- AI is transforming drug discovery, materials science, and mathematics - but results are uneven across domains
- AI hallucinations and bias may be structurally unavoidable, not just engineering problems to fix
- The economic impact on human work is harder to predict than the technology itself
We are building new minds
Science Reader tracks AI in science - and the science of AI.
What happens when artificial intelligence meets scientific discovery? What are the genuine breakthroughs, the stubborn limits, and the human consequences of a technology that is moving faster than our ability to understand it? What do we do when it evolves to solve AI benchmark tests faster than we can develop them?
How is artificial intelligence created and improved in order to become a powerful tool for scientists and scientific research?
AI in Science: What's Real
Across biology, chemistry, and physics, AI in science is reshaping how researchers search for answers. AlphaGenome can now read the regulatory switches hidden in DNA, while AI materials discovery is shortening the gap between simulation and synthesis.
Research also shows that AI can add creativity and novelty to science, but only in fields where knowledge is fragmented. In other fields it can help automate research.
The mathematical front is more contested: real breakthroughs exist in AI-assisted mathematics, but they sit alongside considerable hype. One recent example was OpenAI's successful crack at the Discrete Geometry Conjecture which was disproved by an AI model.
Even the laboratories themselves are changing shape - though self-driving labs are not as autonomous as their name suggests. Drug discovery offers the same pattern: genuine promise, but still largely experimental in 2026.
Artificial intelligence is also becoming important in space exploration, and was a major feature in the Artemis II mission to the moon from start to end.
AI Reliability and Limits
The reliability problem is not a bug to be patched - it may be structural.
OpenAI researchers have shown that AI hallucinations are mathematically inevitable, a finding that compounds earlier work suggesting AI mistakes cannot be fully engineered away. The consequences are already visible in science publishing: AI-generated slop is flooding journals, and fabricated images are slipping through peer review.
Meanwhile, language models amplify the biases already present in their training data. There are proven methods to reduce hallucinations, but none that eliminate them.
AI and Human Work
The economic and human consequences of AI are harder to forecast than its technical capabilities. A rigorous study found that AI coding tools actually slowed experienced developers by 19% - a result that cuts against the dominant narrative.
The labor picture is murky in other directions too: the economics of AI-driven job loss contain a genuine paradox, and artists face displacement pressures that are real but unevenly distributed. At the edges of the field, the founders of deep learning are sounding alarms: Geoffrey Hinton says the world is not prepared, and Yoshua Bengio argues that AI agency - not raw capability - is the real danger.
In education, what students lose when AI does their writing is not just skill - it is a mode of thought.
