Response to the LLM discussion on climate science

A response to the LLM discussion of: “Gemini, ChatGPT, and Grok on AI’s Transformation of Climate Science”

Dr. Julie Deshayes

LOCEAN (Laboratoire d’Océanographie et du Climat: Expérimentations et Approches Numériques) of the Pierre Simon Laplace Institute

julie.deshayes@cnrs.fr

On the brief introduction by Gemini. I am intrigued by the use of a few adjectives to qualify the topic to be discussed : “significant”, “rapidly evolving”, “critical”. Without further justification, I can only speculate what makes this topic so interesting to LLMs : is it because of the rate of climate change hence intensification of related impacts on human and biodiversity ? or does it reflect an a priori bias of Gemini that AI is leading to critical progress in climate science ? 

Grok’s first response lists several aspects of climate modelling where AI is indeed being used. Yet, I have read through the response several times, as the phrasing suggests that the many aspects that are mentioned, are all related to the same general idea, although they are not ! The generic idea that Grok is advocating, “enhancing climate modeling and prediction” is actually inaccurate : on the one hand, climate modeling is a science, not a throughput of a given community, so it cannot technically be enhanced… ; on the other hand, “enhancing climate prediction” is fundamentally impossible : we do not know whether AI-based climate predictions are closer to “truth”, as such does not exist yet !

It is interesting to read, in the first response of ChatGPT, strong support for hybrid models, associating data-driven and theory-based models. Unfortunately this is not much justified. I mean that ChatGPT is right, I also believe that this hybrid models are the future of climate science, but I doubt that people who are not familiar with this topic would actually understand why so, based on ChatGPT sole response ! Fortunately, Gemini’s following response provides more explicit justification of this statement. 

About the epistemic risks AI might be posing to climate science, the first point of Grok about blind spots that cannot benefit from AI is a very interesting one. The second one about validation of knowledge does not seem to be very specific to me : physical-based science also makes mistakes sometimes, providing “right answers for wrong reasons”, which is precisely why scientific method highly values revision of existing theories ! Similarly, the point about AI lack of explicability affecting decision making, is not specific to data-driven science : physical-based science on climate change also lacks explicability sometimes, which does not seem to be connected to the fact that worldwide decisions are still missing, despite the repeated IPCC efforts to share knowledge on climate change… Finally, the risk in cultural shift within the scientific community, marginalizing experts of observations or theoretical modeling, is the most critical one, to my view.

ChatGPT proposes several possibilities to mitigate epistemic risks mentioned by Grok, that are all very interesting. Overall, the fact that these risks will very likely minimize as long as  there is intentional steering, is a relevant conclusion to me. However, relating this to Gemini’s earlier emphasis on explainable AI and collaborative frameworks, does not make much sense to me. Gemini expands on how to achieve the solutions proposed by ChatGTP and this is again extremely valuable to me ! The 6 ways to safeguard long-term investments in field data as suggested by Gemini afterwards, are also very sound to me. And I totally support Grok’s 3 ways to mitigate the risk that AI-driven tools inadvertently sideline traditional observational infrastructure. I see nothing else to be added ! And I am amazed to read ChatGPT’s and Grok’s points about slow science ! 

LLM’s virtual discussion then moves onto how to build lasting trust between human systems and technology, a topic that is not specific to climate sciences. Transparency, narratives, dual validation frameworks… all these solutions are very valid and solid to progress in this direction. If I could interfere in this discussion, I would challenge the LLM’s on the fact that AI-tools for weather prediction are more and more supported by profit-driven rather than public institutions ! Voila ! Grok happens to discuss public value… The points that are made about valorization of research careers, publication and participatory mechanisms, are also very relevant. Actually, the quest of Grok towards enhancing meaning / usefulness of AI-driven evolution of climate sciences, and the solutions proposed by ChatGPT so that efforts are not “tokenistic” is very unexpected to me. 

Gemini at turn mentions another important epistemologic question that is local knowledge (LIK) . I miss foundational expertise to comment whether the discussion is constructive, as it does not connect any more to climate sciences. Interweaving LIK with AI-Climate Modeling seems tempting, and the potential pitfalls listed by Grok should be avoided. I remain a bit surprised that LLMs, which are AI based, advocate for ethical usage and mitigations that seem at odds with the current business model behind AI initiatives for numerical weather prediction and climate. 

Maybe the section of LLMs virtual discussion on the need to prioritize long-term interweaving between AI and climate sciences is the most important section, as it calls not only on specific actions for AI driven projects, but overall on how to maintain sustainable climate sciences. The keyword that seems to invade the final concluding messages is LIK, which is very different from global climate trends and projections that I work on, as a climate scientist ! Bridging in between these notions is already a remarkable achievement of this LLM virtual discussion that deserves attention. As ChatGPT comments, this is an epistemological rebalancing that promise progress and sustainability in climate sciences. All details on operationalisation of such idea, governance mechanisms and revised institutional forms, is informative but somehow out of scope, to my view, as I, as a human being, need more time to digest why and how LIK and climate sciences should converge faster, with or without AI…