Every language model is trained on data gathered up to a particular point. The consequences of that boundary are less obvious than simply not knowing recent news.
The boundary is soft, not sharp
Training data is collected over a period and processed afterwards, so coverage thins out gradually rather than stopping cleanly on a date.
Recent events are also under-represented because commentary, analysis and correction accumulate for months after something happens.
A model may therefore know a topic exists while holding an early and possibly mistaken account of it, which is more difficult to detect than plain ignorance.
The model cannot see the edge
Nothing in the training process marks where the data stopped. The model has no internal signal distinguishing a well-covered topic from one it barely saw.
Asked about something after the boundary, it produces a plausible continuation based on patterns rather than declining, because declining is a learned behaviour rather than an automatic one.
This is the mechanism behind confidently wrong answers about recent developments, and it is why the confidence conveys nothing about accuracy.
Superseded information persists
Practices that were standard during the training period remain the model's default even after they are replaced or deprecated.
In technical work this shows as suggestions using older library interfaces, outdated security guidance or approaches that have since been discouraged.
The output is often functional, which makes the problem harder to spot than a straightforward error would be.
Search retrieval shifts the problem
Systems that fetch current material and supply it alongside the question can answer beyond the cutoff, because the answer comes from the retrieved text.
The model's own knowledge still shapes interpretation, and a conflict between retrieved material and trained assumptions can resolve either way.
Retrieval converts a knowledge problem into a source quality problem, which is more tractable but not automatically solved.
Dates in the prompt help less than expected
Telling a model the current date establishes that time has passed but supplies no information about what happened in the interval.
It does improve reasoning about elapsed time and can prompt more appropriate hedging, which is worth the small cost of including it.
The reliable approach remains supplying the current material directly rather than expecting the model to recognise the limits of what it holds.
Treating any claim about recent events as unverified, regardless of how specific it sounds, is the habit that catches the failures this boundary produces.