AI may be closer to doing original scientific research than we think.
Science has always been a human endeavor, fueled by curiosity, creativity, and a stubborn willingness to question what others take for granted. But what happens when artificial intelligence begins to do the same—not just assisting human scientists, but independently designing experiments, analyzing data, and forming conclusions?
That question became more than theoretical recently, when an AI system from Japan’s Sakana AI generated a hypothesis, designed experiments and wrote a peer-rerviewed scientific paper on its conclusions, all without human intervention.
Titled Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization, the paper was accepted as a Spotlight Paper at ICLR 2025, one of the field’s most prestigious machine learning gatherings. In a quiet way, this event marked a threshold: AI had authored original research deemed worthy by its human peers.
The Rise of the AI Scientist
The system, called AI Scientist-v2, is not just another language model. It is a fully autonomous research agent designed to automate the entire scientific process. Reviewers, unaware that the paper was AI-authored, scored it high enough for acceptance, placing it above nearly half of all submissions by humans.
The implications are profound: a machine not only understood a research domain, but formulated questions, conducted experiments, wrote code, analyzed data, and expressed its findings clearly.
The Promise—and the Problem
At first glance, the achievement suggests we may be inching toward an “intelligence explosion”— the point where AI not only helps with science, but drives it, adding to human knowledge faster than humans can themselves. Some, like former OpenAI researcher Leopold Aschenbrenner, believe this tipping point could come as early as 2027.
But not everyone agrees.
Yann LeCun, Meta’s Chief AI Scientist and a Turing Award winner, has long warned against mistaking pattern-matching for true intelligence. Current AI models cannot form the kind of mental models that underpin real-world reasoning or original discovery, he says.
In other words, the AI Scientist-v2 may have “written” a research paper, but whether it understood what it was doing — or just stitched together patterns from its training — is still an open question.
Sakana’s Cautious Breakthrough
To their credit, Sakana.AI has treated this experiment as exactly that: an experiment. The company withdrew the paper before the conference, acknowledging the ethical gray zone it occupies.
Nonetheless, as AI systems become more capable, they will increasingly play roles in scientific discovery. They are already amplifying the process, accelerating literature reviews, speeding up code development, and generating experiment designs in minutes instead of months.
That, LeCun agrees, is the near-term future: AI as a powerful tool, not an autonomous genius.
Beyond Imitation: Toward Original Thought
So is Compositional Regularization truly original research? In a narrow sense — yes. The paper introduced a novel experimental setup, investigated a fresh angle on generalization, and was judged as worthy of presentation. Still, its findings were incremental, showing that its hypothesis failed. And in the broader philosophical sense, originality in science is not just about novelty; it’s about intuition, question-asking, and an ability to see beyond the data.
LeCun likens this to the difference between solving a math problem and inventing a new branch of mathematics. The latter requires a system to understand the world, make predictions based on experience, and plan actions based on abstract goals. Those abilities, he argues, are still out of reach for AI.
Still, the fact that a machine can imitate the form of scientific thought this well is not trivial. It raises the bar for what’s possible. In the coming years, AI will likely generate hypotheses, automate lab work, and perhaps one day co-author Nobel-worthy research. But LeCun’s caution reminds us: authorship does not imply understanding, and prediction is not the same as comprehension.
The Road Ahead: Collaborative Intelligence
The road forward may lie in hybrid intelligence—where AI systems handle complexity and scale, while humans provide insight, ethics, and conceptual leaps. Sakana.AI’s milestone is not the destination, but a waypoint on a longer journey toward reshaping how we do science.
The Sakana AI experiments reignites the long-simmering discussion of whether AI will soon be optimizing its own architecture, refining its own reasoning capabilities, and accelerate the rate of discovery in ways we cannot yet predict. The AI Scientist’s success does not mean we are at that inflection point—but it does suggest we may be closer than many people think.
