The Time Constraints of AI Access Could Change How We Think

Written by Claire Fu

Imagine you are in a high stress work environment, and you were just assigned yet another project due in two days. Now, imagine that you are given the same assignment, due a month from today. How would you go about completing this project? Which scenario would you be more likely to use AI in?

In the race to understand generative AI’s impact on human cognition, one overlooked variable may be the most important of all: time. Jiayin Zhi, a current second year PhD student at the University of Chicago Department of Computer Science, has always been interested in human-centric design and the idea of technology designed to benefit humanity and society. Thus, she naturally entered the Human-Computer Interaction space, and joined the AI&Me Group at UChicago, led by Assistant Professor Mina Lee. Zhi observed that although many studies have examined the effects of AI on human cognition, one of the most pressing questions today, scientists often compare only “AI” versus “no AI,” a simple comparison that disregards how time may play a role in affecting AI use. What these studies found were inconsistent: AI use may be beneficial for certain tasks, but harmful for others. She set out to do a study that would measure the dynamics between AI use, time constraints, and human cognition.

“We don’t have empirical evidence on AI’s impact on people’s actual performance of critical thinking,” Zhi said. “Previous work included evidence from user research, including some interviews and some surveys, but those studies primarily rely on people’s self reported data. Then, there’s the added issue of social desirability bias. This is the research gap we want to fill. We wanted to provide empirical evidence about AI’s impact on people’s actual critical thinking performance.”

Zhi’s recent paper, Investigating the Effects of LLM Use on Critical Thinking Under Time Constraints: Access Timing and Time Availability, was presented at ACM CHI 2026. In the paper, she looked at the effects of AI on critical thinking through the lens of time. She conducted several pilot studies to identify that participants spend an average of 20 minutes to complete a cognitive task of writing an argumentative essay with a limited set of provided documents as potential sources of information. Approaching this research question with a factorial experimental design, Zhi first determined two categories of time availability for test completion: insufficient time (10 minutes) to complete the cognitive task, or sufficient time (30 minutes). Then, she split the availability of AI to those participants into four variables: no AI access, continuous AI access, and early or late AI access, where the LLM would only be available in the first third or last third of the entire duration of time for task completion.

Strikingly, Zhi found what she called a “temporal reversal”. Under insufficient time, participants that had early or continuous access to an LLM performed better than those that had late or no LLM access. Under sufficient time, this pattern is reversed: participants with late or no LLM access scored better in their essays. To better understand this phenomenon, Zhi began looking at the interaction logs that were being tracked on the interface.

“We were curious if the participants may just directly incorporate the arguments in the AI responses they receive, or if they were inspired by the AI responses,” Zhi stated. “Based on the analysis, participants having access from the start showed very minimal increase in their non overlapping argument with the AI, from insufficient to sufficient time, while participants attempting the task independently first showed substantial increase.”

These results were the first study to provide empirical evidence on how people’s critical thinking performance differs based on AI use, and on the role of time constraints in completing cognitive tasks with AI. It suggested that people with AI access from the beginning may not deliberate further on the argument despite sufficient task time. This was further supported by evidence that those participants, once beginning their writing process, showed less iterative consultation with the source documents. This means that, if given the opportunity to use AI from the very beginning, many people may subconsciously reduce their cognitive engagement on completing a task because of the ease of access to AI.

Zhi finds this research incredibly important in this age, now that there are many AI models that often exceed human capabilities on a number of different tasks. “What is the future of human thinking now that generative AI exists?”, she asks. Zhi hopes this study contributes meaningful evidence to the debate about the benefits and risks behind using AI. AI is great for outsourcing tasks to increase our own efficiency. On the other hand, there is concern that, if our thinking process is outsourced, it may not lead to better thinking outcomes.

For her next steps, Zhi is looking to further test the relationship between human cognition and AI by doing more investigations under real-world, high-stake environments. For example, she is interested in understanding how critical thinking of recruiters changes when making hiring decisions using AI. To learn more about her work, visit the recent Science News article on this paper here, and her personal website here.