A colleague asks an AI assistant to “make this report suitable for management”. Seconds later, the document is shorter, smoother and confidently wrong for the situation. The assistant guessed what management needed: perhaps a decision, perhaps a status update, perhaps evidence for a budget. The prose improved while the task remained undefined.
Generative AI has made immediate completion feel like helpfulness. But when one missing choice would change the whole output, the most capable first response may be a question. The design challenge is not to make AI endlessly inquisitive. It is to spend a small “question budget” where one extra turn can prevent a costly wrong branch.
Fluency hides a branching problem
An underspecified request is not merely missing text. It contains several plausible futures. “Prepare slides for Thursday” could mean a five-minute executive decision, a technical review or a public introduction. Each interpretation changes the length, evidence, language and acceptable uncertainty. Once the system silently chooses a branch, every polished sentence makes that assumption harder to notice.
This is one reason ambiguity is different from ordinary factual uncertainty. The missing information may exist only in the user’s intention. More documents or a larger model cannot discover whether the user wants to persuade, compare or decide unless the surrounding context already makes one interpretation dominant.
Research systems increasingly treat clarification as a selective action rather than a conversational ornament. The 2024 CLAMBER benchmark, built from 12,000 examples covering several forms of ambiguous information need, found that the tested language models had limited practical ability to identify ambiguity and generate useful clarifying questions. Chain-of-thought and few-shot prompting brought only marginal gains in that benchmark. The important lesson is not that every current product will fail identically; it is that fluent completion is not evidence that the intended task has been identified.
When is a question worth asking?
A useful mental model compares two costs. On one side is the interruption cost: another turn, a small delay and the risk of annoying someone who expected a quick answer. On the other is the expected cost of guessing: the chance that the assistant chooses the wrong interpretation multiplied by the effort, consequence and reversibility of that mistake.
A clarifying question earns its place when the second cost is larger. Asking whether a translation is for a friend or a legal filing is worthwhile because the register and risk differ sharply. Asking which of two harmless greeting styles someone prefers may not be. Urgency matters too: in an emergency, a safe provisional action plus an explicit assumption may be better than waiting.
A 2025 NAACL study, “Clarify When Necessary”, frames the decision in similar terms. Its method estimates how widely possible user intentions are distributed and weighs that against the user’s preference for speed or care. Across question answering, translation and natural-language inference, the authors’ intent-based uncertainty method was better than the compared uncertainty approaches at identifying cases likely to benefit from clarification. This is research evidence about a framework, not proof that a general assistant can reliably infer every real user’s priorities.
The question itself must also divide the problem. “Can you provide more detail?” transfers the work back to the user. “Is this report meant to request a decision or document progress?” identifies the fork and offers answerable options. A good clarification changes the output materially, uses language the person understands and asks only for information the system cannot safely infer.
One question, then visible assumptions
A practical interface could adopt a one-question budget for ordinary work. Before generating, the assistant would check for a missing variable that is both consequential and answerable: audience, purpose, jurisdiction, deadline, risk tolerance, output format or authority to act. If one dominates, it asks. If several remain, it asks the question that removes the most important branches.
After the answer—or when no question is justified—the system should state any remaining material assumptions near the output. This prevents clarification from becoming a false certificate of shared understanding. One answer rarely resolves everything, and a user should still be able to correct the frame.
Recent benchmark work is exploring this balance directly. The 2026 AskBench study evaluates both intent-deficient questions and questions containing false premises, rewarding targeted clarification without sacrificing final task performance. Its reported experiments show improvements from rubric-guided training, including on unseen domains. But benchmarks and simulated user replies cannot establish how often questions will help in a particular workplace. Real tasks bring time pressure, status differences, tacit knowledge and people who may not know the answer themselves.
The strongest objection: friction can become obstruction
An assistant that asks before every action would be exhausting. It could turn small requests into forms, use caution as a substitute for competence and burden people who chose automation precisely to reduce coordination. Repeated questioning can also reproduce unequal power: a system may accept terse instructions from insiders while demanding more explanation from people whose language or needs differ from its training patterns.
That is why the goal cannot be “more clarification”. It must be less avoidable rework and fewer hidden consequential assumptions. Systems should learn from corrections, but not silently convert them into permanent personal profiles. They should let people choose modes—fast draft, careful work, high-stakes review—and allow an immediate “make a reasonable assumption” response. For consequential domains, the question policy should be designed with the people who bear the errors, not only those who measure throughput.
This also preserves responsibility. An AI question may reveal a missing decision, but it cannot decide who has authority to make it. If a team has not agreed what a report is for, the assistant has exposed an organisational gap rather than solved it.
Run a twenty-task question audit
A team can test the idea without changing software. Collect twenty ordinary, non-sensitive tasks previously given to an AI assistant: summaries, emails, comparisons, plans or analyses. Remove any confidential material. For each task, run two workflows:
- Ask for the output directly.
- Add: “Before answering, ask at most one question only if a missing detail would materially change the result. Otherwise proceed and state any important assumption.”
Have the task owner score both outputs without being told which workflow produced them. Record usefulness, minutes of rework, whether the question was answerable, whether it changed the result and whether it was unnecessary. Also record high-cost assumptions that neither workflow surfaced.
The experiment may show that questions help only with certain task families. That is valuable. The team can then create a small clarification rule for those tasks—for example, always establish audience before drafting a briefing, or jurisdiction before summarising obligations—rather than adding a generic prompt everywhere.
This continues a wider Alkemata principle: human–AI agreement should not be the goal. Nor should frictionless completion. The better objective is a workflow in which the important branch is visible before effort accumulates behind it.
The next time an AI produces an elegant answer to a vague request, do not judge only the answer. Ask whether there was one question that could have changed the path. If there was, that missing question is part of the system’s output too.
What did this make you think?
Choose a reaction or write a comment. Comments are reviewed before publication.