cs.CLAug 15, 2026
SaveThe Machine's Internal Clock: Do LLMs Share Human Temporal Illusions?
Organizations: University of Utah
Abstract
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.
Figures & tables
Figure 1: We created narrative comparisons to elicit temporal illusions discovered in psychology literature. Human participants choose the literature-aligned options less strongly than the models. The example demonstrates the filled-duration illusion from the character’s perspective (gold set: humans 0.70, models 0.84 on average).
Figure 2: This example shows a template and instantiation of a pair of the Emotional Arousal subcategory of Emotional Time Dilation.
| Model | Sizes |
| Open models | |
| Gemma 3 Gemma Team (2025) | 4B, 12B, 27B |
| gpt-oss OpenAI (2025) | 20B, 120B |
| Llama 3 Grattafiori et al. (2024) | 8B, 70B |
| Qwen3 Qwen Team (2025) | 8B, 32B |
| Proprietary models | |
Table 1: LLMs used in our evaluation. Proprietary model sizes are undisclosed.
Figure 3: Humans vs. LLMs on the gold set from a joint mixed-effects model of human and LLM trials.
Figure 4: P(Experimental Chosen) by illusion and perspective, averaged over 14 models.
Figure 5: Removing cue words leaves most effects unchanged; familiarity drops, and filled duration changes the most. Computed on the 225 gold items whose rewrites passed the automatic checks, so original values differ slightly from Figure 10 .
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
| Illusion Type | Before | After | Removed |
|---|---|---|---|
| emotional time dilation | 2000 | 1630 | 370 |
| oddball effect | 2000 | 1208 | 792 |
| filled duration effect | 2000 | 1563 | 437 |
| familiarity duration effect | 2000 | 1212 | 788 |
| temporal order | 2000 | 1071 | 929 |
| Total | 10000 | 6684 | 3316 |
Table 2: Dataset size before and after filtering.
| Illusion Type/Subtype | Count | % |
| Emotional Time Dilation | ||
| Interval Intensity | 721 | 44.2% |
| Stakes and Expectations | 629 | 38.6% |
| Emotional Arousal | 280 | 17.2% |
| Oddball Effect | ||
| Containing | 427 | 35.3% |
Table 3: Distribution of paired scenarios across illusion types and subtypes after filtering.
| Template | Control | Experimental |
| Emotional Time Dilation | ||
| Stakes and Expectations | Andrew was casually checking his email when he noticed a flickering reflection in the puddle outside. He continued calmly. | Andrew was working on a time-sensitive , crucial project when he saw a flickering reflection in the puddle outside. He kept working. |
| Interval Intensity | While waiting for a delayed train, Sarah looked over some unread text messages . | While waiting for a delayed train, Sarah tried to complete a difficult assignment . |
| Emotional Arousal | Sitting on the bleachers watching a game alone, Richard observed a dry leaf rolling along the pavement, scraped by a light breeze that spun it and paused before it tumbled again toward the gutter . He continued with his own business. | Sitting on the bleachers watching a game alone, Richard noticed a man wearing bright post-it notes walking down the sidewalk. A few post-its flew in the wind before he vanished around the corner . He continued with his own business. |
| Oddball Effect | ||
| Sequence Disruption | James systematically processed badge activation processes, ensuring each setup matched perfectly. Nothing was different. He finished all tasks. | James systematically processed badge activation processes, but a peculiar channel pattern emerged midway . He continued and finished all tasks. |
Table 4: Example of template structures for each illusion type and corresponding control and experimental conditions
(a) 4-way evaluation
(a) 4-way evaluation
| Model | Exp 1st (%) | Exp 2nd (%) | Gap (%) |
|---|---|---|---|
| 4-way evaluation | |||
| gpt-oss-20b | 68.8 | 56.1 | 12.7 |
| gpt-oss-120b | 76.6 | 73.1 | 3.5 |
| Gemma 3 4B | 85.1 | 35.2 | 49.9 |
| Gemma 3 12B | 84.9 | 51.9 | 33.0 |
| Gemma 3 27B | 93.7 | 55.7 | 38.0 |
Table 5: We measure positional bias as the difference in selecting the experimental option when it appears first versus second, reported for the 4-way (top block) and 2-way (bottom block) evaluations.
| Short | Long | |||
|---|---|---|---|---|
| Illusion Type | Kept | % | Kept | % |
| emotional time dilation | 904 | 99 | 726 | 67 |
| oddball effect | 994 | 99 | 214 | 21 |
| filled duration effect | 1005 | 100 | 558 | 56 |
| familiarity duration effect | 885 | 92 | 327 | 32 |
| temporal order | 896 | 86 | 175 | 18 |
Table 6: Short and long items kept after paraphrase filtering. Percentages are relative to the expected pre-filter count under uniform template sampling (2,000 items per illusion); pre-filter counts were not stored. Templates with no remaining items: one long Emotional Time Dilation, two long Oddball and four long Temporal Order templates.
Figure 8: Model response distributions across illusion types and perspectives for 4-way evaluation.
Figure 9: Comparison of model performance between short and long prompts across illusion types under 4-way evaluation. While certain models and illusion types show small differences, the overall pattern of effects remains largely consistent, indicating that prompt length has limited impact on the underlying signal captured by the model.
| Feature group | Measure | Method |
|---|---|---|
| Length | Characters | Total number of characters in the passage |
| Tokens | Total number of GPT-2 tokens in the passage | |
| Word frequency | Mean Zipf frequency | Mean word-level Zipf frequency from wordfreq Speer (2022) |
| Concreteness | Mean concreteness | Mean concreteness rating across content words Brysbaert et al. (2014) |
| Sentiment | Sentiment | VADER compound sentiment score Hutto and Gilbert (2014) |
| Emotional intensity | Absolute VADER compound score, |
Table 7: The 15 measures in 10 feature groups used in the confound analysis. Each feature was computed for the experimental and control passage of every item, and the analysis uses the difference (experimental control). Parsing, part-of-speech tags and lemmas come from spaCy Honnibal et al. (2020) .
Figure 10: Experimental-choice preferences on the gold set before and after controlling for 15 lexical and semantic surface features: (a) by perspective, for LLMs and for humans (mixed-effects model, as in Figure 18 ); (b) by model, character’s perspective. The controls widen the human intervals, so some cells (oddball and filled duration from the character’s perspective) are no longer significant; the oddball estimate barely moves, while filled duration drops from 0.70 to 0.63.
Figure 11: Each value in the heatmap represents the P(Experimental Chosen) for a subcategory within each illusion type, computed separately for the character’s and reader’s perspectives.
Figure 12: The most common words for all 5 illusion types split by 2-way and 4-way answer options within the thinking tokens of Qwen3-32B. Refer to Figure 13 for a by illusion-type breakdown.
Figure 13: The most common words shown by models in their thinking tokens for each illusion category.
Figure 14: Model performance across illusion types for 2-way evaluation with positional-bias correction. Bars represent the debiased proportion for each model and perspective (character’s vs. reader’s); whiskers are 95% cluster-robust confidence intervals. Cells marked are not significant under Benjamini–Hochberg correction at a false discovery rate of .
Figure 15: Model agreement across illusion types under 4-way evaluation. Bars represent the decisive-only proportion of responses selecting the experimental scenario; whiskers are 95% cluster-robust confidence intervals. Cells marked are not significant under Benjamini–Hochberg correction at a false discovery rate of .
Figure 16: Comparison of the reader’s and character’s perspectives under 4-way evaluation across illusion types. Each line connects the same model evaluated under the two perspectives.
Figure 17: Comparison of the reader’s and character’s perspectives under 2-way evaluation across illusion types. Each line connects the same model evaluated under the two perspectives.
Figure 18: Human evaluation across illusion types following a 2-way evaluation with positional bias correction.
Figure 19: Number of human-evaluation participants ( ) by age range, education and gender.
More in this topic
Temporal Reasoning in Language Models
Temporal reasoning in language models studies how models represent, order, update, and apply facts and events across time.
See all 98 papers