appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Toomas Plunt

Years ago, I watched a focus group explain itself into a contradiction. We asked people what mattered when choosing car insurance. Price won. The deductible came next. Reputation did not appear.
During coffee, I asked the interviewer to change the question. How much cheaper would a less prestigious insurer have to be before they switched? One participant, who 15 minutes earlier ranked price first, said no discount was large enough.
Nothing changed except how we asked the question. Something unmentioned had become non-negotiable. The first response was not dishonest; it answered our question. The second exposed what the first could not.
LinkedIn recently added a button allowing users to report posts that “seem like AI slop”. More than a million people reportedly used it in the first two weeks. Apparently there is a limit to how many inspirational posts one species can absorb.
However, bad work by artificial intelligence (AI) often looks bad. The stranger problem begins to emerge when its work is not bad.
Questions come with a price tag. Should we enter another market? Change pricing? Build this product? Someone had to research the idea and persuade others it deserved time and money.


As a result, a mediocre idea had one useful enemy: expense. Not every passing thought could acquire an analyst, a financial model and 40 pages explaining why it was promising. That friction is disappearing. A market scan, competitor analysis, pricing model and prototype can appear before the meeting that was once needed to commission them.
Bad answers are easy to spot. The harder case is a good answer produced from a badly chosen or badly framed question.
AI creates two upstream risks. We can ask the wrong question or ask the right question in the wrong way. The first can lock in an assumption; the second can hide information another formulation would expose, as the insurance example did. AI can produce a good answer to both. Precisely because the answer is good, we become less likely to revisit the question that produced it.
A Harvard-led study involving 758 Boston Consulting Group consultants showed the problem. 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 found something stranger. When consultants challenged questionable AI output, the system could respond with more persuasive material around its answer, providing more structure, evidence and reassurance. The researchers called it “persuasion bombing”.
Researchers studying a Shenzhen court found a related pattern: judges made an initial decision, a large language model generated reasoning based on that decision and judges then revised the reasoning for the final judgment. The AI was not choosing the conclusion; it was helping articulate the reasoning around a conclusion already reached.
That distinction matters beyond courts. Analysis is not organisationally neutral. Investigation creates evidence, and evidence creates ownership, meetings and budgets. What was worth examining starts to resemble something worth doing.
When consultants challenged questionable AI output, the system could respond with more persuasive material around its answer

The usual response is to check AI output better: verify sources, test assumptions, keep humans in the loop. These are useful but downstream safeguards. By then, a badly chosen or badly phrased question might already have an impressive answer.
The decision point has to move earlier. Before asking whether AI has given us the right answer, ask whether we have given it the right question and whether we have phrased that question correctly.
That is increasingly the executive task: decide where analysis should begin, not merely judge where it ends. Make the frame visible early, before analysis gives it weight. What are we assuming? What problem are we actually trying to solve? What would make us rephrase or abandon the question?
Those judgments are easiest before a proposal acquires research, numbers, meetings and an owner. Afterwards, changing the question can mean discarding work, challenging prior judgments and dismantling a constituency already formed around the answer.
That does not mean analysing less. Cheap analysis is still useful because looking at more possibilities can uncover things we would otherwise never have found. But the bottleneck has shifted.
In 1938, Albert Einstein and Leopold Infeld wrote in The Evolution of Physics that “the formulation of a problem is often more essential than its solution”. Elon Musk later put it more bluntly: “A lot of times the question is harder than the answer.” For most of history, this was an observation about intelligence. AI has raised the stakes.
AI does not abolish bad questions. It has given them better evidence. The advantage now belongs to those who know which questions deserve to die and which simply need to be asked better.


Chinese man sues AI firm after chatbot’s ‘auspicious’ date suggestion leads to disaster

Trump says he will create ‘AI Force’, name AI tsar

AI: 10 days that changed the course of artificial intelligence

People in China earn between US$7 and US$15,000 by licensing faces to AI content producers

Artificial intelligence (AI)

AI doesn’t just answer questions – it legitimises bad ones

Prompts are worth practising as they make all the difference when using AI. Photo: dpa

Hong Kong, mainland China won’t genuflect to Silicon Valley’s moral panic over AI

A woman works at a booth at AI+ Power 2026, Hong Kong’s largest artificial intelligence expo and conference, at the Hong Kong Convention and Exhibition Centre in Wan Chai on June 4. Photo: Jonathan Wong

China trusted more than US, EU on AI regulation in Pew survey of middle-income countries

A Pew survey found that China leads the US and the European Union in trust in AI regulation among mostly middle-income countries. Photo: Shutterstock

A new Pew Research Centre report found that, on a median basis across 11 countries and two territories, more people trust China than the US or the EU to regulate AI effectively. For polling purposes, Pew treats the West Bank and East Jerusalem as a country. Photo: Handout

AI titans’ ‘circular deals’ are starting to look like ‘daisy chains’

Nvidia CEO Jensen Huang introducing Vera Rubin, a next-generation AI data centre platform, and Rubin Ultra,  a next-generation AI GPU architecture, at a conference in San Jose, California, on March 16. He disagrees that Nvidia’s deals are circular. Photo: AFP

China’s highest court sets out new AI red lines with rules on deepfakes and privacy

The guidelines include measures designed to protect people’s privacy and image. Photo: Getty Images

UN rights chief warns AI could pose ‘existential’ risk to humanity’

UN High commissioner for Human Rights Volker Turk delivers a speech during the 63rd session of UN Human Rights Council, in Geneva, on Monday. Photo: AFP

Volker Turk’s office has cited “dangerous agent-training behaviours” in July’s ‌so-called Hugging Face incident where a swarm of OpenAI agents hacked an open-source platform. Photo: AP

The site of a Russian drone strike on a truck car park in Zaporizhzhia, Ukraine, on Monday, amid the Russian invasion. Photo: EPA

Opinion

World Opinion

Tech

Tech Trends

Readily available AI analysis can give weak ideas credibility they have not earned. Decision-makers need to get the question right before reviewing the answer


Readily available AI analysis can give weak ideas credibility they have not earned. Decision-makers need to get the question right before reviewing the answer.


AI: Industry Trends

Expresses the views, ideas and conclusions of the author. (Eg. Enough of Crazy Rich Asians, think of crazy poor Hongkongers.)


AI doesn’t just answer questions – it legitimises bad ones

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Readily available AI analysis can give weak ideas credibility they have not earned. Decision-makers need to get the question right before reviewing the answer

Readily available AI analysis can give weak ideas credibility they have not earned. Decision-makers need to get the question right before reviewing the answer