AI Vocabulary

Fine-tuning or prompting: what does a specialised AI really cost?

Fine-tuning means retraining a model on your own data to change its behaviour durably. Prompting means giving it instructions at the time of each request. The first is expensive, takes weeks and ages; the second is immediate, free and can be redone at will.

What fine-tuning does

You start from an existing model and continue its training on a chosen corpus — your instruments, your notes, your wording. The model comes out with different habits: it adopts your style, your vocabulary, your structures more naturally.

That presupposes a significant volume of data, careful preparation of the corpus, computing time, and an evaluation to confirm the result is better and not merely different.

What a prompt does

You give the model, at each request, the useful instructions and examples. Nothing is trained: the model is the same, you steer it in the moment.

The change is immediate, reversible, and corrected in 10 seconds when it does not suit. Professional systems do this in the background, without the user having to attend to it.

What the first does not do, and is credited with

Fine-tuning teaches a way of doing, not knowledge. That is the most widespread and most expensive confusion: retraining a model on your contracts does not teach it what your contracts contain, it teaches it to write contracts resembling yours.

If you want a system to know what is in your matter, you must give it to the system when you interrogate it. No retraining substitutes for that, and a model fine-tuned on your data will invent your clauses in your style, which is more dangerous than inventing them in a neutral one.

When it is nonetheless justified

Two situations make it relevant, and they are rarer than claimed.

The first is a strongly marked and stable requirement of form: a firm whose instruments have followed a particular structure for years, and which wants that structure back without having to ask for it every time.

The second is niche vocabulary that general models handle poorly — certain technical fields, certain specialised disputes where the model's usual terminology is not the right one.

In both cases this is a marginal improvement on a foundation that already works, never a way of making work what does not.

What it does not solve

Fine-tuning is dated by construction. The tuned model corresponds to a state of your practice; six months on, your positions have moved and it has not. Maintaining it means starting again periodically.

It also creates dependency: a fine-tuned model is inseparable from the provider hosting it, and does not travel.

And it raises a confidentiality question that must be asked explicitly: what has served for training is not easily removed from a model.

Why it matters to a lawyer

Because fine-tuning is often proposed as the answer to it does not know how we work, when it addresses only a small part of that need — the form, not the substance.

Before committing to a project of this kind, the question is simple: does what we are looking for concern style or knowledge of our matters? In the second case, fine-tuning is not the tool.