Skip to content
Adrythm
AI and automation

Training data

fine-tuning / pre-training / will you train on our data / model training

In short

Training data is what a model learned from. The question owners ask is whether their own data joins it, and there is no single answer: OWASP names three separate stages, pre-training, fine-tuning and embedding. A commitment worth having names the stage and comes in writing.

The question is always the same and the answer depends on a distinction most conversations skip. OWASP names three: data poisoning can target different stages of the LLM lifecycle, including pre-training, which is learning from general data, fine-tuning, which is adapting models to specific tasks, and embedding, which is converting text into numerical vectors. Your information could touch any one of those and not the others, so a yes or no answer is not an answer.

The reason the distinction is a security matter, not only a privacy one, is in the same document. Data poisoning occurs when pre-training, fine-tuning, or embedding data is manipulated to introduce vulnerabilities, backdoors, or biases. Whoever controls what goes in controls part of what comes out, which is why unverified training data increases the risk of biased or erroneous outputs.

The failure that catches ordinary businesses is not an attack at all. OWASP lists it plainly: users unknowingly inject sensitive or proprietary information during interactions, which could be exposed in subsequent outputs. Nobody set out to hand over a price list or a client name. Somebody pasted it into a prompt because it was the fastest way to get an answer.

What a real commitment looks like is worth seeing once. One vendor writes that your prompts, completions, embeddings and training data are not available to other customers, and are not used to train any generative AI foundation models without your permission or instruction. It goes further, stating that customer data, prompts and completions are not used to improve its own or third-party products or services without explicit permission. That is what a specific answer reads like, and it is a useful shape to hold other answers against.

In practice

The practical control is not a clause, it is a habit. Whatever your agreement says, staff decide what goes into the box, and the box does not look like a filing cabinet. Deciding what is never pasted in, and saying so once in plain words, does more than a paragraph nobody reads.

Not the same as

Retrieval
A system that reads your documents to answer a question is not learning from them. That is a different arrangement with different risks, and the two get described with the same sentence in sales conversations.
Logging
Keeping a record of what was asked is ordinary and often required. Whether those records then feed a model is a separate decision that should be written down separately.

Why it matters to you

This is the question every business asks about AI and the one most often answered in a way that sounds definitive and commits to nothing. Knowing that there are three stages turns a vague reassurance into something checkable, and the vendors who have thought about it will answer stage by stage without being prompted.

What to ask or check

  1. 01Which of the three stages, if any, could our information reach?
  2. 02Is that commitment in the contract, or only in a sales conversation?
  3. 03What have we told staff they must never paste into one of these tools?

What people get wrong

That training is one thing a vendor either does or does not do. OWASP separates pre-training, fine-tuning and embedding, and a commitment that does not say which stage it covers has not answered the question.

Red flags

  • A reassurance about training that appears nowhere in the agreement.
  • An answer that does not distinguish the stages when asked directly.
  • No guidance to staff about what may be pasted into an AI tool.

Who owns it

The vendor owns the commitment and you own the habit. Only one of those is enforceable, and the other is the one that leaks.

Where you will see it

In a vendor's security page, in the paragraph of a contract nobody negotiated, and in whatever somebody pasted into a chat box this morning.

Retrieval augmented generation

Retrieval augmented generation is how an AI answers from your documents without being trained on them. OWASP describes it as combining a pre-trained model with external knowledge sources at answer time. So the documents sit in a store the system reads from, and who can read that store is the question.

Privacy-enhanced AI

Privacy-enhanced AI means an AI system built so your data is not exposed or quietly reused. NIST lists it as one of seven characteristics of trustworthy AI, and states the honest catch: privacy techniques can cost accuracy. The FTC has required firms to delete models built on unlawfully obtained data.

Model drift

Model drift is an AI system getting worse without anyone changing it. Microsoft names the causes plainly: data distribution changes, training-serving skew, data quality problems, shifts in environments and consumer behavior changes can all make a model stale. NIST sets the bar over the entire lifetime of the system, not at launch.

AI agent

An AI agent is a model that has been granted the ability to take actions, not just produce text. OWASP says the damage one can do comes from three grants: excessive functionality, excessive permissions and excessive autonomy. What it is allowed to do matters more than how good it is.

Hallucination

A hallucination is AI output that sounds right and is not. OWASP describes the model filling gaps in its training data using statistical patterns, without understanding the content, so the answer can be fluent and unfounded at once. Its own first example is an airline that was successfully sued over its chatbot.

Prompt injection

Prompt injection is when text an AI system reads becomes an instruction it follows. OWASP notes the text does not have to be visible to a person, only parsed by the model, and that no fool-proof prevention is known. The UK's national cyber security body says the same.

Want this explained against your own numbers?

Twenty minutes, a straight answer, and no follow-up sequence if you decide not to work with us.