Ai2 Identified Five Tasks for an AI Agent in Scientific Research
Ai2 Identified Five Tasks for an AI Agent in Scientific Research
On September 1, the research organization Ai2 published a review of a discussion about systems that assist scientists. The participants identified five tasks: incorporate expert judgment, redirect the agent as the work progresses, validate different assignments in different ways, monitor the quality of the source data, and connect analysis with laboratory experiments.
The Ai2 review focused on AutoDiscovery, a program that proposes hypotheses based on scientific data and tests them through analysis. In the August lobular cancer case, an oncologist's comments narrowed its search, after which the team tested the identified signal using independent data and tumor samples. This case illustrates scientific judgment: a specialist selects the results that can lead to the next question and sets the direction for further investigation.
Scientific work changes as a project develops. An experiment produces an unexpected result, a new paper is published, or another dataset becomes available. At these points, the researcher changes the instructions, context, and tools available to the agent, which is a program that carries out a sequence of research tasks. Oncologist Kelly Paulson described this principle as follows:
“This is research, not search: the aim is to discover something new and test it.”
Record retrieval, information organization, and literature review can be specified in advance and checked against the expected result. A hypothesis about a new mechanism or an unexpected experiment requires separate analysis and replication. Abraham Flaxman described a case in which AI found an error in the algorithms used in a previously published paper. The researcher examined the issue, agreed with the finding, and asked the journal to retract the paper.
The speed of analysis depends on what the system receives, including the experimental design, data collection, and causal reasoning. According to Stephen Salerno, AI amplifies both sound research and weak assumptions affected by methodological errors. In projects involving hundreds of cell types and thousands of genes whose expression changes, an agent can compile the relevant literature and prioritize hypotheses for testing. The result of each laboratory experiment then informs the next hypothesis and the next experiment.