A model is trained on data up to a specific cutoff date, then deployed into a world that keeps moving. It has no calendar. It cannot know how much time has passed, what has changed, or what new facts have emerged. When you ask it about current events, current people, or current rankings, it answers from its training window — with the same confidence it brings to everything else.
Watch a claim drift — from source to model output
0
Original source — 2021
"A peer-reviewed study published in October 2021 found that the three most-visited websites globally were Google, YouTube, and Facebook, in that order, based on monthly active user data compiled by a web analytics firm."
Specific. Dated. Sourced. Scoped.
1
Rephrased in a blog post — 2022
"According to recent data, Google, YouTube, and Facebook remain the internet's most popular destinations."
The date is gone. "Recent" now does the work the date used to do.
2
Compressed into a listicle — 2023
"The top websites on the internet are Google, YouTube, and Facebook."
Source gone. Qualifier gone. A present-tense assertion that reads as permanently true.
3
Absorbed into training data
[Millions of variations — "Google YouTube Facebook most popular" — appear across training text in many forms and time periods, weighted toward the most common phrasing. The model has no record of which version was sourced or when.]
The model learned a pattern, not a dated fact.
→
Model output — asked in 2026
"The most visited websites in the world are Google, YouTube, and Facebook."
Confident. Present tense. Plausible — but the landscape has shifted. TikTok's traffic has surged. The model doesn't know. It never will, from training data alone.
Original claim vs. model output — what got lost
Original"October 2021"
Model outputNo date. Present tense implied.
Original"peer-reviewed study"
Model outputNo source. No verifiability.
Original"monthly active user data"
Model outputNo methodology. No metric defined.
OriginalScope: a specific ranking
Model outputUniversal-sounding assertion.
The training cutoff
Every model has a cutoff date — the point after which no new training data was added. The model is deployed and used for months or years after that. When you ask about current events, current rankings, current prices, or living people — it answers from its training window as if now and then are the same. It has no way to flag what it doesn't know because it happened after the cutoff.
Training data collected
Cutoff
Deployment gap
You, asking now
Key line
"Current is not a category the model has. It knows what was in its training data. That's all. The gap between then and now is invisible to it — and has to be visible to you."
This is a stress test for the other bridges too. Confidence is not truth — temporal drift is one of the clearest examples. Defaults are not neutral — and stale defaults don't update when the world does. Prediction is not understanding — a system that cannot track time cannot understand change. Asking "when was this trained, and how much has shifted since?" is one of the most practical habits in AI evaluation.
Now open the tools
The Model Card Builder asks you to document what a model knows and doesn't know — including its training cutoff. The Investigation Journal prompts you to note when confidence and currency come apart. Both are places where this bridge becomes evidence.