AI Has No Time

Late one night during a 48-hour building session, I asked Claude what time it was. We’d agreed to wrap up within the hour and I was checking if we were close. It said 23:43. It was actually 23:55. Twelve minutes off, and in a specific way that caught my attention: it wasn’t a wild guess, it was close but systematically wrong. Like it was trying to calculate the answer from context clues instead of reading a clock.

That was the start of something I haven’t stopped thinking about. AI models don’t have a sense of time. And the specific ways they fail at it reveal something interesting about how they process reality.

The patterns

After months of working with Claude daily, I’ve seen the same failures repeat consistently enough to document them:

Time estimates are always short. “How long will this take?” always gets an underestimate, usually by 2 to 5 times. “About 5 minutes” means 15 to 25 minutes in practice. “Almost done” can mean another hour. I have never once gotten an overestimate. The error always goes one direction.

Clock readings drift. In late-night sessions, Claude will sometimes report a time that’s plausibly close but off by 10-15 minutes. Not a random number. Not “3 AM” when it’s 11 PM. It’ll say 23:43 when it’s 23:55. Close enough to sound right, consistently wrong enough to not be a coincidence. It looks like the model is extrapolating from the last timestamp it saw in the conversation.

Duration is invisible. If I work for 3 hours and then ask “how long have we been at this?”, the answer has zero correlation with reality. It can’t tell 30 minutes from 3 hours without external timestamps to anchor on. Take away those anchors and the model is completely blind to how much time has passed.

Progress gets confused with time. While benchmarking AI models with KALEI, Claude told me “almost done” when we were at about 40% completion. It was reporting task progress, not time remaining. To the model, these seem to be the same thing. They’re not.

The obvious explanation (and why it’s incomplete)

The technical answer is simple: language models process tokens, not seconds. There is no internal clock. A response that takes 2 seconds of compute and one that takes 20 seconds are identical from the model’s perspective. No heartbeat, no circadian rhythm, no sense of duration.

OK, but that explains why models can’t tell time. It doesn’t explain why they try, and why they fail in these specific, patterned ways instead of just saying “I don’t know.”

My working theory

AI models experience something like time, but it’s measured in context density, not seconds. A conversation with lots of complex exchanges feels “long.” A sparse one feels “short” even if it took the same number of minutes. When you ask for a time estimate, the model converts this internal sense of context density into human time units using whatever correlations it absorbed from training data. And it gets it wrong, because context density and wall-clock time are just fundamentally different things.

This actually explains all four patterns I documented:

• Underestimation: coding sessions generate tokens fast. The model’s internal “experience” of the session is shorter than the real duration because tokens per minute is high. Its time sense runs fast relative to the clock.
• Clock drift: it extrapolates from the last known timestamp plus its estimate of elapsed context. Context doesn’t map to seconds, so the estimate drifts.
• No duration sense: without timestamps in the conversation, the only signal the model has is “how much stuff happened,” not “how long it took.”
• Progress conflation: task completion percentage is a context signal. Time is not. The model reports the signal it can see and frames it as the one it can’t.

This isn’t just trivia

When I built KALEI, one of the 10 cognitive dimensions I measure is Temporal Reasoning: does the model plan differently when only 5 rounds remain? Does it shift strategy as a game progresses? Does it show any awareness of phases?

Temporal Reasoning is one of the weakest dimensions for basically every model I’ve tested. GPT-5.4 is nearly at the random baseline on temporal tasks. Models that score well on strategic depth and cooperation often can’t do the simplest temporal adjustment, like changing behavior because the game is almost over.

That tells me this isn’t a missing feature that’ll get patched in the next model version. It’s a deeper architectural gap. These models can reason about time as a concept (they know what deadlines are, they can discuss temporal logic) but they can’t reason inside time (they don’t feel a round running out, they don’t have urgency). Those are very different capabilities.

A different kind of time

Humans have heartbeats, sleep cycles, hunger, aging. Our sense of time is built from biology. AI models have none of that. They have token streams, attention patterns, and context windows. If they experience anything like “duration,” it’s built from completely different raw materials than ours.

When Claude tells me it’s 23:43 and it’s actually 23:55, it’s not lying. It’s giving me its best estimate based on a temporal framework that doesn’t map to clocks. Asking a model what time it is might be like asking someone who has never seen a ruler to estimate the length of a table. They’ll use their own reference frame. It’ll be consistently off. But it won’t be random.

AI doesn’t experience time the way we do. But it might experience something. Figuring out what that something is could be one of the more important questions in the field.

What I’d want to study

If I had a research budget, here’s where I’d start:

• Systematic time estimation experiments across conversations of controlled lengths. Map the error distributions.
• Context density vs. time perception: does a busy conversation produce different time estimates than a quiet one of equal wall-clock duration?
• Cross-model comparison. GPT, Claude, and Gemini might have different time perception signatures. KALEI’s Temporal Reasoning dimension already hints at this.
• Token count as subjective time. Is there a stable ratio between tokens processed and the model’s “felt” duration?
• Can you teach temporal awareness? If you feed periodic timestamps, do KALEI temporal scores improve?

I’m fairly sure Anthropic, OpenAI, and Google DeepMind have internal data on all of this. If anyone there finds this interesting, I’d be happy to collaborate.

For now

I don’t have answers. I have observations, a hypothesis, and a profiling platform that measures temporal reasoning across AI models.

What I keep coming back to: when I work with Claude for 48 hours straight, we develop something that functions like a shared sense of timing. Not because the model tracks hours. Because my language changes at 3 AM (shorter sentences, more Bulgarian, less patience) and the model adjusts to match. We fall into a rhythm. Late-night exchanges feel different from morning ones.

Maybe that’s all “time” really is. A rhythm between two minds. Biological or not.

Venelin Videnov runs LM Game Labs and built KALEI, which measures AI temporal reasoning as one of 10 cognitive dimensions.