AI Transformation · Glossary

Fine-Tuning

Updated October 25, 2026

Fine-tuning is further training of an existing artificial intelligence (AI) model on a curated set of examples, so that a behavior is built into the model itself rather than supplied in every prompt. Model vendors use the same kind of training stage to shape their own assistants. Andrej Karpathy, a founding member of OpenAI, explaining that vendor training stage in a public talk in early 2025, said you cannot write rules into the model, so the behavior you want is installed by showing it examples of that behavior, including when to decline. In his account of how such models are generally built, not any one vendor's current practice, that is also why two models trained on similar text behave differently: tone, hedging and refusal boundaries come from each vendor's training examples, editorial policy as much as capability.

Fine-tuning changes behavior more than knowledge. What a model learned in training is fixed when it ships and expensive to update, so it is a poor place to keep a company's facts. For those, retrieval-augmented generation (RAG), which hands the model relevant passages from the company's own documents at the moment of the question, together with better instructions and tool access, usually gets there first and cheaper. The case for fine-tuning is a model whose default behavior does not fit the work.

When is fine-tuning worth the cost?

When the problem is how the model behaves, and better instructions have not fixed it. The cost is real in three ways:

1. It is more involved than working in the prompt: data preparation, model selection and a training job.
2. The lever it offers is more examples of the wanted behavior, which means producing a curated dataset rather than writing a specification.
3. Fine-tuning at any level risks degrading what the model already does well, because the new training overwrites some of what it learned. Prefer a model that already performs on the use case.

Fine-tuning is not a cure for AI hallucination either. Training a model to answer "I don't know" at the edge of its knowledge lowers the risk, and Karpathy says models still hallucinate despite it. For an AI agent, a model using tools in a loop, Barry Zhang of Anthropic, a company that sells the computing power agents run on, said in a conference talk that an agent's two real design decisions are which tools to offer and what the prompt says. Those two are usually where to look before retraining the model.

In AI transformation planning, the first question is whether the gap is knowledge or behavior. A knowledge gap usually points to retrieval; a behavior gap is where fine-tuning can earn its cost.