Science’s biggest AI challenge isn’t discovery – it’s proof

AI is making scientific ideas abundant. The hard part is knowing which ones to trust.

4 minute read
Ryan Cory-Wright

Assistant Professor of Analytics and Operations

Two researchers in a lab looking at a computer screen

Article at a glance

  • AI can now generate scientific hypotheses faster than traditional research systems can test them

  • Scientific progress still depends on verification: testing whether promising claims hold up against evidence, theory and domain expertise

  • AI-driven science will deliver on its promise only if we can test ideas as quickly and rigorously as we generate them

Scientific breakthroughs look sudden only in retrospect: a new theory, a medical advance, a technology that changes how we see the world. But before an idea becomes a discovery, it must survive a slower and less glamorous process of verification.

AI-driven science needs scalable verification methods. This is because AI systems can analyse vast amounts of data, identify patterns and generate new hypotheses faster than ever before. But verifying those hypotheses remains slow, costly, and highly domain-specific.

Ideas need proof

Science begins with a hypothesis: a claim precise enough to be tested, challenged and potentially proven wrong. Verification asks whether that hypothesis survives contact with what we already know by testing it against against theory, evidence, experiments, simulations, or formal reasoning. It is this process that separates genuine discoveries from ideas that merely appear plausible.

Scientific advances such as germ theory, thermodynamics and modern agricultural practices transformed society not simply because scientists proposed new ideas, but because those ideas were rigorously tested and repeatedly validated. Verification, then, is the dividing line between a promising idea and a discovery that science can trust.

The age of abundant ideas

For centuries, generating new scientific ideas was one of the most time-consuming parts of research. Over the past decades, major discoveries have become harder to achieve as scientific problems have grown more complex and interconnected.

This is what makes AI so exciting for science: it changes the speed at which researchers can move from data to possible explanations.

Machine learning systems can scan vast datasets for patterns that humans might miss, while large language models (LLMs) can help connect those patterns to possible explanations, making hypothesis generation dramatically faster.

That changes the economics of discovery. When machines can produce candidate explanations faster than researchers can test them, the bottleneck shifts from generating ideas to verifying and filtering them.

Science’s new bottleneck

In many areas of science, the pace of hypothesis generation has historically been constrained by human expertise, data collection and theory-building. AI weakens that constraint: it can propose many more candidates than existing verification systems, such as peer review, can comfortably absorb.

While the number of hypotheses has increased dramatically, verification remains comparatively slow, expensive and often dependent on human expertise. Laboratory experiments, clinical trials and peer review cannot easily be scaled at the same rate as AI-generated ideas. The result is a growing verification bottleneck.

“Generating ideas is no longer the primary constraint. Determining which ideas stand up to scrutiny is becoming the harder problem.”

Without effective verification, scientific progress risks becoming overwhelmed by plausible yet untested results. Some hypotheses may fit existing data while failing to hold up under further scrutiny. Others may appear convincing but remain disconnected from established scientific knowledge.

In other words, generating ideas is no longer the primary constraint. Determining which ideas stand up to scrutiny is becoming the harder problem.

History shows the cost of inadequate verification. Scientific claims that initially appeared convincing have later been overturned because measurements were flawed, assumptions proved incorrect, or results could not be replicated. Without rigorous verification, minor errors can compound, leading to years of wasted effort and resources.

Verification isn’t one-size-fits-all

Verification is not the same in every scientific field. In physics, it often relies on mathematical consistency and reproducible experiments. In biology, it may require laboratory testing and observation. In medicine, verification can involve clinical trials and statistical evidence.

There is no universal framework that can verify every scientific claim. What unites these disciplines is the principle that scientific knowledge advances only when ideas are tested against evidence. Although the methods differ, verification remains fundamentally driven by structured, iterative reasoning.

Trust is the new competitive advantage

AI is changing scientific discovery and may ultimately require us to rethink aspects of the traditional scientific method.

For much of human history, ideas were scarce and difficult to generate. Today, AI is making ideas abundant. The scarce resource is increasingly becoming the ability to determine which discoveries are robust enough to trust, invest in and build upon.

“In an age of abundant AI-generated ideas, trust becomes the scarce resource, and verification will be how science and business earn it.”

As AI continues to accelerate scientific discovery, the greatest opportunities may lie not in generating ever more hypotheses, but in developing better ways to verify them. That means building verification into AI-driven science itself through:

  • Automated checks wherever possible
  • Stronger links to established theory
  • Better benchmarks
  • Human oversight where scientific judgement matters

For businesses, universities and investors, the main lesson is not that AI makes scientists less important. It is that AI makes scientific judgement more valuable. As machines make it cheaper to generate hypotheses, the competitive advantage will shift to organisations that can test claims quickly, rigorously and carefully before building products, making investments or changing practice. In an age of abundant AI-generated ideas, trust becomes the scarce resource, and verification will be how science and business earn it.

Discover insights by Dr Cory-Wright

Meet the author

  • Ryan Cory-Wright

    About Ryan Cory-Wright

    Assistant Professor of Analytics and Operations
    Ryan Cory-Wright is an assistant professor in the Department of Analytics, Marketing & Operations and is affiliated with the I–X initiative on interdisciplinary AI and machine learning.

    His research interests lie at the intersection of optimisation, machine learning, and statistics, and their applications in business analytics and renewable energy.


    Read for more information and publications.