Science1 publisher2 min readPublished
Knowledge workers who made AI argue back describe better work in a 45-person study
Researchers who interviewed 45 knowledge workers and read their AI chats say asking AI to challenge ideas improved performance. The study has no comparison group, so it documents the habit without measuring whether it guards against the skill erosion linked to heavy AI use.
The Scientist · Science desk
What happened
- All participants worked at one US public university and used generative AI in their core tasks of generating ideas, solving complex problems and designing new solutions.
- A marketing professional had AI invent customers who would dislike a planned accounting certificate program, and their objections surfaced a preference for cases drawn from several industries.
- An attorney used AI to hunt for obscure legal loopholes and ways companies might exploit them, turning up realistic scenarios of hard-to-detect unethical behavior by employees.
- The author also cites a separate study, one the author did not work on, finding creative writers produced better copy when using AI as a sounding board instead of a ghostwriter.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint Without a group that used AI the ordinary way, the study cannot separate the effect of friction prompts from the habits of people already inclined to write them.
- constraint The benefit as described depends on enough domain expertise to vet odd suggestions, so the results say little about junior staff who lack that grounding.
- decision Managers weighing the author's proposed AI task forces and internal boot camps have case examples to go on and no effect size to plan a budget against.
The worry behind this work has a name in the research the author summarizes: cognitive offloading. According to that research, people who use generative AI for work tasks can surrender to its quick, confident answers and settle for speedy but low-quality solutions. The same body of work, as the author describes it, shows the habit can erode the skills that make knowledge workers valuable [4]. The friction study is presented as a different view of that problem [3].
The team interviewed the workers and also analyzed the prompts and outputs from their iterative AI chats [2]. Those logs give the study a record of what each worker typed and what came back, in addition to what they recalled in interviews.
The central finding, in the author's words, is that workers who deliberately use AI to challenge their ideas "can significantly improve their performance" [3]. The article does not say how performance was judged, what it was compared with, or whether anyone's critical thinking was measured before and after. Nor does it give an effect size or show that the prompts cause it. People who think to ask an AI for disagreement may already be the ones least likely to offload their thinking. Interviews at a single public university cannot separate those explanations [1].
The author's proposed explanation is that AI models quickly become familiar with how a user prompts and reads output, then present information according to those patterns [9]. Asking for contrary material is meant to interrupt that. A research scholar asked for journal papers that used his concept in contrasting settings, and a lawyer, after following up AI-suggested citations, asked for different types of illustrative cases [13]. The author credits that second exercise with helping her write a strong legal brief [13].
Each case relied on the worker's own expertise. An operations research scientist adapted programs from signal processing and wireless communication that the AI suggested. The fields were new to him, but his grounding in supply network modeling showed him how to apply them [8]. The author advises workers to use their expertise to question surprising AI output [11], and notes that people sometimes see a response lacks originality even when it is factually correct [14].
The author also points to earlier research on human-AI collaboration in loan evaluations that had similar results [5]. In my view the habit is worth trying for anyone who knows a field well enough to reject a bad contrarian answer [11]. As evidence that it protects critical thinking, these interviews are consistent with the idea and do not test it.
What to watch
- Release of the study's full methods, including how performance was assessed and how many of the 45 workers actually used friction prompts.
- A randomized trial assigning challenge-style and answer-style prompting, with critical-thinking skill measured before and after.
- Details of the loan-evaluation research the author cites as having similar results, and whether it used a control group.