Fine-tuning means giving an existing AI model extra training on your own examples, so it gets better at one particular job or style. You do not build a new AI; you take a finished one and adjust it, which takes far more effort than writing a better prompt.
Think of hiring an experienced chef and then training them for a few weeks in your restaurant’s dishes. They already know how to cook. The training teaches them your way: your portions, your plating, your house style. That is fine-tuning.
In a sentence: “We fine-tuned a model on our best past replies, so now it answers customers in our house style without a long prompt.”
What the word means
The plain point: the model already works; fine-tuning nudges it towards your examples.
OpenAI describes it in those terms: take a base model, show it the inputs and outputs you expect, and get a model that is better at your task.1
A base model is the general-purpose model before your changes. “Fine” is the key word: small adjustments on top, not training from scratch.
A simple example
The plain point: fine-tuning suits a fixed format you need thousands of times.
A support team wants every reply in the same structure, tone and length. They collect a few hundred of their best past replies, paired with the customer messages that prompted them, and fine-tune a model on those pairs. Afterwards, the model writes in that shape without a long set of instructions each time.
OpenAI lists jobs like this, such as content in a specific format, among the things fine-tuning helps with.1
Fine-tuning teaches a way of answering. If the model is missing facts, give it the facts instead.
Why it matters to you
The plain point: most people never need it. Try the simpler options first.
- A better prompt is often enough. OpenAI’s own guide says prompt engineering “may be all you need in order to get great results for your use case.”1 See prompt engineering meaning.
- Missing information is a different problem. If the AI does not know your products or policies, hand it the documents. That is RAG or simply pasting them in.
- It can expire. A fine-tune is tied to the model it was built on. When that model is retired, the work may have to be redone. When AI models retire shows how much warning companies give.
Often confused with
- Training a model: building one from scratch on a huge amount of data. That first stage is called pretraining, and it is done by the companies that make the models. Fine-tuning comes after.
- RAG: looks things up in your documents at question time and leaves the model unchanged. See RAG meaning.
- LoRA: a lighter kind of fine-tuning, common with image models, that trains a small add-on instead of the whole model. Covered in the prompting terms glossary.
Questions people ask
What is the difference between fine-tuning and RAG?
Fine-tuning changes the model; RAG leaves the model alone and hands it documents to read. IBM puts it this way: “RAG augments a natural language processing (NLP) model by connecting it to an organization’s proprietary database, while fine-tuning optimizes deep learning models for domain-specific tasks.”2 So for up-to-date or private facts, RAG is usually the better fit. See RAG meaning.
Is fine-tuning better than a good prompt?
Not usually as a first step. OpenAI’s guide says “The prompt engineering process may be all you need in order to get great results for your use case,” and presents fine-tuning as something used together with prompting “to realize a few more benefits over prompting alone”, such as showing the model more examples than fit in one message.3
Learn more on this site
- The full explainer: Prompt, context, skill, or fine-tune explains when each one is worth it.
- Giving the AI material instead: Context engineering.
- How models are made: How an AI is actually built.
- The basics: the AI glossary and AI explained for complete beginners.
- Related: Foundation model meaning.
- Related: What Is a Checkpoint in AI? Checkpoint Meaning in Plain English.
- Related: Fine-Tuning vs RAG vs Prompting.
- Related: What Is Pretraining in AI? Pretraining Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is fine-tuning?
- Building a brand-new AI from scratch
- Writing a longer, clearer prompt
- Giving an existing model extra training on your own examples
Show the answer
Answer: Giving an existing model extra training on your own examples. Fine-tuning takes a finished model and adjusts it with extra training on your own examples, so it gets better at one job or style.
2. If an AI is missing facts about your products, what does the page suggest?
- Give it the documents, for example with RAG or by pasting them in
- Fine-tune it until it learns the facts
- Ask the same question again and again
Show the answer
Answer: Give it the documents, for example with RAG or by pasting them in. Fine-tuning teaches a way of answering; if the model is missing facts, the page says to hand it the facts instead.
3. What does the page say most people should try before fine-tuning?
- Building their own model
- A better prompt
- A bigger screen
Show the answer
Answer: A better prompt. Most people never need fine-tuning: OpenAI’s own guide says a better prompt is often enough.
Related terms: Constitutional AI · Distillation · Emergent abilities · Foundation model · Frontier model · Inference
- 1OpenAI, Model optimization, undated guide, read at source 1 Oct 2026“Fine-tuning lets you take an OpenAI base model, provide the kinds of inputs and outputs you expect in your application, and get a model that excels in the tasks you’ll use it for.” ↑a“Generating content in a specific format”. ↑b“The prompt engineering process may be all you need in order to get great results for your use case.” ↑c
- 2IBM, RAG vs. fine-tuning, undated, read at source 12 Oct 2026“RAG augments a natural language processing (NLP) model by connecting it to an organization’s proprietary database” … “while fine-tuning optimizes deep learning models for domain-specific tasks” and “RAG plugs an LLM into stores of current, private data that would otherwise be inaccessible to it.” ↑
- 3OpenAI, Model optimization, undated guide, read at source 12 Oct 2026“The prompt engineering process may be all you need in order to get great results for your use case.” and “You can use fine-tuning with prompt engineering to realize a few more benefits over prompting alone”, including to “provide more example inputs and outputs than could fit within the context window of a single request.” ↑
The comparison on this page (training an experienced chef in your house style) is this site’s own way of explaining, not a claim from this source.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK