AI is having a remarkable run in science. It’s helping researchers predict protein structures, improve weather forecasts, discover new materials, and speed up drug discovery. 
Researchers are getting results far faster than before. Work that once took weeks or months can now be finished in days, and sometimes hours. In many labs, AI is now used as commonly as simulation software and high-performance computing.
That’s been the easy part. The harder question is whether AI is actually discovering anything on its own. Does it even have the ingenuity or creativity that some of the great scientists and inventors have? Or is it somewhat limited to just recognizing patterns that humans would eventually have found anyway.
A new paper published in arXiv argues that there is an important difference between the two. The authors believe today’s AI systems are becoming exceptionally good scientific assistants, but they’re still missing one of the key ingredients behind the kinds of breakthroughs that changed science forever.
To explain why, they turn to Albert Einstein.
Einstein believed scientific discovery was about more than collecting observations and applying logic. According to the paper, he saw it as a cycle. Scientists start with observations from the real world. At some point comes an intuitive leap – a new idea or way of looking at the problem. Only then does logic step in, allowing those ideas to be tested through experiments.
(Shutterstock/Pressmaster)
The authors of the paper argue that this intuitive leap is exactly where today’s AI still falls short.
We know that AI has become exceptionally good at induction, which refers to finding patterns in vast amounts of data. And it’s also getting much better at deduction, which means using an established set of rules to solve complex tasks.
However, there’s a third form of reasoning – and it often doesn’t get the attention it deserves. In fact, many people may not have even heard the term. Yet the authors argue it sits at the heart of scientific invention. This third form is called abduction.
Abduction is about coming up with a new explanation when the old one no longer makes sense. It’s the kind of thinking behind breakthroughs like relativity and quantum mechanics – ideas that completely changed how scientists understand the universe.
Those theories happened because someone looked at familiar evidence and saw it differently. That “looking at things differently” is where the paper says today’s AI is still missing. 
We know that LLMs are excellent in finding relationships in existing data. That’s enormously useful. But inventing an entirely new scientific framework is a very different kind of problem.
History suggests that the biggest scientific breakthroughs often happen when the evidence is messy or incomplete. In some cases, the evidence is even contradictory. What would AI do in that situation when left to make decisions on its own? When existing theories stop making sense, making better predictions isn’t always enough. Sometimes science needs a completely new way of thinking. Something it hasn’t been trained on. Something it doesn’t have past data to rely on. 
None of this suggests AI has hit a wall in science. Quite the opposite.
The paper itself points to systems like AlphaFold, which has transformed structural biology, along with AI’s growing role in chemistry, materials science, astronomy and climate research. Researchers are making faster progress because AI can sift through huge amounts of information, test more possibilities and rule out dead ends much earlier.
(Shutterstock/fran_kie)
The authors aren’t arguing that AI can’t make discoveries. Their point is more nuanced. Right now, AI seems to be at its best when it’s helping scientists make discoveries, rather than coming up with entirely new theories on its own. It’s a great assistant, but it’s not a great scientist. 
Whether that changes is still an open question.
This is a position paper. It’s not an experimental study. It doesn’t claim AI will never reach that next stage. The paper simply asks: Does the current path of AI research is actually leading us there? 
Bigger models, larger datasets and more computing power have driven remarkable progress over the past few years. But if scientific invention depends on something more than increasingly sophisticated pattern recognition. Simply scaling today’s systems may not get us all the way there.
It’s a thought-provoking argument, especially as AI for science continues to accelerate.
By now, most of us have heard about the “wonder drug” Ozempic. Many experts believe…
Scientists have relied on HPC facilities to help them process large volumes of data, execute…
Modern AI can identify forests all the way from space. It can estimate tree heights…
AI is becoming an increasingly important tool for drug discovery. However, for many pharmaceutical and…
Last week’s TPC26 event brought together many of the world’s preeminent experts in using AI…
The AI industry has spent the past few years trying to give models access to…
Elon Musk’s SpaceX has filed for what could become the largest IPO in history. It…
AWS expanded its enterprise AI portfolio this week with a collection of new tools aimed…
For years, hyperscalers built infrastructure largely for their own cloud ecosystems. However, AI is starting…
The AI boom is creating winners beyond chip manufacturers and cloud providers. One of the…
“Almost right” is not good enough once AI starts making decisions inside a business. That…
Artificial intelligence has redrawn the data center design landscape, and in doing so has propelled…
AI used to be a race to build the best model. Now it’s a race…
Enterprise AI budget forecasting is mostly guesswork right now. Most organizations set their initial allocations…
The AI boom has kicked off a massive wave of data center construction around the…
BigDATAwire recently spoke with Cindi Howson, Chief Data and AI Strategy Officer (CAIO) at ThoughtSpot,…
AI coding assistants have become part of everyday software development. They write functions, generate tests,…
By now, most of us have heard about the “wonder drug” Ozempic. Many experts believe…
Alex Woodie Editorial Director +
HPCwire Managing Editor
Ali Azhar BigDATAwire Managing Editor
Jaime Hampton AIwire Managing Editor
Drew Jolly QCwire Managing Editor
Position: High Performance Computing Engineer
Location: Switzerland
Started in 1989, KDD is the oldest and largest data mining conference worldwide, covering technologies like deep learning, differential privacy, and ethical machine learning.

© 2025 BigDATAwire. A TCI Media Publication (Formerly Tabor Communications). All Rights Reserved.
Use of this site is governed by our Terms of Use and Privacy Policy.