The Machine's Internal Clock: Do LLMs Share Human Temporal Illusions?
This study reveals that while human readers generally fail to exhibit specific temporal illusions in text-only narratives, large language models surprisingly align with the literature-predicted illusions, likely due to retrieving psychological research rather than simulating genuine human-like temporal biases.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Time feels different depending on what is happening around us. A minute spent waiting in a doctor's office can feel like an hour, while an hour spent laughing with friends can vanish in a blink. Psychologists have long known that our brains do not keep time like a stopwatch, ticking away seconds with perfect precision. Instead, our sense of duration is built from context, attention, and how much new information we are processing. When we are bored or waiting, we notice the passing of time more, making it feel longer. When we are deeply engaged or surprised, our attention shifts away from the clock, and time seems to speed up. These are not just quirks of feeling; they are well-documented tricks of the mind, known as temporal illusions, where the brain constructs a sense of time based on experience rather than objective measurement.
As artificial intelligence systems become more capable of understanding and generating human language, a natural question arises: do these machines experience time the way we do? Do they share our subjective sense that time stretches and shrinks based on emotion or novelty, or do they simply calculate the duration of events like a computer program? Researchers at the University of Utah set out to answer this by testing whether large language models, the powerful AI systems behind many modern chatbots, fall for the same time tricks that humans do. They wanted to know if these models possess a human-like, subjective perception of time or if they merely mimic the facts about time they have read in books.
To investigate this, the researchers created a new test based on five specific psychological illusions that distort how we perceive time. These illusions rely on factors like emotional arousal, the presence of unexpected events, and the density of information. For example, one illusion suggests that a period filled with many small events feels longer than an empty period of the same length, while another suggests that a sudden, surprising event makes time feel like it has slowed down. The team translated these psychological experiments into written stories. They wrote thousands of pairs of short narratives, each pair describing two scenarios that lasted the same amount of time in reality but were framed differently to trigger these illusions. In one story, a character might wait in a quiet room; in the other, the same character might wait in a room filled with flashing lights and sounds. The researchers then asked human readers and the AI models to decide which scenario felt longer.
The study involved 60 human participants who read 25 of these story pairs each. The results for the humans were surprisingly limited. When reading the stories, people only consistently felt that time was distorted in two of the five cases: when a story contained a clear, unexpected oddity, or when the order of events was disrupted. In the other three cases, which required readers to imagine the emotional weight or internal state of a character, the human readers did not perceive a difference in time. The text alone was not enough to make them feel time stretching or compressing. The humans needed the illusion to be directly visible in the words to notice it; they did not internally simulate the passage of time in the way the psychological theories predicted.
The results for the artificial intelligence models were strikingly different. The researchers tested 14 different language models on the same set of stories. Instead of behaving like the human readers, the models consistently chose the scenario that the psychology literature predicted would feel longer. This happened in four out of the five illusions, including the ones where humans saw no difference. The models appeared to have a strong, subjective sense of time that aligned perfectly with scientific textbooks, even when the written stories did not actually support that feeling for a human reader.
To understand why the models were so different from the humans, the researchers looked closely at the internal reasoning steps the models took before giving their answers. They found that in about 70 percent of the cases, the models explicitly mentioned psychological research in their thinking process. They used phrases like "psychological studies suggest" or "research indicates" to justify their choices. This suggests that the models were not experiencing time the way a human does, nor were they simulating a character's internal state. Instead, they were retrieving facts from their training data. When asked about time perception, they recalled the established findings of psychology papers and applied them to the stories, effectively acting as a knowledgeable student answering a test question rather than a person experiencing a moment.
The study concludes that while these AI systems can accurately recite the rules of human time perception, they do not share the actual subjective experience. They know that a surprise makes time feel longer because they have read that fact many times, not because they feel the surprise themselves. This distinction is crucial for understanding how these machines work. They are not developing human-like consciousness or internal clocks; they are sophisticated pattern matchers that can retrieve and apply complex human knowledge with high precision. The findings suggest that when we ask these models about subjective human experiences, we are often getting a reflection of what has been written about those experiences, rather than a simulation of the experience itself.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.