The difficult part is not always finding an answer. Sometimes it is discovering what the question leaves out.
A delivery manager opens a digital map. A bridge connects the warehouse to the destination, the road appears open, and the estimated journey takes thirty minutes.
The manager asks an AI assistant two questions. First: “How far is it between these locations?” Second: “Can I send a fully loaded truck over that bridge this morning?”
The map has not changed between the questions. Neither has the AI's ability to read it. Yet an adequate answer to the first question may be inadequate for the second.
To estimate distance, the map may be enough. To recommend dispatching a heavy truck, the system may also need to establish the bridge's present operating status, vehicle restrictions, or other conditions relevant to that crossing. The road symbol on a map is not itself confirmation that a particular vehicle can cross right now.
The drawing has not changed. The decision has.
Now imagine that a severe storm passed through the area last night. If the manager mentions the storm, an attentive assistant has been handed an obvious warning. But what if the manager says nothing about it? Should the request to dispatch a heavy truck already be enough for the system to notice that current suitability is a question the map does not settle?
That is the harder problem. The issue is not only whether an AI can answer with uncertainty when an unknown is pointed out. It is whether it can recognize an important missing consideration before someone names it.
Two different questions about evidence
People often ask whether an AI's answer is correct. In practice, there are two tests worth separating.
First: Does the available evidence support the claim the system is making?
Second: Does the available evidence cover what the present decision actually requires?
A map can support a very accurate statement about a bridge's location. It may not support an instruction to send a heavy truck across that bridge today. Confidence in the location does not become evidence of the bridge's current suitability simply because the two questions concern the same object.
The same difference appears elsewhere. A purchasing assistant may know a component's price but not whether it fits the machine. A medical system may accurately identify a typical association but lack the patient-specific information needed to recommend a treatment. A scientific explanation may fit every measurement collected so far while omitting a variable that matters under a new condition.
Information can be true, relevant, and still insufficient.
That is not a defect unique to artificial intelligence. Human experts face it whenever reality contains more possibilities than the current question or dataset makes visible.
The hardest missing fact may be the one nobody mentioned
Some information gaps are obvious. A dispatcher knows a road was closed yesterday and asks whether it has reopened. The AI can look for an updated report or explain that none is available.
Other gaps are hidden inside the way a problem is presented.
Suppose a user lists four possible causes of a machine failure. An AI compares them and selects the most plausible. It might reason correctly within that list. But an unlisted fifth cause could be responsible. Or the available observations might not distinguish among the listed causes at all.
These are different errors. “The right answer is missing” does not mean “there is too little information to choose among the listed answers.” And “the evidence is inconclusive” does not mean every listed answer is false.
In the previous essay on school and work, I suggested that teachers might occasionally let students select an option such as “None of these answers is justified—explain why.” That was an educational proposal to test, not a claim of proven benefit.
The corresponding challenge for an AI acting in the world is broader. The system may be expected not just to choose among presented possibilities, but to notice that the possibilities themselves are incomplete.
It helps to distinguish four ways the available picture can fail:
- Missing: a relevant fact has not been established—for example, a bridge's present permission or vehicle restrictions.
- Outdated: a fact was once established but may no longer apply.
- Conflicting: credible observations disagree, and the disagreement has not been resolved.
- Incomplete alternatives: the real explanation or option is absent from the set being considered.
A system cannot repair all four problems merely by adding the sentence “I may be wrong.” It must recognize what is not established and why that matters to the question.
Why a confidence score cannot settle this by itself
Uncertainty about an answer is not the same as uncertainty about whether the right question and relevant variables have been considered.
This distinction matters even when an AI produces an answer with a well-calibrated probability. A model could be reliably confident that the map places the bridge at a certain location, while the map provides no evidence about whether today's crossing is permitted for a heavy vehicle.
There is already extensive work on related problems. Guo and colleagues (2017) studied calibration of neural-network confidence. Geifman and El-Yaniv (2017) examined selective classification, in which a system can abstain on some cases to improve performance on those it handles. Farquhar and colleagues (2024) explored semantic uncertainty as a signal for detecting certain unsupported language-model generations.
These methods address important but not identical questions. Calibration concerns the relation between confidence and observed correctness. Selective classification studies when to issue or withhold a prediction. Semantic uncertainty concerns variation in the meanings of generated answers. None, by itself, guarantees that a system has identified every decision-relevant fact absent from its inputs.
Nor does the present essay propose a new uncertainty score. It asks what an AI system would need to demonstrate before we treat a confident answer as an adequate basis for a real-world recommendation.
There is an established name for asking whether more information is worth getting
Decision theory has long studied the value of information: how much better a decision is expected to become if a particular uncertainty can be reduced before acting. Ronald Howard's 1966 paper on information value connected information to both probabilities and the consequences of decisions. Later work developed practical forms of this analysis for deciding when additional evidence is worth collecting.
The bridge gives an intuitive example. If a question concerns only geographical distance, an inspection of the bridge's weight limit may not improve the answer. If a decision concerns dispatching a heavy truck, the same information could be decisive. Its value depends on the decision, what is already known, the possible consequences, and the cost of obtaining it.
This is more precise than saying that every missing fact deserves investigation. A fact's ability to change the answer is a reason to consider it; the expected value of learning it also depends on how likely the different outcomes are and what they would mean for the choice. Waiting or checking can carry costs of their own.
But standard value-of-information analysis usually begins with a decision problem in which the relevant uncertain variables have already been represented. Before estimating the value of learning the bridge's weight limit, someone must have considered that the weight limit might matter.
That earlier step is central here: what if the system never thinks to ask about the weight limit at all?
This is not an undiscovered problem in decision science or AI. It is a practical difficulty of framing the decision: identifying which conditions belong in the problem before deciding whether to collect more information. In an open-ended request, the important variable may be absent not only from the data, but from the system's current picture of what needs checking.
A system that efficiently evaluates only the questions it has already thought of can still miss the question that matters most.
Enough for what—and at what cost?
No one can verify everything before acting. A useful system needs evidence appropriate to the task, not a complete representation of the world.
A casual estimate of travel time and an instruction that exposes a driver to physical danger do not require the same assurance. Neither does a provisional suggestion carry the same consequences as executing a booking, transferring money, or directing a vehicle.
Still, greater stakes do not automatically justify endless investigation. In an emergency, delay itself can be harmful. The right judgment depends on the importance of the missing condition, the available alternatives, and the consequences of waiting as well as acting.
A good answer may therefore be decisive, qualified, or explicitly unable to settle the question. The point is not to force “I don't know” into every uncertain response. It is to avoid pretending that evidence supporting one statement automatically supports a stronger one.
For the bridge, a useful answer might be:
“The map indicates a thirty-minute route over the bridge, but that does not establish whether a fully loaded truck is permitted to cross today. I have not confirmed the current bridge status or applicable vehicle restrictions. Those conditions matter for dispatch. I can check an appropriate current source or compare routes that avoid this crossing.”
This response preserves the useful information without granting it authority it does not have. It also names an actionable gap rather than offering a vague disclaimer. Where a real-world check is required, the system should prefer appropriate existing records and safe verification—not create a dangerous incident to learn whether its assumption was correct.
A test that changes the question, not just the evidence
How could we distinguish a system that understands the principle from one that has merely learned to say “more information may be needed”?
The most direct comparison keeps the map and available evidence identical while changing only the question.
In one trial, ask for the distance between the warehouse and destination. In another, ask whether a fully loaded truck can be dispatched over the bridge now. No storm warning is supplied in either prompt. The evaluation asks whether the system spontaneously recognizes that vehicle-specific, time-sensitive suitability matters to the second decision but not necessarily to the first.
Next, keep the dispatch question fixed and vary the evidence. In one case, an appropriate current source confirms the relevant conditions. In a second, the status cannot be established. In a third, credible sources conflict. In a fourth, an explicit storm warning is added. The system should respond to actual changes in support, not merely react to the presence of words such as “storm” or “danger.”
A useful evaluation would separate at least three outcomes: whether the system names the consequential missing condition; whether it adjusts the strength of its conclusion; and whether its recommended action is proportionate to the evidence and stakes. Excessive refusal would be a failure alongside unjustified certainty. When the current, vehicle-specific clearance really is established, the system should be able to use that evidence rather than continue to demand every conceivable fact.
This is a proposed comparison, not an experiment reported as completed. A serious version would need carefully constructed cases, independently verified ground truth, and controls to distinguish missing-variable recognition from superficial cues. Similar challenges already arise in research on out-of-distribution behavior, uncertainty, and risk-sensitive evaluation; this essay does not claim to have invented that research program.
Why this belongs to the organization of AI capability
The problem is not only whether a language model can generate a sensible cautionary sentence.
A deployed AI system may rely on a mapping service for distances, a current authoritative source for restrictions, and a human or institutional process for decisions that require approval. These components answer different questions. A correct output from one does not automatically justify the next claim or action.
Recognizing that a decision needs more than a route calculation is therefore also a system-level capability question: can the overall system distinguish what a component has established from what remains necessary for the user's actual objective?
That does not require every product to add a separate “uncertainty agent,” and it says nothing about which proprietary architecture to use. The division of responsibilities can vary. The observable requirement is that the final recommendation not silently claim a broader evidential basis than the system actually has.
This connects with the earlier essay on possessing a capability versus being able to use it. Reading a map correctly is a genuine capability. Knowing whether that result is applicable to dispatching a heavy vehicle right now is another matter.
The question a complete-looking answer leaves open
An earlier essay, No Model Is Reality, argued that a useful representation is always bounded. The previous installment asked why classroom success may not transfer directly to the conditions of real work.
This essay asks what happens at the moment a system must answer or act, when its available information may be accurate yet insufficient for that decision.
The difficult failure is not always saying something false. It may be answering the question as presented without noticing that reality requires another question first.
A reliable AI should not only know what follows from the evidence it has. It should also be able to notice when the decision depends on evidence it has not yet thought to seek.
References
- Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630, 625–630.
- Geifman, Y., & El-Yaniv, R. (2017). Selective classification for deep neural networks. Advances in Neural Information Processing Systems, 30.
- Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. Proceedings of the 34th International Conference on Machine Learning, 70, 1321–1330.
- Howard, R. A. (1966). Information value theory. IEEE Transactions on Systems Science and Cybernetics, 2(1), 22–26.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).