AI’s Newest Scientific Breakthroughs Move From Prediction to Experiment

Artificial intelligence is entering a new phase in scientific research. The most important recent advances are no longer limited to predicting protein structures, classifying images or summarizing papers. New systems are being designed to generate hypotheses, propose experiments, improve algorithms and, in some cases, help researchers test ideas in the laboratory.

One of the clearest examples is Google DeepMind’s Co-Scientist, a multi-agent system described in a peer-reviewed Nature paper published May 19, 2026. Built on Gemini, the system assigns different AI agents to generate, criticize, rank and refine scientific hypotheses. Rather than producing a single answer, it runs a competitive process in which proposed ideas are repeatedly evaluated and evolved.

The system was tested in several biomedical areas. In research on acute myeloid leukemia, it identified possible drug-repurposing candidates and combination therapies that showed selective effects in laboratory cell experiments. In liver-fibrosis research, it proposed epigenetic targets and identified compounds that demonstrated anti-fibrotic activity in human liver organoids. One of the compounds, vorinostat, is already approved for another cancer use, making it a candidate for further investigation rather than an established treatment for fibrosis.

Co-Scientist also generated a hypothesis about how certain mobile genetic elements may spread antibiotic-resistance genes among different bacterial species. Researchers reported that the AI’s proposal matched findings from an independent study that had not yet been published when the system produced its result. That convergence is notable, but it does not eliminate the need for experiments, peer review and replication.

Another strand of progress is focused on algorithms rather than biology. Google DeepMind’s AlphaEvolve uses large language models together with automated testing and evolutionary search to generate and improve computer programs. The system has been used to explore mathematical and computational problems, including more efficient algorithms for matrix multiplication and optimization tasks connected to computing infrastructure.

The broader significance is that AI can now search through far more candidate solutions than a researcher could reasonably examine manually. In principle, that makes it useful for problems where the space of possible algorithms, molecules or experimental designs is too large for conventional trial and error. In practice, its value depends on whether the proposed solution can be verified, implemented and shown to work outside the system that produced it.

Research published in Nature in 2026 illustrates both the promise and the limits of this approach. The AI Scientist, developed by Sakana AI and collaborators, was built to handle an end-to-end machine-learning research workflow: generating ideas, writing code, running experiments, analyzing results, preparing a manuscript and conducting an automated review. One AI-generated paper passed an initial review stage at a machine-learning workshop. That result demonstrates increasingly capable automation, but it should not be confused with proof that AI can independently conduct reliable science. The evaluation was limited, and the system’s outputs still require human scrutiny.

Stanford’s 2026 AI Index makes the same distinction from a wider perspective. AI now represents a growing share of scientific research output, with the report estimating that AI-related work accounts for roughly 5.8% to 8.8% of research output depending on the field. Yet the report also finds major weaknesses: leading models can perform strongly on chemistry questions while struggling to reproduce published research, and some scientific agents remain unreliable when answering domain-specific questions or executing code.

The emerging pattern is therefore less “AI discovers everything” than “AI expands the range of ideas scientists can investigate.” The strongest systems act as research collaborators, combining literature search, reasoning, coding and experimental planning. Human experts remain responsible for defining meaningful questions, checking evidence, recognizing unsafe or implausible proposals and deciding whether a result is robust enough to matter.

The next test will be whether these systems can produce discoveries that survive independent replication and translate into practical results. For now, the most credible breakthrough is methodological: AI is becoming capable of participating in the cycle of scientific discovery, while the scientific method remains essential for deciding which of its suggestions are actually true.