Language models don’t understand time, but they talk about it

Language models generate an answer in two seconds, but confidently state: "This task will take two weeks." Behind this strange contradiction lies something more fundamental than just an echo of training data — language models simply don't have what we call time.

How LLM sees time

Language models work token by token, predicting the next word in a sequence. They have no internal clock, don't perceive past and future as a continuum. For them, time is simply words that appear in training data alongside other words. When a model says "two weeks," it's not assessing the real duration of a task. It produces a statistically probable answer based on how often the phrase "two weeks" appeared in contexts similar to the current one. It's like remembering a phrase you once heard, but forgetting what context it was in.

The speed and estimation paradox

Here's the crux of the problem: the model generates an answer faster than any human can write a full answer to a complex question. Yet it confidently names time frames that completely don't match its own speed. This isn't a simple mistake. It's a structural feature of how language models work. They don't model the process of solving a task over time — they only predict what words should come next. A system based on this principle physically cannot "understand" time the way humans do.

Why this matters

This reveals several key problems in using LLMs:

• A model cannot honestly assess task complexity, only guess based on statistics
• Its answers about timelines are not forecasts, but hallucinations, probable patterns
• When planning projects with AI, you must account for the fact that the model physically cannot calculate real time
• For critical assessments, human review is needed, not just model predictions

"A language model doesn't have what we call time," — emphasizing the

fundamental gap between how models work and how people think about them.

What this means

This is the first article in a series about collaborative thinking between humans and LLMs. The conclusion is simple: language models are not mini-humans with a fast processor. They're a completely different system operating by different rules. Using them without understanding these features means expecting errors.