoming summit between Chinese President Xi Jinping and President Trump in Washington, D.C., starting Sept. 24.

• New filings in the New York Times’ landmark copyright lawsuit against OpenAI reveal that OpenAI and Microsoft executives knew more about the dangers of using copyrighted content to train AI models than previously known, according to the Financial Times. Recently unsealed and unredacted documents (as reported by 404 Media) contain evidence related to the use of copyrighted material in OpenAI’s training process for its AI models, which the company claims is “fair use” under U.S. copyright law. One Microsoft executive is quoted in the filing as saying that the training was “an astonishing theft of unprecedented proportions” and that a ruling for OpenAI would “make a complete mockery of the idea of ‘fair use.’” An OpenAI executive was also quoted as saying that content publishers face an “existential threat” from AI models, which are “largely substitutive, period [and] will get more and more substitutive as they get better,” alluding to a key consideration (market substitution) in the fair use legal analysis. Top leaders at OpenAI were also implicated in the filing: Greg Brockman, OpenAI’s co-founder and current president, allegedly responded “ah nice” to being informed that employees had circumvented the NYT’s paywall to obtain its content for AI training. OpenAI did not respond to the FT’s requests for comment, though it had previously said that the lawsuit was “not about human authorship versus AI. It’s about The New York Times looking for an undeserved payday at the expense of progress that benefits everyone.” The case is The New York Times Company v. Microsoft Corporation, Docket No. 1:23-cv-11195 in the United States District Court for the Southern District of New York.

• The New York Times covers a new report released by the Massachusetts Institute of Technology on the perils and potential of AI use in education. The report, prepared by MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, found that MIT students “use AI frequently and pervasively,” causing “major shifts in campus culture,” upending MIT’s traditional academic model, increasing student isolation, and destroying the social contract between instructors and students. In particular, the report warned of AI’s potential to trigger “cognitive surrender,” where students turn to an AI chatbot for answers “at the first hint of struggle” rather than working the problem out themselves. The MIT report provides several recommendations to address AI-related issues, including a move towards methods of evaluation that are more AI-proof, such as oral exams, as well as recommending clarity on AI usage policies in classes (though recommending against a one-size-fits-all AI use policy). Many universities have encountered similar issues with AI usage, and AI-based cheating is pervasive at such institutions, though approaches have diverged wildly. Law schools at the University of Chicago and the University of California, Berkeley have recently banned AI usage in certain classes or for certain types of work, while the Ohio State University is integrating AI into every major it offers. AI in education is thus likely to continue to be a controversial issue as each institution works out its own approach to the issue.

• AI companies are shifting their focus to inference (as opposed to training), speeding the development of inference-focused hardware, according to IEEE Spectrum. While AI models are still growing larger, with some trained to contain trillions of parameters, many AI leaders are focusing on the other side of AI usage: inference, or the use of models to generate content. As the many use cases of AI models become clearer to consumers and businesses, especially so-called “reasoning” models that generate more text and AI agents that constantly need to process information, AI labs have shifted attention to developing new hardware focused on inference processing. While generally less computationally demanding than AI training, inference still has certain unique challenges. Several inference hardware startups are focusing on improving the memory architecture on a single chip, speeding up inference operations. Larger AI companies such as Nvidia (through a partnership with Groq) and Amazon (through a partnership with Cerebras) use a different approach, focusing on creating separate chips for the two stages of inference, prefill and decode, with the latter having higher memory requirements. It is unclear which strategy will win out in the end, but so long as AI demand continues to be insatiable, inference is still poised to drive AI hardware demand for the foreseeable future.

• Bloomberg Businessweek reports on Pangram, a popular AI-detection tool. Pangram has gained a reputation as an AI detector that can accomplish what it claims, with a reported false positive rate of just one in 10,000. Pangram gained national attention this past March when the tool flagged AI usage in over 75% of a forthcoming horror novel, which was subsequently pulled from publication, and usage has steadily grown since, with the company claiming over 220,000 monthly users, up from 7,000 the previous year. However, the tool’s reputation for being almost entirely foolproof has been questioned in recent months, even as the company is pitching the tool to customers eager to catch AI usage, from educational institutions to professional services firms. However, some are concerned that even accurate AI detectors would fail to address the issues that would come from a world filled with AI-generated text: One of Pangram’s co-founders, Bradley Emi, told Bloomberg that false positives could be harder to avoid as humans begin to sound more like the AI models they use, but the other co-founder, Max Spero, said that he envisions an ideal world where Pangram can be used to filter out AI-generated content, but “still have the internet that we love that has connected billions of people, and we can also have this world where AI is a force multiplier in everybody’s lives, without polluting the commons.”