AI Is Making Science Faster. Validation Must Now Catch Up

For decades, the scarce asset in research was insight. Scientists competed to see a pattern first, formulate the right hypothesis, or identify the experiment worth running. AI changes that equation. An agent can read vast literatures, assemble code, suggest mechanisms, simulate alternatives and generate candidate experiments at a pace that no individual lab can match.

That does not mean science has been automated. It means its bottleneck is moving.

Google DeepMind argues in “Conjecture Machines” that agents are beginning to make ideation and candidate generation much cheaper, while validation remains costly, slow and deeply dependent on domain expertise. Its central concern is institutional: scientific infrastructure was built to manage a world in which ideas arrived slowly. It is poorly prepared for a world in which thousands of plausible ideas arrive overnight.

The change matters because modern research already has a productivity problem. As fields become more complex, researchers spend more time absorbing prior work, coordinating teams, preparing data and navigating experimental systems. AI can relieve some of that burden. But if it multiplies the volume of hypotheses faster than labs can test them, it can also create a new kind of congestion: a scientific queue filled with attractive claims that nobody has the capacity to settle.

The discovery engine is already taking shape

“We are moving toward a future of serious 'proof indigestion' where AI generates breakthroughs faster than humans can review them,”

The basic agentic workflow is becoming clear. A research agent begins by retrieving papers, databases, lab notes and structured data. It breaks an objective into subtasks, uses code or specialist models, proposes candidate explanations, ranks them, and produces artifacts for a human researcher to inspect. Unlike an ordinary chatbot, the value is not simply that it answers a question. It can repeatedly act, observe outputs and revise its approach.

DeepMind describes agents as systems able to plan, use tools, run parallel processes and correct some errors during a task. It points to scientific settings where agents can connect large bodies of literature with specialised systems for protein, materials and weather research. The company’s argument is that this combination can expand the range of problems a small team can seriously investigate.

OpenAI’s January 2026 report on AI as a scientific collaborator makes a related case from the workflow level. It describes researchers using AI for literature synthesis, technical coding, mathematical derivation, data analysis, simulation support and experiment planning. OpenAI reported that weekly advanced science and mathematics messages on ChatGPT rose from 5.7 million to nearly 8.4 million during 2025.

This is evidence of adoption, not evidence that every interaction produced science. Still, it identifies the direction of travel. Researchers are already treating frontier systems less as answer engines and more as technical collaborators that reduce the cost of trying an idea.

That cost reduction is most powerful when an agent can operate inside a closed feedback loop. In mathematics, it can propose a proof and send the formal expression to a proof assistant. In computational physics, it can write a simulation and compare outputs against known constraints. In biology, it can propose protein variants, filter them with predictive models, prepare a shortlist and learn from laboratory results. The system does not need to be correct on its first attempt to be useful. It needs to make the next test cheaper and more informative.

Not all validation has the same price

The crucial distinction is between domains where correctness is machine-checkable and domains where it is discovered only through interaction with the physical world.

Mathematics is the cleanest case. A natural-language proof can be elegant, convincing and wrong. But when it is translated into a formal system such as Lean, every required step must be explicit. A proof assistant can reject a hidden gap without caring how persuasive the prose sounds. OpenAI highlights this as a major shift because formal verification converts some of the hardest parts of mathematical checking into a repeatable computational process.

Agent writes plausible argumentInformal proof

plausible prose

Human reads for gapsPersuasive prose is not a guarantee

unresolved uncertainty

?Hidden gap may remain

Agent proposes proofFormal proof

proposed proof

Translator writes Lean statementThe claim and proof become explicit

formal statement and proof steps

Proof assistant checks every stepNo hidden step is accepted

successful check

Verified under stated assumptionsMachine-checkable correctness

The same logic applies, imperfectly, to software and simulation. A program can be tested against held-out cases, checked for reproducibility, inspected for dependency changes and run independently. These tests do not prove that the underlying scientific hypothesis is right. They do establish whether the computational result is robust under specified conditions.

Biology and materials science are harder. A predicted molecule may be chemically plausible yet fail to synthesise. A protein sequence may perform well in a narrow assay but fail in a living system. A model may find a statistical association that disappears when a lab changes sample preparation, cell line, environment or measurement technique.

OpenAI’s report describes a lab-in-the-loop protein engineering effort in which a model generated thousands of candidate sequences, which were narrowed to a few hundred variants for experimental testing. That example illustrates both AI’s advantage and its limit. The model can search a massive design space. The laboratory still decides which candidates were built, which assay mattered, what counts as a hit and whether early results will reproduce.

The practical implication is that validation cannot be one generic stage at the end of an agent workflow. It has to be designed around the error model of each field.

The strongest criticism: throughput is not closure

The optimistic account risks confusing faster activity with faster scientific settlement. In a critical response, Leistiko argued that DeepMind’s essay relies on curated successes without establishing baseline validation throughput, replication failure rates or the true cost of resolving claims.

That is the right challenge. A compelling story about an agent recovering a hypothesis that took a human team years to establish does not reveal how often agents generate false leads that appear equally compelling. Nor does it tell a funder whether an AI-assisted programme produces more independently replicated results per dollar than a conventional programme.

This is especially important because more proposals can make a weak evaluation system look productive. A lab might run more simulations, file more patents, draft more papers or report more benchmark gains while learning little that survives outside its original setting. If each result is tuned to the same data, equipment, evaluation rubric or researcher judgment, apparent progress can compound hidden error.

The supportive view is more ambitious. Conor Griffin, a DeepMind co-author, wrote that the coming abundance of candidate solutions requires stronger validation infrastructure and changes to peer review. That argument is persuasive if the goal is not to remove humans from science, but to redirect scarce human attention toward the claims most worth testing.

The disagreement is therefore not about whether AI can generate ideas. It is about the metric that matters. The important output is not conjectures per hour. It is validated knowledge per unit of experimental, computational and expert-review capacity.

Build two systems, not one

A serious agentic research programme should separate its discovery system from its verification system. The discovery side should be encouraged to explore, combine distant literatures and propose unconventional paths. The verification side should be structurally independent, sceptical and rewarded for rejecting weak claims.

That design starts with provenance. Every output should preserve the model version, prompts, retrieved sources, data versions, tool calls, code, random seeds, assumptions and decision logs. A result that cannot be reconstructed is not ready for scientific reliance, however impressive its narrative may be.

Next comes independent evaluation. The agent that designed an experiment should not grade its own result. Separate models, separate teams or both should attempt to identify missing controls, data leakage, alternative explanations and untested assumptions. Held-out datasets should remain inaccessible to the discovery system until a hypothesis and analysis plan are fixed.

For physical sciences, preregistration can be a useful gate. Before the wet lab begins, the team should record the primary endpoint, inclusion rules, stopping criteria, expected effect and negative controls. This prevents an agent from endlessly reformulating success after seeing results. It also turns failure into useful information: a rejected hypothesis can update the search process rather than disappear from the record.

Finally, institutions need explicit human release criteria. A finding should move from internal conjecture to external claim only when it meets a defined standard for evidence, reproducibility and uncertainty. The threshold can be lower for a speculative research memo than for a clinical recommendation. What matters is that the threshold is declared in advance.

The next scientific race is institutional

AI will make scientific exploration more abundant. That is not inherently a problem. Science has always advanced by producing more possible explanations than any one person can test.

The difference now is speed. The organisations that benefit most will be those that invest in automated laboratories, agent-ready data, reproducible computational environments, adversarial evaluation and independent replication. Their advantage will not come from having the loudest conjecture machine. It will come from having the fastest credible path from conjecture to confidence.

That is the real competitive landscape. In the AI era, discovery may become cheap. Trust will not.