Recent advancements in artificial intelligence (AI) have sparked intense discussions surrounding large language models (LLMs) like ChatGPT and others capable of generating text in response to input. While these tools offer potential benefits for research, there are significant apprehensions regarding their impact—ranging from job displacement and excessive reliance on AI to the dissemination of AI-generated disinformation that undermines democratic systems.
Less explored is the potential constructive use of such technologies to develop tools that can sift through and summarize scientific evidence for policy formulation. Worldwide, science advisers play a crucial role as knowledge intermediaries, providing timely, science-and-technology-related information to presidents, prime ministers, civil servants, and politicians.
Science advisers are tasked with navigating a vast array of information, from solid-state batteries and antibiotic resistance to deep-sea mining. They need to sift through millions of scientific papers published annually, alongside considering reports from advocacy organizations, industry, and scientific academies—each with its own perspective. The pressure to produce policy summaries within tight deadlines is considerable, and the demand for such information by governments is on the rise.
AI-based tools could enhance the capabilities of science advisers and assist policymakers in staying informed. However, it is vital to carefully design these tools to avoid potential pitfalls. Could AI compromise rigor? Might the outcomes be influenced by specific agendas? If AI tools are used by science advisers, what safeguards could prevent AI errors and distorted interpretations from affecting public policy decisions?
Answers to these critical questions are urgently needed. Powerful language models are already extensively used in research and technological development, with increasingly advanced capabilities becoming accessible to a broader user base through commercialization and open sourcing. Policymakers have started experimenting with publicly available generative AI tools. For instance, legislative staff members in the United States are exploring OpenAI's GPT-4 and reportedly other unapproved and potentially less reliable AI tools. This led the US House of Representatives administrators to impose limitations on chatbot use in June.
Our perspective is that, with cautious development and effective management, a new generation of AI-based tools could revolutionize science advice in the near future, making it more adaptable, rigorous, and targeted. However, leveraging these tools for the greater good necessitates the formulation of guidelines by science advisers and policy institutions, and a thoughtful consideration of the design and responsible use of this emerging technology.
In this discussion, we delve into two tasks where generative AI tools hold promise for policy guidance— synthesizing evidence and composing briefing papers. We also emphasize areas that require closer scrutiny.
