Six thousand Tennessee middle schoolers spent 50 minutes practicing fractions. The students who learned the most were those an AI tutor forced to slow down.

That’s the finding in a new working paper from the National Bureau of Economic Research, reported by Jill Barshay and Kristin Fasiang at The Hechinger Report. Students using an AI tutor that required them to answer three consecutive questions correctly after a mistake scored about 3 percentage points higher than peers using conventional software. Lead author Philip Oreopoulos of the University of Toronto is careful not to oversell the results: The gains were modest and appeared mostly on the easiest problems. What stood out to me was the mechanism. Students in that group spent more time on each question than anyone else.

AI as a brake rather than an accelerator is the thread running through today’s edition.

Dana Goldstein asks in The New York Times what students lose when they stop writing. The research she gathers suggests they may lose some of the thinking writing is meant to develop. Cognitive psychologist Ronald Kellogg has compared drafting an essay to digging a ditch, while recent surveys estimate that between two-thirds and 90 percent of students use AI. One MIT study found that people who wrote with chatbots exhibited lower brain activity and had more difficulty remembering what they had written.

Some educators are responding by reintroducing difficulty on purpose. District Administration reports that Denmark will require high school students to defend every written assignment verbally. A Fort Worth teacher has gone almost entirely analog, while the University of Chicago Law School has banned laptops for first-year students.

Writing in Education Week, Texas principal Nicole Humlicek describes a 4-year-old who needed four attempts to write an “S” in the correct direction. Her argument is that the struggle was not an obstacle to learning; it was part of the learning. At least half of states now require cursive instruction again.

Not everyone accepts the friction argument. Grammarly’s Jenny Maxwell tells the Times that educators who withhold AI risk signaling that their coursework is irrelevant to the workplace. The Aspen Institute’s Janice Jackson similarly warns District Administration that taking technology away could put students at a disadvantage.

Michelle Miller, the cognitive psychologist featured on The Atlantic’s podcast, lands somewhere in the middle: The question is not simply whether students use the technology, but who remains in control when they do.

For policymakers, the University of Sydney offers one working model. Its two-lane approach requires some assessments to take place in person under secure conditions, while others are designed with the expectation that students will use AI. The goal is not friction everywhere. It is preserving friction where the work itself develops the skill being assessed.

And for anyone wondering whether students will tolerate the harder path, a junior at John Jay High School offered the simplest answer. After reconsidering her use of chatbots for literature essays, she told the Times: “They’re meant to be hard.”

Looking for your next opportunity in education? W/A Jobs features 3,100+ career opportunities from 300+ organizations across the education industry.