Researchers losing the ability to develop scientific reasoning a big fear with AI: Climate Scientist William Boos

A fundamental fear with artificial intelligence (AI) and its growing use in science is that researchers, particularly students, could become so dependent on AI that they lose the ability to develop scientific reasoning and thinking skills, believes William Boos, climate scientist and professor at the University of California, Berkeley.

“I’m skeptical we will be able to replace human ingenuity totally, but if we get to a level where people are being trained, and students rely on it so much that they lose the ability to develop scientific reasoning and thinking skills, we're going to be in a hard place as a society,” Boos told The Economic Times Digital in an exclusive chat during a conference organised by Ashoka University titled Future of Climate: AI, Science and Society.

The larger question, he said, is what happens if humans stop learning how to do science.


This concern comes despite him seeing AI as a major opportunity for climate and weather forecasting. He describes AI as the “second revolution” in his field, with newer models capable of predicting weather as skillfully as state-of-the-art physics-based models while requiring far less computing power.

Also Read: India can become an AI infrastructure hub for Global South: Microsoft's Puneet Chandok

“The first revolution came in the 1950s and ‘60s, when we realised physical laws could predict the evolution of the atmosphere. But that required supercomputers. Now, instead of solving equations at every latitude, longitude, and altitude, an AI model is like someone who’s watched the weather for decades and just has the intuition. That’s what's happening — a second revolution.”

AI could also make sophisticated weather forecasting more accessible to countries and research labs that cannot afford the supercomputers traditionally needed for it.

“If Google, or India, or the US, or the European weather centers build a model and make it open access, then any country, any lab could use it, and it would be very cheap computationally to predict the future weather,” Boos said.

He said AI weather prediction requires 10,000 times less computing power than traditional weather prediction, although training the model initially requires significant computing power. Once trained, using the model to make predictions is much more efficient.

The reason, as per Boos, is that the initial phase of training the model is what requires the computing power, but once that is done, the usage of that model from then on to predict things, that is much more efficient.

Risks of privatising AI weather forecasts

The possibility depends on access to these models remaining open.

Boos is concerned about complete privatisation of AI-based weather forecasting at the hands of Big Tech companies, which could eventually lead to a closed-access ecosystem.

“If the national centres don’t keep staying in the game, if it does become privatised… and if five years from now the Big Tech companies are saying, ‘We’re not going to keep these open access, we’re going to close them off’, then the whole potential for AI weather prediction to really democratise access to this information could be lost,” he said.

He pointed to developing countries that do not have their own forecasting departments or agencies, including several countries in Africa.

“Think about the many countries in Africa that aren’t running their own weather models. They could run their own… but if everything becomes closed off and locked down, if the governments don’t stay in the AI weather prediction business, that could disrupt the dream of very accessible, state-of-the-art weather predictions.”

AI could change learning science

Boos does not see AI's impact on science as entirely negative. But he believes its adoption could change how researchers learn and practise science, much as computers and numerical modelling changed the field before it.

“I’m concerned that science will be very different and maybe not in ways that are all for the better, and that we'll lose a lot. You look back… with the advent of computers and numerical modeling, scientists, many scientists in my field used to be excellent pen-and-paper mathematicians. We have lost a lot of that… a lot of that has become a lost art,” he said.

“Most people are a lot less mathematical than they would have been in the 1970s and ‘80s, because we rely on computers to do that work numerically.”

Also Read: DeepSeek to brief UN Security Council on AI this week, sources say

At the same time, AI can free scientists from some of the more repetitive work involved in research.

“A lot of science is data analysis and building numerical models — a lot of coding. AI has changed that, and not for the worse: if you can spend your time thinking about the science instead of a week bogged down by bug-finding in code, that's a good thing,” he said.

Why physics-based models still matter

Tapio Schneider, Theodore Y. Wu Professor of Environmental Science and Engineering at Caltech, said that despite the growth of AI weather forecasting, physics-based models remain critical to climate research.

“How do we know Earth is warming and will warm further? It's based on physical models: radiative transfer. We can describe exactly how more carbon dioxide modifies radiative fluxes in the atmosphere; it's the laws of electrodynamics, and that leads to warming, and warming has feedback,” Schneider told The Economic Times Digital during the conference.

“You can audit the pieces of the chain by which you arrive at your conclusion,” he said. That, he added, is a limitation with AI models if they operate as a black box.

“If you just have a black-box oracle, you can’t interrogate the details. If your model produces an unusual prediction, you want to trace it back to a certain process,” Schneider said.

He gave a hypothetical example of an AI model predicting that the Antarctic ice sheet will collapse.

“I can test the mechanics of it — go to Antarctica and probe the ice sheet, or build a laboratory system to test it. But with an AI system, where suddenly the ice sheet disappears, and you have no idea why, how do you produce a new hypothesis to scrutinise?”

So, scrutiny and interrogation are two pillars still critical to scientific research, and physics-based models allow you to do so. “Unlike climate, weather you check every day – if it was right or wrong? If it’s right many days in a row, I'm happy to trust that it'll be right tomorrow too. We don’t have that for climate.”