Japan's first comprehensive AI law puts "promotion" rather than regulation at its core. A new study in Risk Sciences traces this choice to Japan's distinctive social conditions – low AI adoption, low public trust, an ageing population and ageing corporate IT systems – and finds that promotional measures have boosted individual use and corporate trials, but have not yet driven deep integration of AI inside organisations.

A new Caltech-led paper demonstrates that an AI program—specifically a type of neural network called a recurrent neural operator (RNO)—can be used to predict when tipping points occur in simple models, including models of a climate-change system and simulations of turbulent airflow over a plane's wing. 

 

Researchers present RFdiffusion3 (RFD3), a new generative AI system that designs proteins and nucleic acids atom by atom, which could enable the ability to engineer novel biomolecules with precise, complex functions. Recent advances in generative deep learning have made it increasingly possible to design proteins from scratch with specific functions. However, most existing approaches work at the amino acid level, which provides enough detail for designing protein structures and interactions but not for precisely engineering enzyme active sites or other functional regions at the atomic scale. Newer methods have introduced atom-level modeling, but they still struggle to design detailed interactions between proteins and nonprotein molecules such as small-molecule ligands, nucleic acids, or catalytic groups. As a result, current approaches remain limited in their ability to create proteins, particularly functional RNA and DNA molecules, with precisely engineered active sites. Jasper Butcher and colleagues developed RFD3, a generative AI system that designs proteins and nucleic acids by modeling individual atoms, enabling precise control over the molecular interactions that give biomolecules their functions. Trained on a broad collection of biological structures, RFD3 can construct complex functional sites, design RNA molecules, and create enzymes that coordinate multiple catalytic components. In computational tests, Butcher et al. show that the model outperforms previous approaches in designing protein structures and organizing catalytic sites. Moreover, laboratory experiments demonstrated its ability to produce functional enzymes from scratch, including enzymes that break down paraoxon, a toxic organophosphate, and a protease engineered to cleave a peptide associated with transthyretin amyloidosis.

In a Policy Forum, Pamela Samuelson and colleagues discuss whether generative AI causes “market dilution,” i.e., crowding human-authors works out of markets and weakening incentives for people to create original works. Generative AI holds considerable promise for advancing scientific research, but its future depends in part on unresolved questions about copyright law. Because foundation models are trained on vast collections of material, including copyrighted books, images, and other human-authored works, more than 100 active US lawsuits are examining whether this practice is lawful. Although two federal courts have ruled that such training qualifies as “fair use,” a legal doctrine that permits certain uses of copyrighted material without permission, the broader legal dispute remains unsettled. One emerging argument, known as market dilution, proposes that AI-generated content could undermine entire markets for human-created works, even when it does not directly harm any copyrighted work. Here, Samuelson et al. question the conceptual basis of this theory and argue that it oversimplifies the relationship between human and AI creativity. According to Samuelson et al., any assessment of market dilution must account for how AI is used, the extent of human contribution, and the specific creators and markets affected. Claims of harm must be evaluated through rigorous empirical research focused on particular AI systems and creative markets, rather than broad assumptions about AI’s impact on human authorship. The authors suggest that collaboration between AI developers and researchers, alongside longitudinal data and surveys of creative practices, could help clarify these dynamics. If evidence eventually establishes that AI harms particular creative markets, they argue that legislatures, not courts resolving individual copyright disputes, are better equipped to develop targeted policies, such as licensing or funding schemes. They warn that court rulings restricting the use of copyrighted works to train AI models could create widespread legal uncertainty and significantly affect scientific researchers and developers who rely on these technologies. “If courts decide that developers of generative-AI models are likely diluting markets for human-authored works, rulings against fair uses may substantially raise the liability risks for all developers and users of generative AI in ways that are difficult to predict,” write the authors. “This could have considerable impacts on scientist users and builders of generative-AI systems, including open-source and domain-specific models.”