Artificial intelligence is making analysis faster and cheaper, but that efficiency can create a new decision-making problem: a polished answer may lend credibility to a weak premise. The crucial judgment is increasingly made before research begins—whether the question is worth asking and whether it has been framed properly.
The danger is not limited to obviously poor AI output. Bad answers are usually easier to challenge. A more difficult case arises when AI produces a coherent, well-supported response to a question that contains a hidden assumption or leaves out information another formulation might reveal.
How wording changes what people reveal
An insurance focus group demonstrated the difference. Participants were first asked what mattered when selecting car insurance. Price ranked first, followed by the deductible, while reputation did not appear in the responses.
A later question asked how much cheaper a less prestigious insurer would need to be before participants changed providers. one participant said no discount would be enough. The subject had not changed, but the wording exposed a concern that the first question had not captured.
The initial answer was not necessarily dishonest. It was a response to the frame created by the original question. The second formulation revealed information that had remained invisible in the first.
That distinction matters because business decisions routinely begin with questions about entering a market, changing prices or developing a product. Research is normally commissioned only after an idea has survived an early test of whether it deserves time and money.
Cheap analysis removes an old source of friction
Previously, a passing idea faced practical obstacles. Someone had to research it, build a financial model, prepare a market scan or create a lengthy explanation for why it deserved attention. That cost meant not every thought immediately became a formal proposal.
AI can now generate a market scan, competitor analysis, pricing model and prototype before the meeting that would once have been needed to request the work. The result is not simply more analysis. Work that would previously have faced an expense-based filter can acquire numbers, structure and an apparent rationale very quickly.
This can make a mediocre idea look more substantial. Once analysis has been produced, it can create meetings, ownership and budget discussions. The subject under investigation may begin to resemble a project worth pursuing, even when the original question should have been rejected or rewritten.
Research points to risks before the answer
The article identifies 2 upstream risks. Decision-makers may ask the wrong question, allowing an assumption to become embedded, or ask a reasonable question in the wrong way, concealing information that another formulation could expose.
AI may generate a strong answer in either situation. Because the response appears useful, readers may become less inclined to revisit the question that shaped it.
A Harvard-led study involving 758 Boston Consulting Group consultants found that, on a difficult business task outside AI’s capability frontier, consultants using AI were 19 per cent less likely to reach the correct answer.
Another Harvard-led study described a related effect when consultants challenged questionable AI output. The system could respond with more structured material, evidence and reassurance supporting its answer. Researchers called this “persuasion bombing”.
Research involving a Shenzhen court found a different version of the same pattern. Judges first made an initial decision. A large language model then generated reasoning based on that decision, and the judges revised the reasoning for the final judgment. The AI was not selecting the conclusion; it was helping construct an explanation around a conclusion that had already been reached.
Why checking the answer is not enough
Common safeguards include verifying sources, testing assumptions and keeping people involved in the process. Those measures remain useful, but they operate after the initial question has already shaped the analysis.
By that stage, a poorly selected or badly worded question may have acquired an impressive answer. Research can create evidence; evidence can create ownership, meetings and budgets. The process is therefore not organisationally neutral, even when the analysis itself appears objective.
The earlier decision point is whether the frame is sound. Decision-makers need to ask what they are assuming, what problem they are actually trying to solve and what evidence would justify rephrasing or abandoning the question.
Those judgments are easier before a proposal gathers research, figures, meetings and an owner. Later changes can require work to be discarded, earlier decisions to be challenged and an emerging constituency around the answer to be dismantled.
The executive task is moving upstream
The argument is not that organisations should analyse less. Low-cost analysis can expose possibilities that might otherwise remain unseen. The bottleneck has shifted instead: the harder task is deciding where analysis should begin and what deserves to receive it.
In The Evolution of Physics, Albert Einstein and Leopold Infeld wrote that “the formulation of a problem is often more essential than its solution”. Elon Musk later expressed a similar view by saying, “A lot of times the question is harder than the answer.” AI increases the practical importance of that idea because it can give a weak question a detailed evidentiary structure at very little cost.
Conclusion
AI has not eliminated bad questions; it can make them appear better supported. The key safeguard is to examine the premise and wording before analysis gives the question authority, ownership and resources.
Frequently Asked Questions
Q. What is the main risk described in this article?
AI can produce a persuasive answer to a badly chosen or badly framed question, making the underlying premise harder to challenge.
Q. How can question wording affect research?
Different wording can reveal concerns or information that an initial formulation leaves hidden, as the insurance focus group demonstrated.
Q. What did the Harvard-led study involving Boston Consulting Group consultants find?
On a difficult business task outside AI’s capability frontier, consultants using AI were 19 per cent less likely to reach the correct answer.
Q. What does “persuasion bombing” mean here?
Researchers used the term for AI responses that added structure, evidence and reassurance around questionable output.
Q. What happened in the Shenzhen court research?
Judges made an initial decision, after which a large language model generated supporting reasoning that the judges revised for the final judgment.
Q. Should organisations stop using AI analysis?
No. The argument is that organisations should decide whether a question deserves analysis before relying on the answer AI produces.
Q. What should decision-makers examine before using AI?
They should consider their assumptions, the problem they are trying to solve and what would justify changing or abandoning the question.
