How Do Large Language Models Think of Time?

AI can reconstruct the past and predict the future. But does it understand the present?

Large language models understand time through relationships between events, rather than through a continuously unfolding experience. They learn how events are ordered, how long things take, how dates relate, and how the past influences the future. They can reconstruct timelines, calculate intervals, and reason about what might happen next.

But a conventional language model does not have a continuously updating sense of the present. It needs information from its environment, such as a timestamp or an external clock, to establish what "now" means.

This reveals a distinction between knowing about time and experiencing its passage.

The Difference Between Sequence and Time

Imagine reading a biography that begins with someone's death, moves backward to childhood, and ends with an event from middle age.

The order of the sentences is not the order of the events.

Large language models must distinguish between these two structures. They process sequences of text, but the chronology described by that text might move backward, forward, or jump across centuries.

The Transformer architecture, introduced in 2017, uses positional information to distinguish where words or tokens appear in a sequence.

Think of a book with numbered pages. Page 100 comes after page 99, but that tells us nothing about whether the events described happened one minute or a thousand years apart.

Positional information works similarly. It helps a model recognize the structure of the text without providing a clock for the events being described.

Consider this passage:

"Sarah sold her company in 2024. She founded it in 2020, five years after graduating."

Although the sentences describe events in reverse chronological order, the model can reconstruct the sequence. It can infer that Sarah graduated in 2015, founded the company in 2020, and sold it in 2024.

The timeline comes from interpreting the relationships between events, not simply following the order of the words.

How Does AI Know What Time It Is?

Ask someone what day it is, and they can consult a calendar, check their phone, or use memories of recent activities.

A language model operates differently.

Its training provides extensive knowledge of calendars, historical events, seasons, and words such as yesterday, tomorrow, and eventually. But that knowledge does not establish the current date.

A model can explain what happened in 1995. It cannot reliably determine what happened yesterday unless recent information is available in its context.

Consider the question, "Who is the CEO of this company?"

The answer depends on when the question is asked. A correct answer from 2023 might be wrong in 2026.

The same problem applies to "last Friday." Without a reference date, the phrase has no fixed calendar meaning.

Modern AI applications address this by providing timestamps and access to current information.

The model can then calculate dates and describe events relative to the present.

But knowing today's date introduces another distinction: knowing how much time has passed.

Suppose you ask a chatbot a question at 9 a.m., close your laptop, and return at 5 p.m.

Eight hours have passed for you. You have accumulated experiences, formed memories, and observed changes in your surroundings.

In an ordinary chatbot system, the language model does not independently maintain a running account of those eight hours.

When you return, the application can provide the earlier conversation and a fresh timestamp. The model can calculate the elapsed time, but the calculation is not evidence that it experienced the interval.

An AI agent connected to monitoring systems can track events continuously. That capability belongs to the larger system, including its software, memory, and external tools.

Why Is Time Surprisingly Difficult for AI?

A model might accurately explain the fall of the Roman Empire yet calculate a project deadline incorrectly.

Consider a hypothetical contract signed on Friday, March 6, 2026.

Payment is due ten days later.

If the contract specifies calendar days, the deadline is Monday, March 16. If it specifies business days, excluding weekends and assuming no holidays, the deadline is Friday, March 20.

Four days separate the answers.

Now imagine the contract also specifies that payment must arrive before 5 p.m. Eastern time, while the recipient operates in California.

The model must interpret contractual language, count the correct days, and account for the applicable time zone. A plausible answer is not enough.

Researchers have investigated this difficulty directly.

A 2024 study published at NAACL found that language models sometimes contradicted themselves when reasoning about the timing of events, with inconsistencies appearing in at least 27.23% of predictions across the evaluated settings.

One revealing test asked models complementary questions about an event: did it happen before a particular date, and did it happen after that date?

Under the test's conditions, those answers should not both be yes.

Yet models sometimes produced precisely that contradiction.

The study examined earlier systems, including GPT-4, rather than establishing an error rate for all current models. Its underlying finding remains instructive: knowing many facts about events does not guarantee a logically consistent understanding of their chronology.

Can AI Distinguish the Future From the Possible?

Time introduces another complication when we move from describing past events to imagining future ones.

Consider three statements:

"The company launched its product in January."

"The company will launch its product in January."

"The company might launch its product in January."

Only the first describes a completed event. The second asserts an expected future event. The third describes a possibility.

A language model can recognize these grammatical differences. But whether the future claims are justified depends on evidence beyond the wording.

Suppose a company announces that its product will launch next month. A model might present that launch as a certain future event rather than distinguish the company's stated plan from a confirmed outcome.

Development delays, financial problems, or management decisions could prevent the launch.

This distinction matters because the future is not simply another section of a historical timeline.

Past events can be investigated through surviving evidence. Future events must be anticipated through assumptions, models, and incomplete information.

A useful AI forecast should therefore distinguish between scheduled events, conditional outcomes, and speculation.

The challenge is not generating a convincing description of tomorrow. It is identifying which parts of that description are supported by evidence.

Does Understanding Time Require Experiencing It?

For humans, time carries consequences.

We remember what happened, anticipate what might happen, and make decisions knowing our opportunities are limited.

Waiting can feel interminable. A vacation can feel brief. A deadline can change our behavior.

A language model can explain these experiences, recognize their significance in a story, and reason about their practical effects.

But explaining an experience is different from having it.

Whether artificial intelligence could possess subjective experiences remains an unresolved scientific and philosophical question. Accurate temporal reasoning alone would not establish that a system experiences the passage of time.

An LLM can describe a century in seconds, reason about tomorrow using yesterday's evidence, and reconstruct a lifetime from a few paragraphs.

Yet it might still need someone to tell it what day it is.

Can we truly understand time without experiencing its passage?

🧬 QuestionStrings to Practice: The Timeline Audit

Use this sequence when asking AI to reconstruct events, evaluate deadlines, or make time-sensitive recommendations.

What establishes the starting date? → Which events have confirmed dates, and which must be inferred? → What assumptions determine their sequence and duration? → Where could those assumptions change the outcome? → What must be verified before acting?

The goal is to separate established facts, calculated relationships, and uncertain predictions.

---

Explore another question tomorrow at QuestionClass. Ask one question a day forever.

← Home