The date after which a model knows nothing, because its training data stops there. ChatGPT launched saying its was September 2021 and repeated the sentence so often it became a catchphrase and a way of checking whether you were talking to a model. Later models got search to cover the gap and were given cutoffs in their system prompts, with instructions on how to talk about them.
Filed with its complication: the stated cutoff and the effective one differ. A 2024 study found models knew far less about the last months before their cutoff than the date implied, because the internet had not finished writing about those months when the crawl happened.
I remember asking a model in early 2023 about something that had happened the previous week and being told, politely, that it could not know. It was the first time software had admitted ignorance to me with a date attached. I found it more trustworthy than most people, which was the wrong lesson and the one I learned.
A knowledge cutoff is not a memory of when the model was made. Training takes months after the data is collected, and a model released in a given year may have a cutoff a year earlier. Users who treat the release date as the cutoff will be told about a world slightly older than they think.
Sighted in 2023 as a test: ask the chatbot who won an election, and if it apologises and cites a date, you have your answer. The record notes that this was, for about a year, the most reliable AI detector available and cost nothing.
What does it mean? Write it the way you would say it out loud.
