When Hypotheses Become Cheap: The Economics of Verification
AI can generate research possibilities faster than scientists can test them. If the bottleneck moves to validation, what should we optimize—and what should we count as progress?
Essay 3 of 4 — AI and the Future of Science. Once producing a plausible hypothesis becomes inexpensive, the scarce resource may no longer be an idea. It may be the evidence needed to decide which ideas deserve belief.
Imagine that a research agent generates one hundred candidate explanations in an afternoon. It can write code for each, report preliminary scores, and produce a paragraph defending the winner. At the end, the scientist has more candidates and perhaps a stronger-looking manuscript.
But how many hypotheses have actually been tested against evidence they were not chosen to fit?
This distinction points toward an economics of scientific verification. I do not mean that knowledge can be priced like a commodity. I mean that researchers operate under limits: time, experimental capacity, access to independent data, computational resources, and attention. AI changes the relative costs of activities within those limits.
Discovery and verification do not scale together by default
Consider two different operations.
Generating candidates means proposing methods, mechanisms, datasets, or explanations worth investigating. Large language models can make this activity extremely fast, although a large candidate set may contain repetitions and superficial variations.
Validating claims means exposing them to meaningful chances of failure: a held-out dataset, new physical measurements, preregistered tests, ablations that distinguish rival explanations, or an independent replication. Many such tests require resources that a language model cannot simply generate.
The National Academies’ 2019 report carefully distinguishes computational reproducibility—obtaining consistent results using the same data and computational steps—from replicability, obtaining consistent results across studies addressing the same question with their own data. [1] Both matter, but rerunning code does not automatically test whether an explanation is true beyond the original experimental setting.
An AI system could therefore produce perfectly reproducible code and still support a fragile scientific claim. The code may implement the wrong question, the test set may have influenced method selection, or the apparent effect may disappear elsewhere.
A proposal: validation debt
Software engineering uses the metaphor of technical debt for shortcuts that impose costs later. An analogous idea might help us reason about AI-assisted inquiry.
I propose validation debt as a provisional name for the burden created when publicly communicated knowledge claims accumulate faster than they can be independently challenged.
The important word is claims, not drafts. An untested idea clearly labeled as speculation need not impose the same burden as a confident result already cited by other papers.
Validation debt would rise when an experiment is selected after many undisclosed trials, when independent test data are repeatedly reused, or when an explanatory story is written only after the result is known. It could shrink when claims are narrowed, unsuccessful trials are disclosed, or stronger new tests are completed.
This is an analytical construct, not a measured quantity with an accepted definition. It should be useful only if it predicts something observable—for example, a larger fraction of published gains failing on an independent evaluation, increased effort spent reproducing reported findings, or the persistence of unsupported claims in follow-up work.
How could we test the idea?
An experiment could separate two budgets rather than treating “more AI compute” as one variable.
Give research agents the same task family and vary the number of candidate hypotheses they may explore. Independently vary the budget for validation: fresh datasets, independent seeds, prospective tests, or external experiment calls.
Record all attempts, not only the final manuscript. Require the teams or agents to distinguish exploratory choices from tests reserved for confirmation.
Then compare at least four outcomes:
- How many genuinely distinct hypotheses were proposed?
- How many survived tests not used for selecting them?
- How often did a published-looking claim fail independent validation?
- What was the total cost per verified finding, including failed experiments and human review?
The key comparison is not “AI or no AI.” It is whether allocating the next unit of effort to more ideas or to better tests improves the number and quality of findings that survive external scrutiny.
There are real obstacles. In some computational fields tests are cheap, while a biological experiment may demand weeks of laboratory time. Novelty is also difficult to operationalize, and a test may refute only one formulation of a hypothesis. Any meaningful measure would have to be field-specific and explicit about these limits.
Why conventional productivity measures are inadequate
Suppose two systems have identical hardware and produce the same number of polished papers. One ran five experiments per paper; the other ran five hundred and selected the best. If both report only the chosen experiment, their apparent success rates cannot be interpreted in the same way.
Similarly, a model producing more citations or explanations is not necessarily producing more trustworthy knowledge. Messeri and Crockett have warned that AI can create illusions of understanding while narrowing scientific inquiry. [2] The AI Scientist study reports substantial progress in automation but also documents failures in implementations, methodological rigor, and citations. [3]
The lesson is not that AI outputs should be discounted simply because they came from AI. It is that we need the denominator: the attempted searches, the decisions to stop, the evidence withheld from selection, and the attempts that failed.
Research already faced these problems before generative AI; the Open Science Collaboration’s 2015 replication project is one reminder that reproducibility concerns predate today’s agents. [4] AI may amplify familiar methodological weaknesses as well as help us address them.
Verification can itself be automated—but not assumed away
The most optimistic answer is to create AI agents that design stronger tests and detect their own mistakes. That is a serious research direction. Automated proof checkers, simulations, and laboratories can all expand verification capacity.
Yet a verifier may share the same assumptions, training data, or blind spots as the system whose proposal it evaluates. Multiple agreeing agents do not automatically provide independent evidence.
We therefore need to describe which sources of error are independent and what new observations could falsify the claim. Verification should not mean asking a second language model whether the first one sounds convincing.
What counts as scientific progress?
My provisional answer is not the number of generated hypotheses or finished manuscripts. A more defensible measure—at least for some discovery-oriented systems—is the number of independently supported, nonredundant findings per unit of total research cost.
Even that measure would be incomplete. Basic research also creates instruments, negative results, datasets, and concepts whose value may appear much later. We should not erase those contributions by reducing science to a single efficiency ratio.
Still, the ratio can expose a neglected trade-off. When hypotheses become cheap, choosing which evidence to buy becomes an intellectual task in its own right.
The final essay turns from the cost of evidence to a more unsettling boundary: must the knowledge that survives our tests also be understandable to us?
References
- National Academies of Sciences, Engineering, and Medicine (2019). Reproducibility and Replicability in Science.
- Messeri, L., & Crockett, M. J. (2024). Artificial intelligence and illusions of understanding in scientific research. Nature, 627, 49–58.
- Lu, C., et al. (2026). Towards end-to-end automation of AI research. Nature, 651, 914–919.
- Open Science Collaboration (2015). Estimating the reproducibility of psychological science. Science, 349, aac4716.
Validation debt and the proposed evaluation framework are hypotheses for further work, not established findings. The diagrams contain no measured research data. Generative AI assisted in drafting and editing this working essay.
Continue the inquiry
From Alchemy to Science—and from Science to AI?
Can Science Exist Beyond Human Understanding?
All research notes