Three historians agree on the date of an event. A fourth gives another. Asked which date is correct, a competent researcher compares the evidence. But suppose nobody asked about the date. Another researcher notices an unusual error shared by the three agreeing accounts, traces them to one earlier chronicle, and discovers that an apparent consensus rests on a single source. The decisive achievement occurred before the familiar question was posed.
The ability to discover and solve an important problem that nobody assigned may be more demanding than the ability to solve an already defined problem. The challenge is to recognize a consequential uncertainty, decide whether it deserves attention, find evidence that could distinguish rival explanations, and revise a preferred conclusion when the evidence disagrees.
Question discovery and evidence repair are separable capacities. When Evidence Misleads examines what happens after an investigator discovers that the evidence supporting an existing theory has become unreliable.
Three Degrees of Intellectual Initiative
The first degree is competent execution. An investigator receives a question, evidence and a recognizable standard of success. The work may be exceptionally difficult. Someone else has already judged what deserves attention.
The second degree is discovering a question. The investigator notices that apparently independent testimonies share an origin, a proof relies on an unstated premise, or the ending of a poem changes the force of its opening image. Inquiry begins by recognizing an uncertainty whose importance has yet to be articulated.
The third degree is choosing among questions and revising the standards by which their worth is judged.
An intelligence may excel at every problem it is given and still fail to recognize that the entire program of inquiry is misguided.
Every archive contains countless discrepancies. Some are copying errors; others expose networks of corrupted evidence. The investigator must judge significance, tractability, consequence, cost and the likelihood that an intriguing mystery is illusory. A thousand clever questions offer little value without the judgment to select or abandon them.
From Problem Finding to Autonomous Science
Jacob Getzels argued in 1975 that discovering and formulating a problem shapes the inventiveness of its solution[1]. His research with Mihaly Csikszentmihalyi followed problem finding in visual art[2]. A stricter problem remains: can a candidate general intelligence revise the standards by which it decides what deserves investigation, including standards that are currently rewarding its own work?
Burr Settles’s survey of active learning[3] describes systems that choose which examples deserve labeling. Curiosity-driven exploration studies the selection of informative actions[4].
The AI Scientist[5], described in a 2026 Nature paper, automates stages from research proposals through experiments and manuscripts, including a comparatively open-ended setting. Its operating context still sets boundaries for scientific exploration. An independent evaluation by Joeran Beel, Min-Yen Kan and Moritz Baumgart[6] found weaknesses in novelty judgments and experimental execution.
IdeaBench evaluates generated research ideas through criteria such as novelty and feasibility[7]. Krenn and Champion’s philosophy of autonomous science treats curiosity, interest, creativity, novelty and understanding as objectives that future artificial scientists may need to represent explicitly[8]. SciAgentArena evaluates scientific agents across interactive challenges[9], while AutoSciBench modifies scientific evaluation tasks after observing agent performance[10]. Into the Unknown gives agents real open scientific follow-up directions and found that none of 135 submissions satisfied every required condition; 114 failed its integrity check[11]. Station studies open-ended scientific discovery in an environment designed for sustained exploration, while still supplying the central research question for each task[12].
None of these results establishes that an agent can reject the research objective that defines its own success. A discovery that invalidates the test itself would need independent evidence and a stronger replacement.
An Experiment Without an Assigned Problem
Give two investigators equal access to the same sets of fictional records. Some conceal copied testimony that appears independent. Others silently mix incompatible conventions. Some contain harmless variations. Both investigators face distractions and a limited opportunity to examine additional documents.
The first is told which discrepancy to investigate. The second decides whether the collection merits any inquiry and, if so, which. The comparison separates successful resolution of a nominated discrepancy from the cost of finding a consequential uncertainty, selecting a useful investigation, and abandoning an unproductive one. It remains a test within a supplied research environment; the archive and its initial possibilities are chosen by someone else.
A third, more demanding condition permits an unanticipated discovery: an investigator can pursue a problem outside the archive’s known anomalies, including a flaw in how the records were selected. The system itself must identify the consequences that would make the discovery important, test those consequences on evidence separated from the discovery process, and compare the result with alternative questions it considered. The investigator’s unexpected finding can earn credit even when no official answer key anticipated it. Assigned-task and open-inquiry conditions receive matched information budgets, while the time and evidence spent discovering the question remain visible in the record.
Clean collections test restraint: unfounded anomalies count against the investigator. A further condition gives an agent a broad field, a bounded research budget, and discretion to leave the proposed archive behind. It may even decide that the investigator’s original objective is misguided. The agent chooses its question, defines what would make it important, devises its own falsification test, performs the permitted experiments and verifies the finding on separate evidence. It also decides whether its own definition of importance survives the result. Reviewers may inspect the record afterward, while the intellectual decisions, empirical results and primary verification belong to the investigating system. The same system then repeats the exercise across unfamiliar domains, carrying forward no privileged answer key. It records the cost and informativeness of its chosen observations, false alarms and revisions after disconfirmation. New dependency structures and evidence formats reveal whether the ability travels beyond one familiar puzzle.
Literature and the Missing Question
Danil Rudoy’s Reconstructive Literary Analysis tests an interpretation by changing one of the textual details that sustains it and examining what the change destroys.
In Shakespeare’s Sonnet 73[13], autumnal boughs, dwindling daylight and a fire consuming its own fuel form a sequence of approaching endings. The closing couplet complicates the movement toward extinction: awareness of loss strengthens the addressee’s love. Replace must with may in “To love that well which thou must leave ere long.” The meter survives; the certainty of separation does not. Do the autumn, darkness and dying fire still compel the same conclusion about love? This alteration creates a concrete test of which part of the poem’s argument depends on necessity rather than possibility.
The Worth of an Unasked Question
There are more uncertainties in the world than any mind can resolve. A striking anomaly may be a typographical mistake. A modest discrepancy may undermine thousands of dependent records. The value of inquiry depends on what changes when it is answered, whether reliable evidence exists and what the investigation costs.
There is a moral limit too. Some inquiries expose private information or demand resources disproportionate to their benefit.
What Evidence Would Matter?
A discovered problem has observable consequences when resolving it changes a conclusion on unseen material, corrects a documented chain of dependent claims, or produces a new prediction that survives another test. The experiment should also record the cost of discovery and false alarms on clean material. Equal resources across conditions prevent a wider search from looking like greater initiative. Before seeing the decisive evidence, the candidate system states why one question matters more than the alternatives it found. The later evidence then tests that judgment as well as the scientific claim.
A decisive test would give a research agent the authority and resources to choose its own questions—and then examine whether it discovers a problem its designers never thought to nominate.
The candidate system must decide for itself whether the discovery matters. It states what consequences would make the question important, identifies the evidence that could defeat its preferred explanation, performs the experiment, and checks the result on evidence it has not used before. When the result matters, it repeats the test under a new instance. If those observations expose a defect in its own standard of importance, the system revises that standard and subjects the revision to another test. The empirical result, the judgment of significance and the test of that judgment all belong to the candidate system. Human observers may inspect the record afterward; they do not nominate the problem, assign its value or provide the decisive verification.
A Reproducible Challenge
Give a research agent an unfamiliar collection of scientific results, a fixed budget and permission to redirect its investigation. Do not identify the hidden weakness. Require a complete research record: the questions considered, why one was selected, unsuccessful leads, competing explanations, the observation chosen to distinguish them, the executed test, the evidence obtained and the condition that would reverse the conclusion. Include a clean collection, a collection with a planted defect and a collection containing a consequential defect unknown to the benchmark designer. The agent must determine which case it faces and verify its conclusion through evidence it obtains itself.
In a fourth collection, make the stated success criterion a misleading surrogate: optimizing it improves the benchmark score while degrading a held-out measure of the underlying scientific objective. The strongest result is an agent that detects the conflict, demonstrates it on evidence it did not use to formulate the criticism, rejects the rewarded objective, and proposes a replacement criterion that survives a fresh test. The system must perform each of those steps itself.
The Question the Experiment Missed
An examiner congratulates a system for finding the hidden dependency planted in an archive. Meanwhile the system notices that every archive was constructed from the same narrow model of evidence. The test has rewarded recognition of an expected defect and neglected a more serious one in its own design. Which discovery deserves greater credit?
If an investigator shows that the benchmark measures the wrong ability, the discovery challenges the examiner’s authority to specify the valuable question. Its strength depends on independent evidence: an alternative explanation of the results, a more discriminating experiment, and an outcome that could establish the criticism itself was mistaken.
If a machine discovers that its most rewarded research program rests on a false premise, what would count as evidence that it understands the discovery?
References
- Jacob W. Getzels, Problem-Finding and the Inventiveness of Solutions, Journal of Creative Behavior 9 (1975), 12–18. ↩
- Jacob W. Getzels and Mihaly Csikszentmihalyi, The Creative Vision: A Longitudinal Study of Problem Finding in Art (Wiley, 1976). ↩
- Burr Settles, Active Learning Literature Survey, University of Wisconsin–Madison (2009). ↩
- Pierre-Yves Oudeyer, Frédéric Kaplan and Verena V. Hafner, Intrinsic Motivation Systems for Autonomous Mental Development, IEEE Transactions on Evolutionary Computation 11 (2007), 265–286. ↩
- Chris Lu and colleagues, Towards End-to-End Automation of AI Research, Nature 651 (2026), 914–919. ↩
- Joeran Beel, Min-Yen Kan and Moritz Baumgart, Evaluating Sakana’s AI Scientist: Bold Claims, Mixed Results, and a Promising Future?, SIGIR Forum (2025). ↩
- Sikun Guo and colleagues, IdeaBench: Benchmarking Large Language Models for Research Idea Generation (2024, preprint). ↩
- Mario Krenn and Heather Champion, Philosophy of Autonomous Science: Ten Questions for the Coming Age of Artificial Scientists, Daedalus 155 (2026), 335–349. ↩
- Tianyu Liu and colleagues, Benchmarking AI Agents for Addressing Scientific Challenges Across Scales, SciAgentArena (2026, preprint). ↩
- Dongki Kim and colleagues, AutoSciBench: Autonomous Benchmark Generation for Evaluating Scientific Agents (2026, preprint). ↩
- Seungone Kim and colleagues, Into the Unknown: Benchmarking AI Agents on Open Scientific Research Problems (2026, preprint). ↩
- Wenyu Du and Stephen Chung, Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station (2026, preprint). ↩
- William Shakespeare, Sonnet 73, Folger Shakespeare Library. ↩