Artificial intelligence has become an integral part of research. Using AI algorithms, researchers fold proteins, search for new drugs, study climate change, and analyze acoustic data from the underwater world to track marine mammals. There are virtually no limits to research using AI. But can we also conduct research in collaboration with AI? Can AI actually do science?
This question inevitably brings to mind the new generation of generative AI, which we encounter almost daily in the form of large language models like OpenAI’s GPT or Google’s Gemini. For example, anyone experiencing a problem with their internet connection today is usually directed to the provider’s chatbot, which attempts to resolve the issue. Couldn’t we simply entrust such chatbots with scientific tasks—along the lines of, “ChatGPT, solve my research problem!”?
Anyone who tries this quickly runs into limitations. Language models like GPT are, in a sense, generalists. During their training, they processed a vast amount of text from a wide variety of sources and fields, mostly without any specific qualitative or thematic pre-selection. You can therefore talk to such chat models about almost anything, whether it’s recipes or medieval courtly love poetry. However, the models lack the depth of understanding required for scholarly engagement with a topic—they are not experts.
There is a second, serious problem. Scholarly work is characterized, among other things, by verifiability. Scholarly publishing plays a particularly important role in this regard. Quality control mechanisms ensure that the results published in books or scientific journals are considered sufficiently substantiated—that is, true in a certain sense—and can thus serve as a basis for further research. Language models, however, at least so far, have no internal criterion for truth; they are merely trained to generate plausible language patterns. The statements produced in this way are usually factually correct, but there is no guarantee of this. Occasionally, the models “hallucinate”—that is, they make statements that sound plausible but cannot be substantiated.
The current language models are therefore only of very limited use in supporting research. We at the Max Planck Institute for Geanthropology are working to change that and are developing an AI assistant designed to support our institute’s staff in their future research. To this end, we’ve developed our own language model—GeaCop. The name stands for “Geanthropology Cooperation Partner.” During its training, GeaCop analyzed thousands of scientific publications from fields relevant to our work and has become a true expert in the process. Let us explain how a language model works and how it’s trained. Maybe you’d like to pepper GeaCop with your questions or test its limits. GeaCop may know almost everything about the Anthropocene, but does it also know a good recipe for strawberry cake?
Given the flood of new scientific findings, even established experts in a field can’t know everything—and the same goes for GeaCop. That’s why we’ve paired GeaCop with a partner—Kantropos. When users submit queries to GeaCop, Kantropos first searches a vast collection of scientific literature for information that might be useful in answering the question. These snippets of information are then forwarded along with the question to GeaCop, which processes them into a meaningful answer. This principle is known as RAG—Retrieval of information to support Augmented Generation. Thanks to Kantropos, GeaCop can not only access scientific literature but also return it to users along with its answer. This allows users to verify GeaCop’s answers at any time, which solves the problem of hallucinations. However, you can also delve deeper into the literature in a targeted manner, with GeaCop providing support once again. You can learn exactly how this works from us, or you can simply engage in a scientific dialogue with GeaCop and Kantropos—perhaps on a topic you’ve published on yourself.
GeaCop and Kantropos are far from being able to solve scientific problems on their own. It is the dialogue between our researchers and GeaCop and Kantropos that fosters the emergence of new scientific insights—conducting research together with AI: We don’t just think this is possible; we’re already doing it.