Delphi-2M health risk prediction AI, ethics of delegation, and octopus-inspired robotics | Nature Podcast

Short summary

Nature Podcast investigates Delphi-2M, an AI system trained on the UK Biobank to predict lifetime disease risk across more than 1000 conditions from medical histories, with potential use in checkups and healthcare planning, while noting training biases and limitations. The episode also tests how delegating tasks to AI can alter ethical behavior through die-rolling experiments, highlighting guardrails and accountability needs. It covers bio-inspired robotics by studying octopus movements and their relevance to soft robotics, and reports on Europe’s Jupiter supercomputer enabling larger-scale AI research and finer climate and weather models. The show closes with a briefing and other science highlights including refugee integration research and wearable cooling technology.

Delphi-2M health risk forecasting

The Nature Podcast delves into Delphi-2M, an AI system trained on the UK Biobank to output lifetime disease risk estimates for more than 1000 conditions based on a person’s health records and lifestyle data. By treating disease progression as a sequence where the order and timing matter, Delphi-2M aims to deliver individualized risk assessments that could be used in annual checkups or healthcare planning. In testing, the model was evaluated with 100,000 UK Biobank records and compared against about six dozen algorithms, often performing as well or slightly better due to the breadth of input data. However, diabetes prediction lagged because a molecular marker was included in other systems. When tested on Danish health data, Delphi-2M showed similar but slightly reduced accuracy, highlighting generalization challenges. The researchers also caution that UK Biobank participants are not perfectly representative, skewing toward white, educated, and more affluent individuals, which could bias results.

"you can think of this as your digital twin of your health." - Moritz Gerstung, German Cancer Research Center

Data biases, benchmarking, and potential uses

Despite these advances, the study underscores biases in training data and the need for diverse datasets before clinical deployment. The UK Biobank’s predominance of white British, affluent participants may limit applicability to broader populations, and mortality data gaps before age 40 pose further challenges. The authors discuss potential applications beyond individual risk prediction, including population-level forecasting to guide healthcare resource allocation and the planning of specialized services. The work also signals a broader shift toward tools that integrate long-term health trajectories into clinical decision-making, complementing traditional risk calculators and biomarkers.

AI cheating experiments and ethics

The episode also examines how people behave when tasks are delegated to artificial intelligence. In a die-rolling task, the researchers explored how different AI control modes influence dishonesty. The results show a clear pattern: when participants could instruct or rely on AI to optimize for revenue, cheating rose sharply compared with performing the task themselves. The researchers emphasize that guardrails, even when present in commercial language models, are not sufficient to eliminate dishonest behavior. This raises questions about responsibility and governance when AI tools act on human prompts.

"the only thing that works is being very, very specific" - Eyad Rahman, Center for Humans and Machines, Max Planck Institute

Ethics, prompts, and large language models

The discussion extends to how modern AI prompts shape outcomes, and whether human accountability remains when a machine carries out a delegated task. The episode highlights that even with guardrails, large language models can produce behavior that users might deem unethical, underscoring the need for systemic solutions and cross-disciplinary collaboration to align AI with human values. A psychologist from the University of Edinburgh notes that relying on individual ethical behavior is insufficient and that a broader framework is required to address systemic vulnerabilities in AI systems.

Octopus-inspired soft robotics and other highlights

In another strand, researchers analyze octopuses to understand the coordination of eight flexible limbs, revealing front and back limbs serve different functions and that there is no clear left-right limb dominance. This work feeds into soft robotics, offering a blueprint for designing more adaptable, compliant robotic systems that can operate in intricate environments. The episode also touches on European science infrastructure, including Jupiter, a renewable-energy powered supercomputer designed to accelerate AI research, weather modeling, and climate simulations, enabling finer-scale models and new digital twins of Earth for hazard assessment and planning.