WAIC 2026: Can AI truly become a scientist?

[00:00:00]Artificial intelligence is already helping [music] scientists analyze data, predict molecular structures, and search through decades [music] of published research.

[00:00:10]Now researchers are asking whether it can take part in the discovery process.

[00:00:16]Traditionally, scientists form a hypothesis, design an experiment, study the results, and then decide what to do next. In a closed-loop system, the AI essentially becomes the scientist, proposing an idea, directing or analyzing an experiment, learning from the outcome, and then beginning the cycle again. The promise is not simply faster research, but a process that can keep refining itself.

[00:00:39]But completing that loop requires very different capabilities. [music] Researchers at the World Artificial Intelligence Conference say current systems are already highly effective at reading scientific papers, [music] connecting information, and summarizing existing knowledge. Generating a genuinely original scientific idea remains much harder.

[00:00:59]>> [music] >> Any new hypothesis must then be tested first through calculations or simulations, and ultimately through experiments in the physical world.

[00:01:32][music] Scientists hope AI agents and automated laboratories could eventually design those experiments, operate around the clock, and feed the results directly back into the model.

[00:01:43]But that raises a more fundamental question.

[00:01:46]>> [music] >> Is producing a correct prediction the same as understanding why it's correct?

[00:01:53]>> The way we used [music] to do science with is to reproduce the way the biological molecule works.

[00:02:06]>> [snorts] >> AI is based more on correlation. So, these are two, I would say, complementary direction.

[00:02:15]One strive more to understand, the other strive more to predict.

[00:02:21]>> For Nobel laureate Arieh Warshel, the boundary might not be as clear as it once appeared. He says [music] that given enough experimental evidence, a sufficiently capable system might be able to infer an entirely new scientific theory, even if humans couldn't always follow its reasoning.

[00:02:39]>> Six years ago, I would be very negative about AI.

[00:02:44]Now, I'm a little I'm I'm using it, but this does not matter really.

[00:02:51]I I think it has more potential.

[00:02:54]So, the answer is that >> [snorts] >> clearly, the hype might not be justified because I mean, I grew up on the ideas that understanding >> [music] >> will tell me what to do. Now, I'm not sure about it anymore.

[00:03:17]>> [music] >> Maybe just by AI, you will know the answer.

[00:03:22]>> [music] >> And you will know it much faster than somebody who understand the problem.

[00:03:28]>> Building these systems will require enormous quantities of data, [music] specialist models, computing power, and advanced facilities.

[00:03:36]Those resources are currently concentrated in relatively small number of universities, [music] companies, and countries.

[00:04:01]>> [music] >> Wu says scientific AI will need to be offered through shared platforms, allowing researchers to access models, computing, [music] and specialist agents as services rather than having to build the entire infrastructure themselves.

[00:04:15][music] She's also calling for international funding and training programs to ensure developing countries can participate.

[00:04:22]The debate now is how far that partnership should go and what parts of scientific discovery should remain in human hands.