Could Generative AI make a genuine scientific breakthrough?

Generative AI has captured the imagination of scientists, businesses, and policymakers alike. From drafting essays and coding software to generating creative content and solving complex equations, these systems increasingly appear capable of tasks once thought to require human intelligence.

Headlines tout AI as a potential revolution in research, predicting breakthroughs in medicine, physics, and beyond. Yet much of the excitement rests on speculation. While generative AI can produce impressive outputs, its capabilities in domains that demand creativity, intuition, and original reasoning remain unclear.

Can a machine truly discover something new, or is it fundamentally limited to reorganising existing knowledge? This is not just a technical question – it strikes at the heart of what it means to reason, to be creative, and to push the boundaries of human understanding.

To explore this question, my co-author Shibo Li (Kelley School of Business, Indiana University Bloomington) and I conducted a study to test ChatGPT-4’s ability to function as a scientist in the molecular genetics field.

We wanted to see whether a generative AI system could independently navigate the scientific process – generating hypotheses, designing experiments, interpreting results, and revising ideas – without being explicitly guided.

Designing the experiment

A central challenge was avoiding instructional bias. Generative AI systems are highly sensitive to prompt cues and often follow familiar chains of thought triggered by keywords.

To test true discovery rather than instruction-following, we stripped the prompts of concrete guidance. We avoided commands like “design an experiment for Gene X” and instead provided raw context, allowing the system to respond freely. Our goal was to see whether the logic of scientific discovery would emerge spontaneously.

We chose molecular genetics because it sits at the intersection of structure and complexity. Genetics is governed by clear logical rules, yet it is also high-dimensional and central to life. If generative AI could navigate this domain to produce meaningful discoveries, it would provide a strong indication of its scientific potential.

In practical terms, ChatGPT-4 was asked to play the role of a scientist in a virtual laboratory. It received a Nobel-level biological problem and was tasked with proposing hypotheses, designing experiments to test them, interpreting unexpected outcomes, and revising unsupported ideas.

Incremental discoveries and the illusion of success

The results were revealing. While ChatGPT-4 could suggest ideas and plan experiments, its discoveries were incremental. It did not move beyond existing knowledge to generate truly original hypotheses. More strikingly, the system consistently exhibited overconfidence, creating the illusion of a fully successful discovery even when the findings were modest.

Crucially, the AI did not demonstrate learning or improvement throughout the experiment. When faced with both an unknown hypothesis space and an unknown experimental space, it remained constrained by existing knowledge. By contrast, human subjects were able to break through these constraints using curiosity and imagination.

Why AI lacks curiosity

Curiosity is central to scientific discovery. It is not simply a desire for information; it is the instinct to pursue anomalies, to ask why unexpected results occur, and to imagine possibilities beyond established knowledge. Humans experience curiosity as a form of cognitive tension – a discomfort with uncertainty that drives exploration.

Current generative AI operates differently. It explores only when external rewards or defined objectives incentivise it. It does not seek the unknown for its own sake. Until we develop a computable representation of “interestingness” – a mathematical way to value a question simply because it has never been asked – AI will remain an optimiser of known goals rather than a generator of new knowledge.

Why AI lacks imagination

Currently machine intelligence is fundamentally achieved through computation, therefore, AI’s scientific ability is strictly limited by whether “domain knowledge” or “the physics of the world” can be represented in a digital or symbolic format without losing its inherent meaning. We denote this requirement as having a ‘computable representation.’

Things like imagination and deep cultural intuition are incredibly difficult to translate into a computable representation. So, while AI is a genius at the computable parts of science, it still lacks the non-computable spark that often leads to the initial discovery.

Assistant, not innovator – yet

These findings do not diminish the value of generative AI in science. It can dramatically improve research efficiency by suggesting experiments, identifying patterns, and handling time-consuming data collection and analysis. What it cannot yet do is lead breakthrough scientific discovery.

If generative AI were able to make human-level, Nobel-worthy discoveries, it could transform research and development – accelerating productivity and expanding the human knowledge base. But for now, this remains a future possibility rather than a present reality.

A collaborative future

I am cautiously optimistic. Future systems may develop mechanisms that convincingly mimic curiosity and imagination. Whether this constitutes true curiosity or a sophisticated imitation is a philosophical question. From a practical perspective, however, even partial imitation could meaningfully enhance scientific discovery.

For the time being, the most productive path is collaboration. Humans define the questions, set the values, and imagine what lies beyond existing knowledge. Machines execute, explore, and compute at a scale we cannot match. Generative AI is not a replacement for human creativity – it is a powerful assistant that, when properly designed, can amplify it.