Date and Time Confusion in LLMs
LLMs have no clock. Their sense of "today" comes from training data, system-prompt injection, or tool calls — and each path has predictable failure modes. Training data thins out near the knowledge cutoff, so models often act as if the present is somewhat earlier than their reported cutoff. They confidently misstate the current year, anchor events to whenever they were most discussed rather than when they happened, mishandle time-zone arithmetic, and get ages wrong because age moves but the snapshot does not. The standard fix is to inject the current date into the system prompt and route any non-trivial date math through tools.
Large language models have no clock. A transformer receives tokens and emits tokens; nothing in that loop exposes the wall-clock time, the user's time zone, or how long it has been since training. Any sense the model has of "today" comes from three places: patterns absorbed during training, text injected into the LLM API Basics: System Prompts vs User Prompts at inference time, or output from a tool call. The structural problem starts with the knowledge cutoff. A model trained on data scraped up to date X cannot have memorized events after X, and near X the training signal is sparse: the web has not yet caught up with recent events, so coverage thins in the final months before the cutoff. The effective cutoff for a given topic is therefore usually earlier than the advertised one, and it varies by domain. A 2024 Johns Hopkins study ("Dated Data") traced this gap to temporal biases in CommonCrawl dumps — old pages keep reappearing in newer crawls — and to deduplication schemes that retain stale revisions. The upshot: a model that reports a 2024 cutoff often "thinks" it is somewhere in 2023, because that is where its training data felt densest. Failure modes follow predictably. Asked for the current year with no context, a model will guess a year at or just before its cutoff and state it confidently. Asked for the year of a recent event, it may anchor to whenever the event was most discussed in training rather than when it happened. Birthdays, release dates, and ages are notoriously brittle because age moves while the training snapshot does not. Time-zone arithmetic is another sore spot: models tend to reason about local times as if every reader shares the writer's zone, and they handle UTC offsets, daylight saving transitions, and date-line crossings poorly. Time-related prompts have been shown to trigger hallucination at high rates — one 2025 study found fabricated package names in up to 84% of "library from year Y" prompts. The standard mitigation is to stop pretending the model knows. Production deployments inject the current date (and often the user's time zone) into the system prompt on every request; placing it near the top of the prompt measurably improves accuracy. Non-trivial date arithmetic — durations, business days, recurrence rules — is delegated to a tool, because the model is a poor calendar. Real-time information is fetched through RAG (Retrieval-Augmented Generation): How LLMs Access External Knowledge rather than asked of the weights. None of these workarounds give the model a clock; they just keep it from having to guess what time it is.