Training data: Training data is the collection of examples an AI model learns from, such as web pages, books, code, images, or labeled records. What is in that data, and what is missing, shapes everything the model knows, how it writes, and which mistakes it makes.
A model has no knowledge except what it extracted from its training data. For large language models that means an enormous slice of the public internet, digitized books, code repositories, and licensed datasets, followed by smaller curated sets of example conversations used to teach helpful behavior.
Three consequences follow. The model's knowledge stops at the date the data was collected. Topics that are common online are handled well, while niche or specialized subjects are thinner. And whatever biases, errors, and stereotypes exist in the source text can show up in the output, which is why AI bias is a practical concern rather than a theoretical one.
Your own data can become training data. Free consumer AI tools may use your conversations to improve their models unless you opt out; business and API tiers typically do not by default. Before pasting anything sensitive, find out which applies to the tool and account you are using, and follow your organization's policy.
Example at work
A recruiting team evaluates an AI screening tool and asks the vendor what data trained it. Learning it was trained largely on the vendor's past placement records, the team asks how those records were checked for bias against older candidates and non-traditional backgrounds before agreeing to a pilot.
Why it matters
Every 'why did the AI say that?' question ultimately traces back to training data. Knowing this helps you predict where a model will be strong, where it will be shaky, and what questions to ask a vendor about data sources, bias testing, and whether your inputs feed future models.
Related terms
- Pre-trainingPre-training is the first and largest phase of building an AI model, in which it learns general patterns from a massive dataset, such as predicting the next word across trillions of words of text. Later phases then shape it into a useful assistant.
- AI biasAI bias is a systematic tendency for an AI system to produce outputs that are unfair or skewed toward certain groups, usually because the data it learned from reflected historical patterns, gaps, or human prejudice.
- Knowledge cutoffA knowledge cutoff is the date after which a language model's training data stops. Unless it is given live information through search or documents, the model knows nothing about events, prices, or rules that changed after that date.
- Data privacyData privacy, in the context of AI, is the set of rules and practices that govern what information you put into an AI system, who can see it, how long the vendor keeps it, and whether it is used to train future models.
- Fine-tuningFine-tuning is extra training that adjusts an existing AI model's weights using a smaller, targeted set of examples so it performs better on a specific task, style, or domain. It changes the model itself, unlike prompting, which only changes the instructions.