The Machine's Internal Clock: Do LLMs Share Human Temporal Illusions?

The gist The LLMs pick literature-predicted scenarios across four of the five illusions, diverging from human behavior, as reasoning traces suggest this alignment reflects retrieval rather than human-like temporal biases.

Benchmark Design

The authors present a new benchmark that uses temporal illusions to study whether LLMs produce judgments aligning with previously studied human distortions of time. This benchmark spans five illusions, translated into 6,684 narrative pairs generated from 111 templates. Narrative framing is used to separate what a character in a situation experiences from how a reader processes it. The benchmark evaluates both perspectives separately, allowing the study of whether models track a character’s subjective time, or a readers, or both.

Human Study Findings

The human study revealed that human readers reliably perceive only two of the five illusions through text and only two of the five illusions transfer to a purely textual narrative format. Specifically, Only Oddball and Temporal Order have a statistically significant preference towards the experimental sentence because they disrupt the text directly, creating a task similar to a ‘spot the difference’ question. The authors conclude that only manipulations detectable directly in the text produce reliable effects.

LLM Performance and Alignment

The LLMs evaluated on the benchmark replicate the pattern predicted in the literature across four of the five illusions. The reasoning traces suggest this alignment is consistent with explicit retrieval of published findings rather than human-like temporal biases. Specifically, ∼70% of responses explicitly evoke psychology research, suggesting that this alignment is consistent with retrieval of published findings rather than human-like temporal biases.

Illusion Strength and Perspective Differences

The strength of the illusion varies across types; Oddball, Familiarity, and Temporal Order reach high pexp across nearly all models (0.85–1.00, confidence intervals well clear of chance). Emotional Time Dilation is markedly weaker than the other types because it uses emotionally salient contexts (e.g., surprise or stress) while Oddball relies on more structured and systematic changes. Furthermore, there are clear differences between perspectives for Filled Duration: the model’s selection changes from the experimental option when moving from the character to the reader perspective.

Model Behavior and Uncertainty

Models exhibit systematic differences in uncertainty; smaller open-source models show stronger positional bias. GPT models tend to avoid decisions under ambiguity, while Qwen models are far less likely to avoid a decision, staying around ∼10% non-selective overall and primarily selecting can’t tell rather than defaulting to same. This suggests that positional bias and noncommittal behaviors may be linked to model uncertainty, potentially serving as heuristic strategies when a clear answer cannot be determined. The authors also note that for the Filled Duration effect, longer narratives are more likely to yield selections aligned with the experimental option.

Limitations and Future Work

A limitation of this work is that we only study the model outputs and reasoning traces, not the internal representations. Future work would require probing the model’s internal representations during decision-making to better test underlying mechanisms. The study concludes that readers show a reliable effect for only oddball and temporal order, while models for four of the five illusions. This work contributes to behavior modeling in LLMs, where capturing subjective time may be important for building systems that interact with humans.

Conclusion

The authors introduced a narrative benchmark of 6,684 paired scenarios spanning five temporal illusions to ask whether written text can evoke illusions. The human study finds no reliable evidence that they transfer, except where the manipulation is directly visible in the text. Models select the literature-predicted scenario for four of the five illusions, and reasoning traces frequently invoke psychology literature explicitly. This suggests that model alignment is consistent with retrieval of published findings rather than human-like temporal biases.

References

The authors present a new narrative benchmark that uses five temporal illusions to study whether LLMs produce judgments aligning with previously studied human distortions of time. This benchmark spans five illusions, translated into 6,684 narrative pairs generated from 111 templates. Narrative framing is used to separate what a character in a situation experiences from how a reader processes it. The benchmark evaluates both perspectives separately, allowing the study of whether models track a character’s subjective time, or a readers, or both.

Psychological research on temporal illusions has studied visual or auditory stimuli over short durations. In this work, we focus on five illusions, selected for their reliance on attention, density, novelty, and association (Eagleman 2008), which make them suitable for a narrative representation. They fall into two categories: (a) Arousal-based illusions (Emotional Time Dilation and the Oddball Effect), and (b) Information-processing illusions (Filled-Duration, Temporal Order, and FamiliarityDuration).

Emotional Time Dilation arises from physiological arousal and attention; researchers typically study it using visual cues of high-arousal stimuli, such as pictures of phobias or physical mutilation. The Oddball Effect demonstrates this: when an unexpected stimulus appears within a series of repeated expected stimuli, the duration of the unique item seems longer (Pariyadath and Eagleman 2007). Stimulus density and complexity affect illusions based on processing information. Intervals containing discrete elements, such as tones and visual markers, are perceived as longer than empty intervals of the same duration (Thomas and Brown 1974; Wearden et al. 2007). Neural encoding causes the Familiarity-Duration Effect. Familiar items cause a shorter neural response and subjective time (Manahova, Spaak, and de Lange 2020; Avant and Lyman 1975; Schiffman and Bobko 1977). Temporal Order Effects prioritize the sequence of events over the durations. As the brain diverts resources, the mental effort required to process sequence events decreases the accuracy of duration judgment (Brown and Smith-Petersen 2014).

Our evaluation focuses on deciding if LLMs exhibit these illusions, not if they should. They show that models and humans diverge: readers show a reliable effect for only oddball and temporal order, while models for four of the five illusions. Model behavior aligns with the psychology literature rather than with our human participants. Understanding how models interpret and represent temporal illusions can inform the design of agents.

We introduced a narrative benchmark of 6,684 paired scenarios spanning five temporal illusions to ask whether written text can evoke illusions<ref:2608.

1. Precise Temporal Illusion Detection via Narrative Text Analysis: AI systems can be improved to identify temporal illusions by analyzing reasoning traces for explicit literature retrieval, as ∼70% of responses explicitly evoke psychology research suggests that LLM alignment reflect[s] retrieval of published findings rather than human-like temporal biases.

2. Perspective-Aware Temporal Modeling: Systems can be enhanced to distinguish between subjective and objective time by modeling both character and reader perspectives, specifically noting that For the character perspective, where the narrative frames the experience as lived through, models favor the empty interval as longer while The reader’s perspective, evaluated retrospectively after processing, which allows for more nuanced temporal judgments.

3. Contextual Sensitivity Calibration: LLMs can be fine-tuned to recognize that only two of the five illusions transfer to a purely textual narrative format for human readers, thereby preventing over-reliance on text alone for complex temporal reasoning tasks.

Human perception of time is subjective. Temporal illusions show that the brain relies on context and relational cues when judging duration. Prior studies established these effects with visual and auditory stimuli. We ask whether written narratives alone can cause these illusions, and whether LLMs respond as human participants do. We introduce a benchmark of 6,684 paired scenarios spanning five illusions, each judged from the character's and the reader's perspective. In our human study (60 participants), participants' judgments matched the predicted directions for all five illusions in at least one perspective (7 of 10 conditions). Across 14 LLMs, models reproduce every effect humans show in the same direction, but on average more strongly. Most effects persisted even after cue words were removed and after controlling for 15 lexical and semantic features. These results show that models and humans agree in direction but not in strength, which cautions against treating LLM judgments as substitutes for human responses for subjective temporal reasoning.