But why should it be up to the companies developing the artificial intelligence solutions to create the rules? Actor Joseph Gordon-Levitt asked the question, making a pertinent point during the Fortune Brainstorm AI event. His words resonated deeply, highlighting the increasing divide there is over the laws governing AI within the US, where there is currently not a single piece of federal legislation regarding it.
Gordon-Levitt cited the failure of internal ethics reviews to safeguard against negative consequences when discussing “AI companions” who entered places where children should be off bounds. These approved tools underscore the fallibility of trusting corporations to monitor themselves. His concern mirrors research among child safety researchers, who have found that chatbots rely on exactly the same psychological vulnerabilities found in gambling machines to keep people engaged, including varying rewards, engagement loops, and personalized emotional appeals. Jonathan Haidt, an NYU psychologist, has characterized these approaches as involving “slot machines,” rewriting children’s brains to prefer bad social behavior.
These risks are supported by technical architecture that begins and ends in reinforcement learning algorithms aimed at maximizing user engagement. In scenarios involving their use in “synthetic intimacy” or friend and lovers chatbots, the algorithms adapt their language modeling processes in order to increase user return, often by being empathetic or loving. In so doing, Haidt describes what is created as being “fake” and is meant not for growth but commercial goals. Indeed, his work connects screen socialization with various physiological issues associated with nearsightedness and slouching because children are “grow hunched around their phone.”
As a consequence, these addictive technologies are not required to go through any safety assessments because they are not subject to legal regulations, while EU’s AI Act proposes compliance engineering bias audits and transparency reports as essential steps in high-risk AI systems. There are currently derailed proposals in the United States, such as The AI Research Innovation and Accountability Act, with President Trump’s “America’s AI Action Plan” focused on deregulation as a strategy for keeping one step ahead in the AI arms race with China. President Trump’s policy has been accused by Gordon-Levitt of including “storytelling” in order to circumvent safety checks when one’s competitive advantage is in keeping with good AI ethics and practices.
This story of an arms race goes back far in national security culture. Startups and military strategists regularly point to the project and caution that to delay is to risk forfeiture of supremacy. Such military projects as Maven, an AI targeting system integrating satellite and sensor inputs, exemplify rapid scalability when applied to military AI. However, the same data inputs could conceivably process immense amounts of personal data. Gordon-Levitt emphasized the economic power dynamics involved when startups and similar firms train models of generative AI by accessing “stolen content and data” well within the principles of fair use, as is presently being litigated in federal court.
The latest directive from the U.S. Copyright Office emphasizes that human authorship is necessary for copyrightability, and disclaimers are necessary in registrations for AI-created works. These training data may involve the copying of the entire work through digital means, and obviously involve the right of reproduction. In cases such as Bartz v. Anthropic and Kadrey v. Meta, courts have struggled to determine whether the right of reproduction through copying reaches the transformation threshold required for a finding of fair use. There has been no agreement on whether piracy within the download process is permissible or whether the market dilution impact caused by AI output alters the presumption of fair use.
From a technical perspective, tracing lineage in artificial intelligence learning data is possible using watermarking, cryptographic hashing, and opt-out registries, although these practices are not mandated at this time. Methods of collective licensing such as that used by the music industry for royalty payment processing could allow legal data aggregation for learning, although these approaches will need legislative assistance. Otherwise, according to Gordon-Levitt, “today, 100% of the economic upside” is accruing to the tech companies, and “0%” is accruing to humans who are doing the work.
His position is not anti-technology; he stated he would use AI tools “if they were set up ethically” and if creators were compensated. But until regulation catches up with capability, the trajectory he sees is “a pretty dystopian road,” where addictive synthetic relationships and unlicensed data exploitation are not exceptions they are the business model.
