Beyond the Chatbot: AI as a Scientific Collaborator
1 October 2026
That is the idea behind Collaborative Human-AI Dialectics for Scientific Research, a new research project led by Dr Bradley P. Allen of the Intelligent Data Engineering Lab (INDElab) at the University of Amsterdam. In collaboration with the Alliance for Data Science and AI (ADSA), Vrije Universiteit Amsterdam and Old Dominion University, the project uses a system called Elenchus to explore whether AI can support scientific work by challenging researchers to make their reasoning clearer, more precise and more consistent. The project is supported by the Alfred P. Sloan Foundation.
Making scientific knowledge computer-readable
Scientists often need to turn specialist knowledge into formal descriptions that computers can work with. For example, a biodiversity researcher may need to define exactly what counts as a biome, how it relates to habitats and species, and which exceptions matter. These structured definitions are called ontologies. They help researchers connect datasets, compare findings and make research more reproducible.
Scientific concepts are rarely as neat as they first appear. Definitions may have hidden assumptions, overlap with one another, or lead to conclusions that do not quite fit together. Elenchus is designed to uncover those weak spots. An expert enters a claim or definition, and the AI can identify a possible tension: a point where two statements may conflict, or where a conclusion may not follow from what has been said before. The expert, however, remains in control. They can reject the challenge, revise their claim or explain why the apparent conflict is not a problem. Only the decisions endorsed by the expert become part of the final knowledge record.
The approach takes its name from the Socratic method: learning through careful questioning and challenge. But the researchers aren’t assuming that more disagreement will automatically lead to better science. It will test that claim.
Putting the approach to the test
Thirty researchers in biodiversity will complete comparable knowledge-formalisation tasks in two different ways: once using a conventional free-form AI chat interface and once using Elenchus’s structured dialogue. The same person, task and language model are used in both cases, allowing the researchers to focus on one key difference: the structure of the interaction. Independent experts will then assess the results without knowing which method produced them.
Designing better AI for research
The outcome could help answer a growing question in AI research: when does an AI system become more than a tool? If structured challenge improves the quality of scientific knowledge, it could inform the design of future AI systems that support research. If it does not, that result is equally valuable: it would show where an ordinary AI chat may be just as effective.
Elenchus is being developed as open-source software, so that others can inspect, use and build on the platform.