What is a knowledge cutoff?
Updated · By Robert Breen
A knowledge cutoff is the point in time when an AI model's training data ends. The model knows nothing that happened or changed after that date unless a tool looks it up or you tell it, so its answers about recent events, prices or products can be out of date.
Why it matters for a small business
For a business, the most important things are usually the newest: this month's offer, this season's prices, the service you launched last week, a rule that just changed. Those are exactly the details a model cannot know from training. Ask without supplying them and it either says it does not know or, worse, fills in something that was true once or never was.
Some AI products can search the web, which helps with public news but still will not find your private price list or internal policy. The reliable fix is the same either way: put current facts in front of the model every time it needs them, through the prompt, a knowledge file or a tool.
In a real lesson: Build a Custom GPT Ad Copy Writer for Search and Social Ads
In the Custom GPT ad copy lesson, you build an ad writer for Harbor & Pine Coffee Roasters, a made-up small-batch roaster. Offers and products change all the time in a business like that, and no model's training data could contain Harbor & Pine's current lineup.
So the instructions you paste say: "Only use products, prices and offers from the Brand Voice & Ad Guide. Never invent a discount or a sale." Then you use Upload files to attach Harbor-Pine-Brand-Voice.pdf, which the voice-over describes as containing tone rules, character limits, current products and offers, claims you never make and sample ads.
The guide is the answer to the cutoff problem. When the offer changes, you update one document instead of hoping the model somehow knows. The instruction to never invent a sale covers the gap if something is missing.
Try this lesson free or read the step-by-step guide.
Common confusions
Knowledge cutoff vs web search
The cutoff is a property of the model. Web search is an extra tool some apps add, which can fetch newer public pages. A model with search can still have an old cutoff, and search never sees your private files.
Knowledge cutoff vs hallucination
A cutoff means the model lacks recent information. A hallucination is the model stating something false with confidence. A cutoff often causes hallucinations when the model guesses instead of saying it does not know.
Tips
- Never rely on a model for current prices, offers, deadlines or regulations; paste them in or keep them in a knowledge file.
- Date your reference documents so you know when a Custom GPT's information was last updated.
- If you need recent public information, ask the tool to search and show sources, then check them.
Related terms
Where you use it: free lessons
- Build a Custom GPT Ad Copy Writer for Search and Social Ads (ChatGPT, 8 min)
- Prompt Writing 101: Role, Context, Task, Format (ChatGPT, 8 min)
Frequently asked questions
- How do I find a model's knowledge cutoff?
- Providers usually list it in the model's documentation. You can ask the model too, but treat its answer as a hint rather than a guarantee.
- Does a knowledge cutoff matter if I give the AI my own facts?
- Much less. When your prompt or knowledge file holds the current details, the model works from those. The cutoff mainly bites when you ask about things you did not supply.