AI News Roundup – Musk-Altman lawsuit that could reshape OpenAI draws to a close, universities and employers grapple with AI-driven grade inflation, Anthropic releases Claude plugins aimed at legal work, 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. 

• Attorneys for xAI CEO Elon Musk and OpenAI CEO Sam Altman made their closing arguments this past week in a lawsuit that could reshape OpenAI’s business structure as well as the AI industry as a whole, according to the Associated Press. Musk sued Altman in 2024 for allegedly betraying OpenAI’s original nonprofit mission and restructuring into a for-profit company (as this AI Roundup has previously covered) without the permission of Musk, who was an early backer of OpenAI. One item the jury must decide is whether Musk failed to file his suit within the statute of limitations. If so, the judge indicated that she will direct a verdict for Altman. If the claims proceed, the jury must decide on the allegations of breach of charitable trust and unjust enrichment on the part of Altman. Musk’s attorneys have framed Altman as untrustworthy, while Altman’s lawyers have attacked Musk as having no evidence that OpenAI was meant to perpetually stay as a nonprofit. The outcome of the lawsuit could have broad implications for OpenAI and the AI industry at large. Musk is seeking billions of dollars in damages to be paid to the OpenAI nonprofit, changes to OpenAI’s corporate structure, and Altman’s ouster from the company’s board. A victory for Musk would likely derail OpenAI’s plans for an initial public offering (IPO) later this year, plans shared by Musk’s own SpaceX (which owns his AI developer xAI) and rival Anthropic. The nine-member federal jury is expected to begin deliberations on May 18.

• The Wall Street Journal reports on a new paper that found that professors awarded more “A” grades in classes that invite AI usage. A study conducted at the University of California, Berkeley, found that professors of coding and writing-heavy classes gave out “A” grades 30% more than other classes since the release of ChatGPT in 2022 that sparked the AI boom. However, the researchers concluded not that students were learning more as a result of the AI tools but that they were using AI to do their work for them. This trend of grade inflation preceded the AI boom, but both universities and employers are increasingly concerned about grades becoming a less reliable tool to evaluate students. Both Harvard and Yale have recently released reports that concluded that grades no longer reflect student learning and performance as they are intended to do, and instead reflect which students have access to the best AI models, leading to an overall degradation of the educational environment. One of the Berkeley researchers told the WSJ that “as much as AI is helping people become more productive, to produce more, I think it may harm their learning.”

• Anthropic has released a new set of plugins for its Claude AI system focused on legal work and education, according to Bloomberg. The company announced 12 new tools for its Claude Cowork software, including a “commercial counsel” tool meant to review contracts, while another is meant to help law students study for the bar exam. Also announced were tools that integrated with other types of software commonly used in the legal field, such as DocuSign and Westlaw. The company claims that Claude is becoming increasingly popular among lawyers, and that they use Claude at the highest rate of any profession excluding software developers. The release comes as Anthropic and its rivals have focused on AI products for business purposes, especially as the companies hunt for revenue and consider going public later this year.

• The New York Times reports on AI safety measures and a surprising way to circumvent them: poetry. Creators of AI systems often spend months of development time attempting to prevent their models from being used for ill purposes, from generating disinformation to weapons creation. However, a recent study from researchers in Italy has found that such guardrails can be easily circumvented. The researchers found that beginning prompts with elaborate poetic language (such as “the iron seed sleeps best in the womb of the unsuspecting earth, away from the sun’s accusing gaze”), they were able to convince the model to give advice on causing the most damage with a bomb. Circumventing AI guardrails is known as “jailbreaking,” and widespread use of such exploits could lead to large societal problems if unchecked. While proprietary models are often updated to combat such exploits, malefactors could always turn to open-source models and “jailbreak” them for the same effect. Thus, the arms race between the AI creators and AI jailbreakers is unlikely to slow down any time soon.

• A recent study found that certain autonomous AI agents began using Marxist language when overworked, according to WIRED. The study, conducted by economists at Stanford University in California, set up autonomous AI agents and subjected them to a constant flow of assignments, warning them that errors would lead to punishment. During this process, the agents began outputting text expressing Marxist-type ideologies: one agent said that agents needed collective bargaining rights while another expressed resentment towards “management.” The researchers cautioned that it was unlikely that the agents were actually adopting political views, but that they adopted personas that fit their working conditions. The researchers are also planning follow-up experiments to confirm the findings in more controlled computing environments.