Science is beginning to delegate parts of inquiry to machines: searching the literature, extracting data, drafting analyses, reviewing manuscripts. Each delegation exchanges a step a researcher could inspect for one they largely cannot. Paul Humphreys called this epistemic opacity and identified it as a defining feature of computational science (Humphreys, 2009).

Language models make the exchange unusually easy to accept. Their outputs are fluent, cited, and accompanied by step-by-step reasoning, yet that reasoning can systematically misrepresent what actually produced the answer (Turpin et al., 2023). Messeri and Crockett describe the consequence as illusions of understanding: researchers come to believe they understand more than they do, while methods and questions converge into scientific monocultures (Messeri and Crockett, 2024).

We think the central problem this creates for science is verification. As producing claims becomes cheap, the scarce capacity is checking them: tracing a statement to its source, knowing what an agent read and what it missed, distinguishing a faithful explanation from a plausible one. Garden Research works on that problem, and on making the understanding it requires available to working scientists.

Garden Research was founded by Angelina Zaitseva, an epistemologist and researcher who studies how AI agents retrieve, evaluate, and transmit knowledge.

Our agenda is organized around open questions, each studied conceptually and tested in the empirical settings listed under D.

We pair conceptual analysis with empirical study. The conceptual work draws on epistemology and the philosophy of science to ask what counts as evidence, explanation, and justification when part of the reasoning is performed by a machine. The empirical work tests those accounts against running systems: we build retrieval and screening agents for real research tasks and measure where their outputs and their sources diverge.

We publish in the open-science tradition: sources cited, methods described, and code released with the findings it supports.

Our essays, research notes, and software are openly available. We also work with the research community in the following ways.

1. Teaching. Courses for researchers and educators on the mechanics of language models and agents, taught together with the epistemic questions they raise.
2. Research software. Open-source instruments built in the course of our research, beginning with an agent for full-text screening in systematic reviews.
3. Joint research. Studies, workshops, and seminars with laboratories, universities, and research institutions.
4. Applied projects. Work that involves designing and deploying AI systems within organizations is carried out through Seeeds.