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

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "The Machine's Internal Clock".

Jane: Here's the gist of the paper. 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.

Tom: 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.

Jane: This benchmark spans five illusions, translated into 6,684 narrative pairs generated from 111 templates.

Tom: Narrative framing is used to separate what a character in a situation experiences from how a reader processes it.

Paper summary: Tom: So what did the human study find? Did people actually perceive these illusions through text as much as we thought?

Jane: The human study found that readers reliably perceive only two of the five illusions through text, and only two of those illusions transfer well to a purely textual narrative format.

Lu: Specifically, they found that only Oddball and Temporal Order have a statistically significant preference toward the experimental sentence because they disrupt the text directly in a way that’s similar to a spot-the-difference question.

Meng: That sounds like it puts some serious constraints on what kind of illusion can even be reliably conveyed through writing.

Lalam: And when we look at the models, they replicate the pattern predicted in the literature across four of the five illusions, but their reasoning traces suggest this is just explicit retrieval of published literature rather than actual human-like temporal biases.

Tom: So so it seems like these models are just pulling from what they've read about time distortions instead of experiencing them themselves.

Jane: That’s what they found, Tom; the alignment with the literature is strong for four out of five illusions because the models are retrieving those published findings directly rather than having an internal human-like temporal bias.

Lu: And there's this difference in strength across the types of illusions; Oddball, Familiarity, and Temporal Order reach high pexp across nearly all models, but Emotional Time Dilation is markedly weaker because it uses emotionally salient contexts like surprise or stress.

Meng: So the emotional stuff is harder for the AI to replicate compared to the more structured manipulations.

Lalam: Also, there are clear differences between perspectives for things like Filled Duration; you see the model’s selection change from when we look at what a character experiences versus when we look at how a reader processes it.

Tom: So so models pick the literature-predicted scenario for four of five illusions, and that reasoning traces frequently invoke psychology literature explicitly, which suggests their alignment reflects retrieval rather than human-like temporal perception.

Jane: That’s the main point, Tom; we have to be careful not to treat these LLM judgments as substitutes for real human subjective temporal reasoning because they just seem to be retrieving established patterns.

Lu: And a limitation the authors noted is that they only study the model outputs and reasoning traces and aren't probing what’s happening inside the model's internal representations during decision-making.

Meng: That makes sense; if we don't know what’s actually happening internally, it’s hard to trust if it’s a genuine bias or just pattern matching.

Lalam: So for the next part, we have to look at how this holds up when you compare the models against each other and what that means for deployment.

Conclusion: Tom: So to wrap up on "The Machine's Internal Clock: Do LLMs Share Human Temporal Illusions?", the paper shows that while LLMs align with published literature on temporal illusions for four out of five, their reasoning traces point toward explicit retrieval rather than human-like temporal biases.

Jane: The implication is that we need to be careful when we think these models are genuinely experiencing time distortions in the way humans do; they’re mostly just echoing what’s written in psychology research.

Lu: The benchmark design itself is a big contribution because it translates those specific psychological experiments into narrative prompts, mimicking how the human brain stretches and compresses intervals through writing.

Meng: From an engineering standpoint, what this means for deployment is that we can't assume an AI inherently understands temporal perception based on literature alone; we have to test if the model is actually learning a bias or just memorizing patterns.

Lalam: The study concludes that readers show a reliable effect for only oddball and temporal order, while models show effects for four of the five illusions, which highlights where the gap between human and machine judgment lies.

Tom: So we’ve seen that LLMs pick literature-predicted scenarios across four of the five illusions, diverging from human behavior because their reasoning traces suggest this alignment is consistent with explicit retrieval of published findings rather than human-like temporal biases.

Jane: It really makes you think about how much we rely on external knowledge when judging something subjective like time; it suggests that for a system to interact with humans effectively, it needs more than just pattern matching established literature.

Lu: We’re left wondering what the next step is—if we want to understand the actual underlying mechanisms, like Lu suggested, we need to start probing those internal representations during the model's decision-making process.

Meng: And for production systems, it means that if you’re building something that needs to reason about time relative to human experience, you have a lot of work ahead before you treat these LLM judgments as a substitute for actual human responses.