1. In molecular biology, smaller models are outperforming larger ones.

MSAPairformer, a 111-million-parameter protein language model, outperformed previous leading methods on the benchmark, ProteinGym; and GPN-Star, a 200-million-parameter genomics model, outperformed a model with 40 billion parameters.

2. Virtual cell models emerged as a new frontier in 2025, with major releases including Evo 2 from the Arc Institute, STATE, and DeepMind’s AlphaGenome.

These models aim to predict cellular responses to drugs and genetic perturbations without running wet-lab experiments, though current systems still require experimental validation.

3. Like other areas of AI, biological model development is increasingly bottlenecked on data rather than architecture.

With cofolding models now representing all structure types in the Protein Data Bank, 2025 saw a turn toward distilled datasets of AI-predicted structures and training on combined experimental data sources, expanding training sets from hundreds of thousands of entries to tens of millions.

4. AI tools that automatically generate clinical notes from patient visits saw broad adoption in 2025.

Across multiple hospital systems, physicians reported they were spending up to 83% less time writing notes, experiencing significant reductions in burnout, with one hospital system reporting a 112% return on investment.

5. The FDA authorized 258 AI medical devices in 2025, most through pathways that do not require new clinical trials.

The vast majority entered the market via device-modification pathways that rely on existing safety and efficacy evidence rather than new randomized trials, with only 2.4% of devices with clinical studies supported by randomized trial data.

6. A multi-agent AI system scored 85.5% on complex published case studies, versus 20% for unaided physicians.

Microsoft's AI Diagnostic Orchestrator, paired with OpenAI's o3, was tested on challenging cases drawn from the medical literature against physicians working without their usual tools. Multi-agent frameworks more broadly have shown diagnostic accuracy gains of 7% to over 60% over single-agent baselines.

7. AI-generated summaries now appear at the top of 84% to 92% of health-related Google searches.

Symptom and common health questions trigger an AI Overview 92% of the time, followed by treatment and condition queries. These summaries are now a routine feature of health information searches, shaping the initial interpretation of users' questions.

8. Ethics discussion in medical AI publications more than doubled in 2025, but the conversation is narrow.

Governance dominates the discourse, while algorithm accountability, biosecurity, and global health equity remain underexplored.

9. Research interest in medical digital twins is growing fast, and where trials exist, early results are promising.

In a randomized trial of 150 diabetes patients, 71% achieved healthy blood sugar levels over one year while safely reducing their medications.

Support the Stanford Institute for Human-Centered AI (HAI) in our mission to advance ethical and impactful advancements in artificial intelligence.

Your support helps foster research, education, policy, and collaboration across diverse fields. Whether you are an individual, a corporation, a foundation, or a family office, together we can ensure AI serves humanity’s best interests.