What is fine-tuning in AI?

What fine-tuning means

An AI model such as the one behind ChatGPT or Claude is trained once, on a vast amount of text. That gives it general ability: it can write, summarize and answer questions on almost anything. Fine-tuning is a second, much smaller round of training that starts from that finished model and pushes it toward one specific behavior, using examples you supply.

Each example is a pair: what goes in, and what a good answer looks like. Show the model enough of them and it starts producing that kind of answer on its own, without being told how every time.

What “fine-tune” means as a verb. To fine-tune a model is to run that second round of training on it. You will also hear people say they are “fine-tuning a prompt” or “fine-tuning the output.” That is the everyday sense of making small adjustments, and it has nothing to do with changing the model. If someone in a meeting says “we should fine-tune it,” ask which they mean. The two cost wildly different amounts of effort.

Fine tuning in one example

“Fine tuning” and “fine-tuning” mean the same thing. Say a support team wants every reply in the same short format with the same sign-off. They collect a few hundred past replies they are happy with, each paired with the customer message it answered, and train the model on those pairs. Afterwards the model writes in that format without being reminded. If the same team wanted the model to know this month’s return policy, fine-tuning would be the wrong tool: they would give it the policy document instead (see RAG below).

The short version: fine-tuning changes how a model writes. It does not reliably teach it new facts. A controlled study by Gekhman et al. found that large language models struggle to acquire new factual knowledge through fine-tuning.

How it works, in plain English

Think of a new hire who already has a strong education. You do not send them back to school. You give them 500 sample cases from your team, each with the answer your best person gave, and they absorb how you do things.

Technically, the model has billions of internal numbers (its “weights”) that were set during the original training. Fine-tuning nudges a small portion of those numbers using your examples, so the model gets better at your pattern. Three practical consequences follow:

• It needs examples, not documents. A folder of PDFs is not training data. Training data is pairs of “input” and “ideal output,” usually saved in a structured file.
• The change is baked in. The model behaves differently everywhere you use that version of it, with no instructions needed. That is the benefit, and also the risk: if your needs change, you retrain.
• It is run by developers. Even where it is offered, it is done through a developer platform, not a button in the chat app.

Fine-tuning vs prompting vs RAG vs custom GPTs

These four are the options people mix up. They solve different problems.

	What you do	Best for	What it costs you	Changes the model?
Prompting	Write clear instructions and a few examples in the chat	Almost any task, tone and format	Minutes	No
RAG (give it your documents)	The AI looks up your files before answering	Answers based on your policies, contracts, knowledge base	Setup of a tool, clean documents	No
Custom GPTs and Projects	Save instructions and reference files as a reusable assistant	Repeating the same task across a team	An hour or two to set up	No
Fine-tuning	Train the model on hundreds of examples	A narrow, high-volume task needing consistent style or labels	Examples, technical work, ongoing upkeep	Yes

Notice that three of the four leave the model alone. That is why they are the place to start: they are fast to try, easy to undo and easy to update. See RAG, custom GPTs and prompt engineering for each, or the side-by-side fine-tuning vs RAG vs prompting page.

Where fine-tuning makes sense, and where it doesn’t

It can make sense when:

• A company labels thousands of support tickets a day into fifteen categories and has years of tickets already labeled by staff. A fine-tuned model can copy that labeling at low cost per ticket.
• A team must produce hundreds of documents a week in one rigid format, such as a standard claim summary, and a prompt still drifts off format.
• A product needs a very specific voice or output structure, in volume. OpenAI’s guide notes that you can train a smaller, cheaper, faster model to excel at a particular task where a larger model is not cost-effective.

It usually does not make sense when:

• You want the AI to “know our company.” Give it the documents instead; they stay current and can be cited.
• You write a weekly report, a handful of emails or a few proposals. A good prompt with two sample examples does this.
• You have fewer than the 50 well-crafted demonstrations OpenAI suggests starting with. A small sample is hard to evaluate.
• Your information changes monthly. A fine-tuned model has to be retrained to update. A document can be replaced in a minute.
• You want to keep company data private. Fine-tuning requires uploading your examples to a provider, so the same data-handling questions apply. See is it safe to put company data in ChatGPT.

A useful test: if you can describe what you want in a paragraph and show it two examples, you do not need fine-tuning.

What it takes

• Examples: the supervised fine-tuning guide from OpenAI says the minimum number of examples you can provide for fine-tuning is 10. OpenAI recommends starting with 50 well-crafted demonstrations. In the GPT-3 paper, Brown et al. say the pre-train-then-fine-tune method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. Each pair should be checked by a person who knows what “good” is. Preparing them is typically the slowest and most expensive step.
• Technical work. Formatting the data, running a training job through a developer platform, and testing the result against examples the model has not seen.
• Time. Gathering and cleaning examples takes days to weeks. The training job itself is often minutes to hours.
• Money: pricing differs by provider and model and changes, so get a current quote before planning. For small jobs the platform charge is often modest compared with the staff time spent preparing data.
• Upkeep. When the provider retires the base model, or your needs change, you may have to retrain.

Can you fine-tune ChatGPT, Claude or Gemini today?

As of November 2026, and with the caveat that this changes often:

• ChatGPT and OpenAI. You cannot fine-tune inside ChatGPT. OpenAI’s developer platform did offer fine-tuning of models in the GPT-4.1 family, but OpenAI is winding the platform down. On May 7, 2026, creating fine-tuning jobs became unavailable to organizations that had not previously run fine-tuning, and on July 2, 2026 it also ended for organizations that had not run inference on a fine-tuned model in the past 60 days. Now eligible existing customers can create training jobs until January 6, 2027, unless the base model shuts down earlier. Fine-tuned models keep working until their base model shuts down: gpt-4.1-nano and its fine-tuned versions shut down on October 23, 2026, and gpt-4.1, gpt-4.1-mini and their fine-tuned versions shut down on January 14, 2027.
• Claude and Anthropic. There is no fine-tuning in the Claude apps. AWS says that on Amazon Bedrock you can train a foundation model to improve performance on specific tasks. The only Anthropic model in the AWS fine-tuning table is Claude 3 Haiku, listed for us-west-2 only. Anthropic’s post says fine-tuning Claude 3 Haiku in Amazon Bedrock is generally available, in an update dated November 1, 2024.
• Gemini and Google. Google documents supervised fine-tuning for Gemini models on its Vertex AI developer platform. The Google page lists Gemini 3.5, 3.1 and 2.5 models, including 2.5 Pro, as supporting supervised tuning. The Google page has separate tuning guides for text, image, document, audio and video, and for function calling. The list changes, so check it before planning.
• Open models. Some open-weight models can be fine-tuned too. AWS’s table lists Meta Llama 3.1 70B Instruct among the Bedrock models you can fine-tune, which suits teams with engineers.

The headline for managers: OpenAI is closing its fine-tuning platform, and the Claude route is one older model on AWS. For most office work the practical path is good prompts plus your own documents.

Common misconceptions

“We need to fine-tune so the AI knows our business.” Usually false. Fine-tuning teaches style and patterns, and it is a poor and expensive way to store facts. Give the AI your documents when you ask.

“Fine-tuning removes hallucinations.” It does not. A fine-tuned model can still invent things, sometimes in your preferred format and tone, which makes the errors look more convincing. The same 2024 study says its results highlight the risk in introducing new factual knowledge through fine-tuning. See hallucination.

“Fine-tuning is the same as training a model.” Training from scratch is a project for a handful of large labs. Fine-tuning is a small adjustment to an existing model.

“It is only a technical decision.” The hard part is deciding what “good” looks like, collecting examples and agreeing on how to measure them. That is a business task before it is an engineering one.

“Our chats with ChatGPT are fine-tuning it for us.” No. Using a chat assistant does not retrain your own private version of it. Whether your conversations are used to improve a provider’s models is a separate settings question.

What to do instead

Work up this ladder and stop when the problem is solved:

1. Rewrite your prompt with a clear role, the steps and two sample outputs. See how to write AI prompts that work.
2. Give the AI the documents it needs when it answers.
3. Save the setup as a reusable assistant for your team, such as a custom GPT or a Claude Project.
4. Only then, with an engineer and a clear success measure, consider fine-tuning.

Most office workflows never need step 4. If you would rather practice steps 1 to 3 on email, spreadsheet and document exercises, that is the focus of Learn AI Fast, a two-day in-person class in Boston for non-technical professionals.

Sources

• arXiv, Gekhman et al., Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?: “large language models struggle to acquire new factual knowledge through fine-tuning”
• OpenAI, Model optimization guide: “You can train a smaller, cheaper, faster model to excel at a particular task”
• OpenAI, Supervised fine-tuning guide: “We recommend starting with 50 well-crafted demonstrations”
• OpenAI, Supervised fine-tuning guide: “The minimum number of examples you can provide for fine-tuning is 10”
• arXiv, Brown et al., Language Models are Few-Shot Learners: “this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples”
• OpenAI, Deprecations, update to self-serve fine-tuning: “Creating fine-tuning jobs or training is not available to organizations that have not previously run fine-tuning”
• OpenAI, Deprecations, update to self-serve fine-tuning: “Creating fine-tuning jobs is no longer available to organizations that have not run inference on a fine-tuned model in the past 60 days”
• OpenAI, Supervised fine-tuning guide: “Eligible existing customers can create training jobs until January 6, 2027, unless the base model shuts down earlier”
• OpenAI, Supervised fine-tuning guide: “and its fine-tuned versions shut down on October 23, 2026”
• OpenAI, Supervised fine-tuning guide: “and their fine-tuned versions shut down on January 14, 2027”
• AWS, Amazon Bedrock User Guide, Customize a model with fine-tuning: “you can train a foundation model to improve performance on specific tasks”
• AWS, Amazon Bedrock User Guide, Customize a model with fine-tuning: “Anthropic Claude 3 Haiku anthropic.claude-3-haiku-20240307-v1:0:200k us-west-2”
• Anthropic, Fine-tune Claude 3 Haiku in Amazon Bedrock: “Fine-tuning Claude 3 Haiku in Amazon Bedrock is generally available”
• Google Cloud, Tune Gemini models with supervised fine-tuning: “The following Gemini models support supervised tuning”
• Google Cloud, Tune Gemini models with supervised fine-tuning: “Text tuning Image tuning Document tuning Audio tuning Video tuning Tune function calling”
• AWS, Amazon Bedrock User Guide, Customize a model with fine-tuning: “Meta Llama 3.1 70B Instruct meta.llama3-1-70b-instruct-v1:0:128k us-west-2”
• arXiv, Gekhman et al., Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?: “highlight the risk in introducing new factual knowledge through fine-tuning”