Next time you’re about to ask an AI chatbot to help you solve a hard problem, you might want to slow your roll.
Misryoum newsroom reported on new research presented April 14, 2026 at the CHI conference on Human Factors in Computing Systems in Barcelona. The headline result is pretty straightforward: people who waited to consult a chatbot until they had partially worked through a problem on their own performed better on a critical thinking task than those who used the chatbot from the start. But there’s a twist—when time is tight, using AI early can still give a boost.
The study was run by computer scientist Mina Lee of the University of Chicago and colleagues, who randomly assigned 393 people to one of eight categories. First, participants were split into two groups—those given sufficient time (30 minutes) and those given insufficient time (10 minutes). Then, within those time buckets, they were further divided based on when, or if, they could use the OpenAI’s GPT-4o chatbot: early, continuous, late or no access. Each group had roughly 40 to 50 participants.
After that setup, participants took on a role: they were told to act as city council members and decide, using a set of seven documents, whether to accept or reject a company’s proposal meant to mitigate a water contamination problem. Each person had to write an essay explaining their decision. It’s the kind of task that sounds simple until you’re staring at the clock—one participant in Misryoum’s newsroom imagination (not this study, obviously) might still hear the faint hum of their laptop fan while they try to connect the dots.
The researchers scored the essays based, in part, on how many valid arguments and textual references they contained, and found that participants given 30 minutes performed better across the board than those given only 10 minutes. The top essay scores went to participants who had enough time to complete the task and had access to the chatbot later in the process. In other words, the best performance wasn’t just “use AI,” it was “use AI at the right moment,” after some thinking had already happened.
When the team looked at memory for information in the provided documents, the most successful group was different again: participants with sufficient time who never had access to the chatbot. They also scored “myside bias,” measuring how many perspectives participants incorporated in their arguments. Here, the group with sufficient time and late chatbot access did best.
The pattern lines up, according to Barbara Oakley, a systems engineer and education expert at Oakland University in Rochester Hills, Mich. The results reflect two kinds of learning: slow, effortful reasoning, and fast, automatic thinking. Slow learning means building understanding carefully and weighing options; fast learning leans on habits and quick judgments with little reflection. Participants who had time to reason through the material on their own before using AI did best because they had already engaged in that slower learning, Oakley said.
Of course, real life is messier than a lab. Under time pressure, the “insufficient time” groups told a different story: among those four categories, the group with access to the chatbot early on scored the highest on their essays. Misryoum analysis suggests that early assistance may help people get unstuck just enough to produce a coherent response. But the study’s authors didn’t frame this as permission to rush.
Lee cautioned that the result doesn’t mean people should sprint toward AI. “When you are under time pressure and use AI to boost your performance, then you are basically risking [just taking and using the] AI’s framing, and that reduces the kinds of arguments that you make and your engagement with the documents or different pieces of information,” she said. You have to “at least be aware of what you’re signing up [for].”
That awareness—timing awareness, mental awareness—may be the practical takeaway. Lee says people need strong AI literacy and also a way to understand their own thinking patterns so they can weigh risks and benefits of chatbots in different scenarios and different points in problem-solving. “I think our work kind of targets time constraints as the first step towards [that] understanding.” Not a bad place to start, really. But the whole question—how to use AI without outsourcing your reasoning—still feels like it’s only half answered.
White House roadmap aims for lunar nuclear power by 2028
Artemis II astronauts return to Houston after splashdown off San Diego