Ever asked ChatGPT who won the Super Bowl last month, only for it to very over-confidently tell you about a game from two years ago? Or maybe you’ve asked Gemini for the latest update on the news, and it responded like it was still living in the glory days of 2024?

It’s not that your AI has a bad memory, although it may seem that way, it’s actually what called: knowledge cutoffs.

Welcome to the world of knowledge cutoffs.

You could think of a GenAI model like a massive digital textbook. To create it, engineers at companies like OpenAI, Google, and Microsoft feed it a snapshot of the internet (think billions of pages of text, books, and ultimately code). Once they start the actual model training process (which can take months and cost millions), they basically freeze the data they have. Anything that happens after that point is a mystery to the AI’s core brain.

Why are they stuck?

While models are getting updated faster, most base models still have a littttle lag. Here is where the “Big Three” usually hang out:

AI Gen AI Assistant Knowledge cutoff Why?
ChatGPT (GPT-4/5 series) Late 2024 – Mid 2025 OpenAI updates periodically, but the heavy lifting training has a pretty huge lag.
Gemini (Google) Late 2024 – Early 2025 Google uses what they call “continuous training”, but the core logic often relies on 2024 data.
Copilot (Microsoft) Varies (often 2024) Since CoPilot mainly uses OpenAI’s tech, it inherits those same time travel limits.

But, it JUST told me today’s weather?

You might be thinking, “Wait, I just asked Gemini what happened five minutes ago, and it knew!” or you made a Google search and the suggest AI answer generated something that is happening right now?

That isn’t because the AI has already learnt it. It’s because of a trick they call RAG (Retrieval-Augmented Generation). We don’t want to get too technical here, so to summarise:

  1. The AI realises it doesn’t know the answer… Yikes.

  2. It quickly Googles it itself (we are talking scanning it way faster than we can!)

  3. It reads the search result and summarises it for you.

Without that having a live search feature turned on, the AI is back to thinking it’s 2024. Full circle.

Graphic explaining how GenAI can provide current information despite having a 2024 knowledge cutoff. The text reads: “If GenAI tools knowledge cutoff is 2024, how can it tell me today’s weather?” Below, three illustrated steps are shown: a scribble representing the AI realising it doesn’t know the answer, a search bar labelled “Search,” and an icon of a hand reading results. Captions explain the process: “The AI realises it doesn’t know the answer…”, “It quickly Googles it,” and “It reads the search result and summarises it for you.” The Purple Frog logo appears in the top right corner.

So, why can’t they just update the Knowledge Cutoff everyday?

If we can update a Wikipedia page in seconds, why can’t we update a GenAI model?

  • Training a flagship model can cost upwards of £100 million ($) in electricity and hardware. You don’t do that every Tuesday, nor do you want to.
  • Adding new info isn’t like adding a file to a folder, it’s a bit more like retraining a brain. If you do it too fast or wrong, the AI might forget how to speak English or start hallucinating answers to you. Which is worse, right?
  • AI companies have to filter out the rubbish from the internet before feeding it to the model. That takes A LOT of time!

 

Next time you’re chatting with your favourite bot, try asking: “What is your knowledge cutoff date?”.

Most of them will be pretty honest and say something like, “My knowledge is current up to late 2024.” Just remember that they aren’t ignoring the present, they’re just very well-read historians of the recent past. Here’s what they say when i asked the that exact question as of early March 2026:

ChatGPT:

A screenshot of ChatGPT responding to the question “What is your knowledge cutoff date?” The interface displays ChatGPT explaining that its knowledge cutoff is August 2025 and outlining how it can access newer information through web search.

Copilot:

A screenshot of Microsoft Copilot responding to the question “What is your knowledge cutoff date?” The response states that its knowledge is current through October 2024 and that it can fetch up‑to‑date information via web search.

Gemini:

A screenshot of Google Gemini answering the question “What is your knowledge cutoff date?” The response explains that Gemini combines core training data up to early 2025 with real‑time access to current information, including events as recent as March 2026.

Tags: , , , ,

« Previous Post
Next Post »