Artificial intelligence has become an integral part of research. Researchers use AI algorithms to fold proteins, search for new drugs, study climate change or analyze acoustic data from the underwater world to track marine mammals. There are virtually no limits to research with AI. But is it possible to conduct research with AI? Can AI actually do science?
This question immediately brings to mind the new generation of generative AI, which we encounter almost daily in the form of large language models such as OpenAI's GPT or Google's Gemini. For example, anyone who has a problem with their internet connection today is usually referred to a chatbot from the provider, which attempts to solve the problem. Couldn't such chatbots simply be entrusted with scientific tasks - along the lines of: "ChatGPT, solve my research problem!"?
Anyone trying to do this will quickly reach their limits. Language models like GPT are universalists in a sense. During their training, they have processed a large amount of text from a wide variety of sources and areas, usually without any particular qualitative or thematic pre-selection. You can therefore talk about almost anything with such chat models, be it cooking recipes or medieval minstrelsy. However, the models lack the depth of penetration that is necessary for a scientific discussion of a topic - they are not experts.
There is a second, more serious problem. Scientific work is characterized, among other things, by verifiability. Scientific publishing plays a special role here. Quality control mechanisms ensure that the results published in books or scientific journals are considered sufficiently reliable - in a sense, true - and can therefore serve as a basis for further research. However, language models do not have an internal truth criterion, at least so far; they are merely trained to generate plausible language patterns. The statements generated in this way are usually factually correct, but there is no guarantee of this. Sometimes the models "hallucinate", i.e. they make plausible-sounding statements that cannot be substantiated.
The existing language models are therefore only suitable for supporting research to a very limited extent. We at the Max Planck Institute for Geanthropology are in the process of changing this and are developing an AI assistant to support staff at our institute in their future research. To this end, we have developed our own language model - GeaCop. The name stands for Geanthropology Cooperation Partner. During its training, GeaCop has analyzed thousands of scientific publications from scientific fields relevant to us and has become a real expert in the process. Let us explain to you how a language model works and how to train it. You may also want to ask GeaCop your questions or test his limits. GeaCop knows almost everything about the Anthropocene, but does he know a good recipe for strawberry pie?
Given the flood of new scientific findings, even proven experts in a field can't know everything, and the same goes for GeaCop. That's why we've given GeaCop a partner - Kantropos. When users submit queries to GeaCop, Kantropos first searches a huge collection of scientific literature for information that could be useful in answering the question. These snippets of information are then forwarded together with the question to Geacop, which processes them into a meaningful answer. This principle is known as RAG - information retrieval that supports answer generation (augmented generation). Thanks to Kantropos, Geacop can not only access scientific literature, but also return it to the user together with its answer. You can check Geacop's answers at any time, which solves the problem of hallucinations. However, you can also delve deeper into the literature with Geacop's support. You can find out exactly how this works from us or you can simply conduct a scientific dialog with Geacop and Kantropos, possibly on a topic on which you have published yourself.
GeaCop and Kantropos are far from being able to solve scientific problems on their own. It is only the dialog between our researchers and GeaCop and Kantropos that supports the emergence of new scientific insights - researching together with AI: We don't just think it's possible, we're already doing it.