The dynamic frontier of artificial intelligence

Nature Materials volume 25, page 161 (2026)Cite this article As artificial intelligence tools continue to develop, their impact is growing.Most of us have been exposed to artificial intelligence (AI) tools in the past couple of years, following the release of generative AI chatbots such as ChatGPT. We now hear of AI on a nearly daily basis in the mainstream media, from its impact on the stock market to its vast energy consumption. Scientists are engaged in the development of AI tools, the application of AI to their research questions and output, and in the development of solutions to problems raised by AI. As publishers, we are also aware of the impact of AI on science.A common endpoint of the scientific research life cycle is publication in a peer-reviewed journal. The policy of Nature Portfolio journals on the use of AI in the publishing process can be found here: https://www.nature.com/nature-portfolio/editorial-policies/ai. We expect these policies will evolve as AI continues its rapid evolution. We permit authors to use AI to edit their manuscripts; this may help to improve the clarity of their language but can also lead to undesirable homogenization of style. Other uses of AI should be declared in the Methods, with a description of the model used and how it was trained. As AI cannot be held accountable for its output, we do not allow AI to be listed as an author. We ask that peer reviewers do not upload manuscripts into generative AI tools and declare the use of any AI tools in their peer-review report.On 23 October 2025, a seminar given by Francisco Villaescusa-Navarro to a full house at the Flatiron Institute in New York City demonstrated the use of the Denario multi-agent AI platform to perform the entire research cycle, from ideation to calculation to writing a publication1. Even in this extreme case, which we note has not been peer reviewed, our present policies would not qualify AI for authorship as AI still cannot be held accountable for its output.A tremendous advance in the development of AI tools for science, as recognized by the 2024 Nobel Prize in Chemistry, was the release of the AlphaFold method to predict protein structures2. Trained on experimental protein data, AlphaFold predicts structures with quality approaching those determined using difficult and time-consuming experimental methods. AI methods to predict inorganic crystalline materials are also highly desirable. In an Article in this issue of Nature Materials, Ryotaro Okabe and collaborators describe a diffusion-based generative model focused on honeycomb and kagome lattice geometries in inorganic materials, with an eye on realizing quantum states in such structural motifs. Their generative AI model Structural Constraint Integration in a GENerative model (SCIGEN) was trained on 45,229 materials from the Materials Project database, an open-access repository of inorganic materials data3. SCIGEN predicted millions of materials that, after screening and sampling, converged to 24,473 materials. These materials were then separated depending on whether they were magnetic, and it was found that, for specific elements, the predicted magnetism matched reasonably well with the ground truth. For example, 19.1% of Nd-containing materials were predicted to be magnetic compared with a 26.61% ground truth ratio. Guided by their SCIGEN results, the authors synthesized and characterized two new materials that show magnetic behaviour.At Nature Materials, we have witnessed the impressive application of AI to varied fields, including electrocatalysis4, drug delivery5, peptide and protein design6,7,8, and nanocrystalline structural determination9. Agentic AI acting as an experimentalist in autonomous laboratories is tremendously exciting10. Nevertheless, it is unclear how far these AI tools can go and how much good they can do. During an engaging panel discussion hosted by Andrew Millis at the Flatiron Institute following the Denario seminar, one panelist commented (in this centennial year of Schrödinger’s equation) that we have not yet seen Schrödinger’s equation emerge from AI. Audience members expressed their frustration at the proliferation of AI slop and inaccuracies, noting that ultimately you cannot trust its output without substantial human oversight, suppressing their appetite for using it as a tool in their work. A recent study suggests that while AI increases a scientist’s output and impact, it narrows the scope of their work and decreases engagement with each other11.It is easy to feel lost in this wave of rapid development. If you are looking for a good place to start or to hone your understanding, we include in this issue a Review from Mingda Li and collaborators that surveys developments in AI-driven materials discovery.Adapted from the Reviewby Li and collaborators, Springer Nature Limited.Villaescusa-Navarro, F. et al. Preprint at https://doi.org/10.48550/arXiv.2510.26887 (2025).Marx, V. Nat. Methods 19, 5–10 (2022).Article CAS PubMed Google Scholar Horton, M. K. et al. Nat. Mater. 24, 1522–1532 (2025).Article CAS PubMed Google Scholar Moon, J. et al. Nat. Mater. 23, 108–115 (2024).Article CAS PubMed Google Scholar Li, B. et al. Nat. Mater. 23, 1002–1008 (2024).Article CAS PubMed Google Scholar Liu, H. et al. Nat. Mater. 24, 1295–1306 (2025).Article CAS PubMed Google Scholar Rankovic, S. et al. Nat. Mater. 24, 1635–1643 (2025).Article CAS PubMed PubMed Central Google Scholar Wang, S. et al. Nat. Mater. 24, 1644–1652 (2025).Article CAS PubMed PubMed Central Google Scholar Guo, G. et al. Nat. Mater. 24, 1726–1734 (2025).Article CAS PubMed Google Scholar Zhang, Z. et al. Nature 647, 390–396 (2025).Article CAS PubMed Google Scholar Hao, Q., Xu, F., Li, Y. & Evans, J. Nature https://doi.org/10.1038/s41586-025-09922-y (2026).Article PubMed Google Scholar Download referencesReprints and permissions The dynamic frontier of artificial intelligence. Nat. Mater. 25, 161 (2026). https://doi.org/10.1038/s41563-026-02504-xDownload citationPublished: 03 February 2026Version of record: 03 February 2026Issue date: February 2026DOI: https://doi.org/10.1038/s41563-026-02504-xAnyone you share the following link with will be able to read this content:Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative