What is AI training data?

How it works

During training, a model’s internal settings (its parameters) are fitted to the data it is shown. The EU AI Act defines training data as “data used for training an AI system through fitting its learnable parameters” (AI Act, Article 3). ISO/IEC 22989:2022, as quoted by a secondary source, uses similar wording: “data used to train a machine learning model” (itSMF).

The EU AI Act and ISO/IEC 22989 separate three kinds of data:

• Training data: what the model learns from.
• Validation data: used to evaluate the trained system and tune settings it does not learn on its own, among other things to prevent underfitting or overfitting (missing patterns, or memorising training data and failing on new data) (AI Act).
• Testing data: used to independently confirm the system performs as expected before release (AI Act).

The AI Act also separates all three from input data, which is what a finished system receives and responds to. Pasting a document into a chatbot is input; using that document to adjust the model is training. Updating a trained model with different training data is what ISO calls retraining (itSMF).

In security glossaries

NIST’s security glossary index lists no plain entry for training data, but lists attack-related terms such as “training data extraction”, and notes its entries are not official definitions (NIST). That term means an attacker prompting a generative model with specific inputs to pull out its training data (NIST).

Why it matters to sellers

If you license data for AI training, how you obtained it and what your contracts and privacy policy promised customers and employees can matter later. In January 2021 the FTC announced a proposed settlement with photo app developer Everalbum requiring express consumer consent before using facial recognition on their photos and videos, and deletion of models and algorithms built from users’ uploaded photos and videos (FTC). As general information, points to review with counsel include:

• Promises to customers. FTC staff said it may be unfair or deceptive to start using consumers’ data for AI training and disclose it only through a hidden, retroactive change to terms or a privacy policy (FTC).
• Personal data. The European Data Protection Board (EDPB), in Opinion 28/2024, says AI models trained on personal data “cannot, in all cases, be considered anonymous”, and calls for case-by-case assessment (AI for Science).
• Disclosure. The European Commission’s AI Office published a template on 24 July 2025 for the public summary of training content of general-purpose AI models (WilmerHale). Providers of general-purpose AI models first placed on the EU market from 2 August 2025 must publish a completed summary straight away; providers of earlier models have until 2 August 2027 (Paul, Weiss). The template’s source categories reportedly include “private/licensed datasets” (regulations.ai).
• Licensing. The US Copyright Office’s Part 3 report, released in pre-publication form on 9 May 2025 (McDermott), notes licensing markets for training data are developing, and that where licences exist or are likely feasible, unlicensed use weighs against fair use (a US copyright doctrine allowing some unlicensed use of protected works) (Mishcon de Reya).

On SellMyData, each dataset goes through a rights check and de-identification before listing; consumer personal data, health records tied to people and payment card data are not accepted. Buyers see only a redacted sample, and nothing is licensed without the seller’s sign-off on buyer, terms and price. You can request an estimate or book a call; neither commits you to anything.