Training Data
Data used to develop and train AI models to perform their intended functions.
How the laws define this
Regulators and standards bodies consistently frame training data as the underlying datasets and content inputs used to train, develop, fine-tune, or optimize artificial intelligence and machine learning models. In these framework definitions, inputs shape model behavior and include sources ranging from open web data to personal and synthetic records. The most critical divergence involves whether the scope extends to validation inputs. ISO/IEC 22989:2022 strictly separates training data from validation or test datasets based on its primary role in model learning. Conversely, the Draft Bill on Artificial Intelligence explicitly includes validation datasets within its definition alongside training and fine-tuning data. Furthermore, instruments like California AB 2013 and the Generative AI Ethics Guidebook limit their focus specifically to generative AI systems, whereas ISO/IEC 22989:2022 applies to general artificial intelligence models.
Synthesised from the 9 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first. Showing 8 of the 9 instruments that define this term; all are listed under Definitions below.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| China | Provisions on the Administration of Deep Synthesis of Internet-based Information Services | In Force | Limits to data for deep synthesis models, templates or algorithms, explicitly including personal or biometric information like faces and voices. |
| California, United States | California AB 2013 — Generative Artificial Intelligence: Training Data Transparency | In Force | Adds mandatory public disclosure of detailed dataset information including sources, acquisition methods, processing steps, volumes, temporal scope and synthetic data use. |
| South Korea | Generative AI Ethics Guidebook (NIA / KCC – sectoral guide for generative AI) | In Force | Emphasizes documenting, minimizing and managing training data based on its provenance, licensing and content to address legal, ethical and technical risks. |
| Ireland | Data Protection Commission Guidance on AI and Large Language Models | In Force | Treats training data as potentially containing personal data, stressing provenance, retention, minimisation and lawful basis under data protection law. |
| ISO | ISO/IEC 22989:2022 - Information technology — Artificial intelligence — Concepts and terminology | In Force | Distinguishes training data by its role in parameter learning during development, separate from validation or test datasets, and notes it may be labelled or unlabelled. |
| Australia | Harnessing data and digital technology — Inquiry report (Productivity Commission, No. 111) | Under Review | Specifies inclusion of open web material and licensed content, framing it for licensing, transparency and copyright considerations. |
| Denmark | Strategic Approach for Artificial Intelligence | Adopted | Specifies examples like Danish text corpora and describes clarifying licensing, rights, privacy and provenance, signalling open-release under governed terms. |
| Indonesia | Draft Bill on Artificial Intelligence | Draft | Includes annotations and preprocessing steps, requiring data governance, quality and personal data protection to mitigate bias and privacy risks. |
What it triggers
- In California, developers of generative AI systems must publicly disclose detailed information about training datasets including sources and processing steps.source
- Under China's Provisions on the Administration of Deep Synthesis, training data containing personal or biometric information is subject to data-protection, consent and security requirements.source
- ISO/IEC 22989:2022 defines training data as distinct from validation or test datasets by its primary role in model learning.source
Definitions (9)
Material used to train machine learning models, including open web material and licensed content; encompasses the datasets and content inputs used to develop AI systems' behavior. The report uses this term to frame licensing, transparency and copyright considerations for AI development.
Datasets used to train, retrain, or fine-tune generative AI systems, including data used for reinforcement learning from human feedback (RLHF) and other methodologies; the law requires disclosure of sources, acquisition methods, processing steps, volumes, temporal scope, and whether synthetic data was used.
Described as the datasets used to train AI and generative models (e.g., Danish text corpora); the strategy requires clarifying licensing, rights, privacy and provenance of training data and signals open-release of selected corpora under governed terms to support domestic model development.
All datasets and records used to develop, pre‑train, fine‑tune or update an LLM, including public, proprietary, scraped and third‑party sources; the guidance treats training data as potentially containing personal data and emphasises provenance, retention, minimisation and lawful‑basis considerations.
The dataset or collection of data used to train, fit, or optimize an AI model's parameters during the development phase; it supports the learning process and is distinct from validation or test datasets by its primary role in model learning. Training data may be labelled or unlabelled depending on the learning paradigm.
Datasets and data sources used to train, validate, or fine-tune AI models, including annotations and preprocessing steps, which must meet data governance, quality, and personal data protection requirements to mitigate bias and privacy risks.
Collections of data used to train generative AI models, whose provenance, licensing, and content determine legal, ethical, and technical risks and therefore should be documented, minimized, and managed for compliance and quality.
Collections of data used to train or tune models, templates or algorithms for deep synthesis, including datasets containing personal or biometric information (e.g., faces, voices), which are subject to data-protection, consent and security requirements under the Provisions.
Data sets used to train or fine-tune machine learning models, encompassing sources, provenance, and characteristics of the records used to develop model behaviour; treated as a lifecycle stage that may engage privacy obligations (lawful collection, consent, quality, bias).
Not to be confused with
General-Purpose AI refers to models trained on broad data for diverse tasks, while training data is the specific datasets used in that training process.
Technical documentation records compliance details about AI systems, whereas training data consists of the actual datasets used to develop those systems.
Questions and answers
What must be disclosed about training data under California law?
California AB 2013 requires disclosure of sources, acquisition methods, processing steps, volumes, temporal scope, and whether synthetic data was used for generative AI training datasets. This applies to developers effective January 1, 2026.
Does training data include personal information and what are the rules?
Several definitions treat training data as potentially containing personal or biometric data, subjecting it to consent, protection and lawful basis requirements. China's Provisions explicitly include faces and voices with security mandates.
How does ISO distinguish training data from other datasets?
ISO/IEC 22989:2022 defines training data as the dataset used to train, fit or optimize model parameters during development, distinct from validation or test datasets by its learning role. It may be labelled or unlabelled.
Related Terms
General-Purpose AI (GPAI)
AI models trained on broad data capable of performing diverse tasks across domains....
Bias Mitigation
Measures to identify and reduce discriminatory outcomes in AI systems....
Technical Documentation
Detailed records demonstrating AI system compliance with regulatory requirements....