How does an AI model get its knowledge?
An AI language model learns in two steps. In pre-training, it reads very large amounts of text and learns to predict the next word; this is where language, general knowledge and patterns of reasoning come from. In fine-tuning, it learns from examples and human ratings to answer questions and follow instructions. After that, its knowledge is frozen. It only gets current information and knowledge from your organisation if it is given to it with the question.
What does a model learn in pre-training?
In pre-training, the model receives very large amounts of text: books, specialist articles, public websites, program code. Its only task is to predict the next word. When it gets it wrong, its billions of internal weights are adjusted a tiny bit. This is repeated an unimaginable number of times.
To predict well, the model has to learn a great deal along the way: grammar, facts, specialist language, how arguments are built, how to calculate and translate. That is the “intelligence” you experience later. It is not stored in the model as a database, but as patterns in its weights.
How does it learn to answer questions?
After pre-training, a model can continue text, but cannot yet answer well. In fine-tuning, the maker shows it examples of good answers, and people rate which of several answers is more helpful, more correct and safer. From this, the model learns to follow instructions, ask back and decline harmful answers.
The maker of the model does both, pre-training and fine-tuning, once and with great computing effort. After that the model is finished and no longer changes.
Why does the model not know everything?
- Knowledge cut-off: the model does not know what happened after its training.
- No knowledge of your organisation: your contracts, guidelines and processes were never part of the training.
- No looking things up: the model generates text that is likely to fit. If it lacks knowledge, an answer can sound convincing and still be wrong.
How do current and in-house knowledge get into the answer?
You give it to the model with the question. The platform searches your documents for the relevant passages and presents them to the model together with the question. The model answers on that basis and names the sources, so you can check. For current information from the internet there is web search. The model itself learns nothing in the process: for the next question without these documents, it again only knows what it learned in training.
Who has to disclose what a model was trained on?
The EU AI Act obliges providers of general-purpose AI models, among other things, to publish a summary of the training content and to document the model technically (Article 53). Anyone who merely uses such a model is not its provider.
How does this work with basebox?
basebox does not develop or train its own models. We run open-source models ourselves, in our cloud in Germany or on your server, and choose the best models that suit the hardware. Which models these are is listed in the model register. A new model arrives with an update.
Your data trains no model. In-house knowledge is added to every question from the documents in your apps, with sources; current information comes from web search if your organisation enables it. What the AI model in basebox does
Frequently asked questions
Where does the AI model in basebox get its knowledge?
From its training by the model maker. basebox does not develop or train its own models; we run open-source models ourselves, in our cloud in Germany or on your server. Your organisation's knowledge is added to every question from the documents in your apps, with sources.
Does the model learn from our questions and documents?
No. Your inputs and documents do not change the model, and basebox trains no model with your data. The model reads your documents only for the answer at hand.
Why does the model not know about current events?
Its knowledge ends with the state of its training. For current information there is web search, which your organisation can enable; the model then fetches public information from the internet.
Why does a model sometimes make things up?
Because it generates text that is likely to fit and does not look things up in a database. If it lacks knowledge, an answer can sound plausible and still be wrong. Answers with sources from your documents can be checked.
Sources
- Brown et al. (2020): Language Models are Few-Shot Learners, arXiv ↗
- Ouyang et al. (2022): Training language models to follow instructions with human feedback, arXiv ↗
- Regulation (EU) 2024/1689 (AI Act), Article 53: obligations for providers of general-purpose AI models, EUR-Lex ↗
- basebox: model register ↗
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 16 November 2026.