Why generative-AI tools fail in R&D research

View profile for Yiannis Kiachopoulos

CEO, Co-Founder at Causaly

Most generative-AI tools — ChatGPT, Claude, Gemini — excel when the data is already containerized. Give them a spreadsheet, a protocol, a dataset, and they analyze brilliantly. But in R&D, most knowledge isn’t in a dataset. It lives in PDFs, methods sections, toxicology tables, ELNs, and internal reports. Here, generic LLMs fail. They can’t index full-text papers, they can’t retrieve evidence across 200-page documents, and they can’t generate traceable citations reliably. That’s why purpose-built scientific retrieval and reasoning systems matter. Real research requires auditability, grounding, and governed multi-agent reasoning — not just chat.

Ivan Krstic

Senior Director of Portfolio Marketing | Elsevier Life Sciences

3d

Completely agree, this really highlights the core challenge of applying AI to science. Most generative models perform well when information is already structured but scientific knowledge rarely comes that way. In research, insight hides in unstructured evidence: experimental methods, results tables, and full-text papers that demand both context and traceability. That’s why the real opportunity isn’t in general chat models, but in domain-specialized systems that can understand, retrieve, and reason across the scientific record, while providing transparent citations and grounding in source data. That’s why we at Elsevier are focused on exactly that — building AI agents that work with scientific evidence, not around it. Our goal is to help researchers ask complex questions and receive auditable, evidence-based responses that support real decision-making in R&D.

Tommy Doyle

Global Head of Corporate & Law GLG

3d

110% agree Yiannis. Have wrestled with this challenge at some of the biggest STM providers and data aggregators. Glad you are raising awareness (and going after it!).

Vitalii Honcharenko

Software Development Services at Yael Acceptic | AI/ML/NLP, Web & Mobile, IoT, E-commerce/learning, Fin/Insur/PropTech, Logistics

3d

Absolutely, context and traceability are everything in R&D.

Like
Reply
Chris Kachris

Assistant Professor, Edge AI, Accelerated Computing

1d

Context optical compression like deepseek ocr could be a game changer on this domain https://arxiv.org/html/2510.18234v1

Like
Reply
See more comments

To view or add a comment, sign in

Explore content categories