How is AI changing science?

Why did you decide to look into the impact AI is having on the way scientists work?

We’ve all heard about the negative impacts AI is having on the climate, mental health, democratic processes, and the labour market. One of the most common justifications for these and future harms made by big tech CEOs is the positive potential of AI. We can’t really complain about energy intensive data centres and recommender algorithms if at the same time AI cures all human disease, solves world hunger and averts the climate crisis.

This puts us in a complicated situation. The most direct benefit of AI is profit, which goes to big tech companies. The second most direct benefit is not to the public, but to the scientists who use it to solve important problems. The public benefits from this in an indirect way. Hopefully. The upshot is that an accurate appraisal of AI in terms of social impact cannot be done without an evaluation of the impact of AI on science.

What are the potential safety risks of AI adoption in science that you want to explore?

Many scientists are genuinely excited about AI. Others worry about de-skilling, data privacy, biases and new forms of malpractice. There is a lot of anecdotal evidence to be found in social media posts, op-eds in science journals and blogs. And there is some quantitative evidence concerning scientist opinions toward AI. But there is very little qualitative data on how decisions about whether to adopt AI are actually made and how the landscape of scientific abilities is changing. That’s the gap the project aims to fill. Perhaps the main worry is de-skilling, or what is now called “never-skilling.” Early adopters of AI tend to be people at the beginning and end of their careers. Well-established older scientists aren’t in danger of de-skilling. Experimenting with AI to see where it might play a role in their work is low-risk because they can easily spot mistakes and unhelpful suggestions. Younger scientists have grown up with AI and they accept it as my generation did the internet. They never felt the need to gain the skills which middle-career scientists just finished gaining the hard way. 

Is this worrying? I think so. An autopilot handles 90-95% of flight time, but we passengers are assured in the knowledge that airplane pilots must learn and maintain the skills required to safely fly a plane from take-off to landing. If something happens, they can take over. Surgeons and healthcare professionals, on the other hand, may soon be assisted to a large extent by digital copilots and these people, who are also responsible for our well-being, aren’t required to learn or maintain the skills that AI is replacing in order to renew their licenses.

A second worry about AI adoption in science concerns which projects scientists decide to pursue. It is becoming more common for funding bodies to prefer projects with AI somewhere in them. Industry scientists face mounting pressure to include AI in their solutions. What will this do to science? In the short term at least, it is likely to lead some scientists to exclusively pursue AI-solvable problems. They want to use the new hammer. There is a good chance that some of these projects are not the ones which would best profit the public. More data is needed.

Are there some disciplines where the safety concerns are more acute?

Everyone would agree that science can positively affect our lives. But it is notoriously difficult, for any given positive effect, to track just where it came from. While we might credit a medical company for inventing a new life-saving device, the inventors themselves will often give credit in turn to earlier breakthroughs in medicine, chemistry, engineering, biology, or physics. This is often how governments justify spending on “pure” science. So, we must never count out the possibility that AI applications will have potential ethical consequences, even in the “purest” areas of science.

Still, there are areas of science whose impact on our well-being is very clear and direct, and this project assumes that concentrating on those areas as a first priority makes sense. So we will be looking at a number of cases which might include AI applications in a) clean energy, especially fusion power (as a way of addressing the climate crisis), b) automated chemistry or materials science (as these might lead to more efficient solar power cells, agriculture breakthroughs, new medical devices, etc.), c) medicine and pharmacology (as these may lead to new drug discoveries, especially for neglected diseases), and d) space science (as these extend humanity into space, drive new inventions relevant to daily life, help track earth’s climate and agriculture from above, etc.). The concern with such cases is not safety in the sense of mitigating a direct risk to humans, but in the sense of best directing our scientific focus responsibly into the future for the benefit of all.

You're aiming to create the first detailed picture of how AI is currently used in research laboratories. How could this research make AI adoption in science safer? 

The project will produce new empirical data about how AI is changing science, especially concerning research trends and how scientific skills and abilities are being augmented, eliminated, or extended with AI. The focus on ability will be helpful in addressing concerns about de-skilling and never-skilling. Using this data, the project will enable new ways to evaluate scientific practice and pedagogy and help explain what “good science” looks like in an age shaped by tools that we don’t yet fully understand. For each case study, we will interview participants from upper-level management all the way through to lab technicians, learning about people’s backgrounds, favourite methods, daily activities and recent successes and failures. We will triangulate this interview data against what we observe in workshops, seminars, lab meetings and work along protocols.

From all of this, we will generate hypotheses about how things are changing and how those changes differ across contexts. Insofar as those hypotheses appear generalisable, we can then test them using large sample size, survey-based, quantitative methods. Then, we can help provide concrete recommendations, not only for the labs we study, but for all scientists considering adopting AI and looking for best practices. We also hope to offer clear structure and direction for science educators who want to teach skills that the next generation of scientists will require. Making science better requires making better scientists. And better scientists will make a safer, better world.

The project, Scientific Progress and Artificial Intelligence: a Capabilities-Based Ethnographic Epistemology, is supported by an ERC Consolidator grant.