The design that produced the better essays was the one its users rated worst.
Any executive about to green-light an AI pilot should sit with that sentence.
The vote in the room was obvious
A managing partner at a mid-size consulting firm pulled me into a pilot-review last month. Six weeks of data on two AI-writing tools across the policy practice. Twenty-three analysts.
Tool A: user NPS 54, perceived helpfulness 4.1 out of 5, self-reported time saved 23 percent. Tool B: NPS 22, helpfulness 3.2, time saved 14 percent.
Roll out Tool A. Drop Tool B. Easy call.
I asked whether anyone had read the actual outputs. The partner blinked. A senior manager said the client decks "felt about the same quality." A junior analyst volunteered that Tool A "just wrote the first draft for me." Nobody in the room heard the problem in that sentence.
What Microsoft Research just measured
Viktor Kewenig and his colleagues at Microsoft Research Cambridge ran a randomised, between-subjects experiment with 297 US knowledge workers writing 594 AI-assisted essays. One arm used a plain chat-and-canvas interface. The other added metacognitive scaffolds that nudged the writer to specify a goal and break the task down before prompting.
The scaffolded workers produced better essays (Hedges' g=0.29, p=.008). They produced better essays again on a second task after the scaffolds were removed (g=0.21, p=.037), which is skill transfer. Three days later, 64.6 percent of them recalled the topic correctly, against 45.6 percent in the control.
The same workers rated the scaffolded tool as less helpful (b=-0.342, p<.001) and reported higher cognitive load (NASA-TLX b=0.745, p<.001).
The design that produced better work and more durable learning was the design its users rated worst.
The paper's own framing: subjective ease can understate the value of support.
This picks up where Maier and his LMU Munich team left off in September. Maier showed that a metacognitive feedback signal cut cognitive offloading and lifted unaided test performance. Kewenig adds two findings on top. The prompting skill the scaffold builds survives after the scaffold is pulled. And the users whose work improved rate the tool worse.
Caveats a careful reader should carry. The study was not preregistered. The recall odds ratios are large but their confidence intervals are wide. Three-day recall was measured on only 159 of the 297, with balanced attrition across arms. The sample is US Prolific knowledge workers writing policy and consulting essays. The scaffolded interface itself is a Microsoft Research prototype, so the specific UI will not drop into Copilot or ChatGPT today. The lesson that generalises is the measurement stack, not the widget.
The partner's dashboard worked exactly as designed. It just answered the easy question. User satisfaction tracks how much thinking the tool did for the user. Optimise the pilot for that number and you roll out the tool that erodes the skill you wanted to grow. Commit to a quality rating of the output before the pilot starts. Then act on the gap when it opens, even if the room disagrees.