Large Language Models (LLMs)
AI systems trained on vast text corpora to generate and analyze text.
How the laws define this
Regulators consistently frame Large Language Models as artificial intelligence systems or statistical models trained on very large text corpora to understand, transform, predict, or generate natural language. Most frameworks, such as the Generative AI Framework for HMG and the Guidance to civil servants on use of generative AI, treat them as a subset of generative AI operating through pattern-based token selection. The central divergence lies between technical descriptions and procurement-focused compliance mandates. The Egypt National Artificial Intelligence Strategy (Second Edition 2025-2030), National AI Strategy Roadmap 2.0 (NAISR 2.0), Generative AI Framework for HMG, and the Joint AI plan for the safe and effective use of AI in the Norwegian health and care services 2024–2025 emphasize technical capabilities, national integration, or specific risks like hallucinations and privacy. In contrast, Executive Order 14319 — Preventing Woke AI in the Federal Government omits technical descriptions entirely, defining LLMs solely as a targeted model type required to adhere to Unbiased AI Principles when procured by federal agencies.
Synthesised from the 6 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 4 of the 6 instruments that define this term; all are listed under Definitions below.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| United States | Executive Order 14319 — Preventing Woke AI in the Federal Government | In Force | Uniquely ties LLMs to federal procurement requirements under Executive Order 14319, mandating adherence to Unbiased AI Principles to avoid political or ideological biases. |
| United Kingdom | Generative AI Framework for HMG | Withdrawn | Identifies LLMs as a subclass of generative AI with specific risks like hallucinations and assurance needs for government use. |
| Norway | Joint AI plan for the safe and effective use of AI in the Norwegian health and care services 2024 - 2025 | Superseded | Limits LLMs to the health and care services sector, treating them as a distinct category needing dedicated resources and risk assessments for hallucinations, privacy, and misuse. |
| United Kingdom | Guidance to civil servants on use of generative AI | Superseded | Specifies LLMs produce text via probabilistic token selection based on learned patterns, referencing public tools like ChatGPT and Bard. |
What it triggers
- In the United States, federal executive agencies are prohibited from procuring LLMs that embed political or ideological biases and must adhere to Unbiased AI Principles.source
- The Generative AI Framework for HMG sets out ten core principles and identifies specific risks and assurance needs for LLMs in UK government use.source
- The Guidance to civil servants sets out general principles for safe, lawful and responsible use of LLMs until it was superseded.source
Definitions (6)
AI systems trained on very large amounts of textual data that can generate human-like text, perform translation, sentiment analysis, and other language tasks. The strategy explicitly defines LLMs as key technology components for both the national foundational model and subsequent domain-specific models.
AI models trained on very large text corpora that can understand, generate, and transform human language in fluent and context-aware ways. NAISR 2.0 highlights LLMs as a key class of models with broad economic and social impact requiring stewardship and integration into sectoral strategies.
Statistical models trained on large text corpora to generate or predict natural language; the Framework identifies LLMs as a prominent subclass of generative AI with specific risks (e.g., hallucinations) and assurance needs for government use.
AI systems trained on very large text corpora to generate, summarise or transform natural language outputs; in this plan they are treated as a distinct application category requiring dedicated knowledge resources and targeted risk assessments (e.g., for hallucinations, privacy and misuse).
Large Language Models (LLMs) are a common subset of generative AI that produce text by probabilistic selection of tokens based on learned patterns; examples referenced include public tools such as ChatGPT and Bard.
A type of artificial intelligence model specifically targeted by Executive Order 14319, which are required to adhere to the Unbiased AI Principles when procured by federal agencies.
Not to be confused with
Large/fundamental models are defined by parameter size (≥1 billion) as bases for many tasks, while LLMs focus on text corpora training for language generation.
Language models predict or generate text in general, whereas LLMs are specified as large-scale versions with training on vast corpora and additional attributes like risks or principles.
Questions and answers
What obligations apply to LLMs in US federal procurement?
Executive Order 14319 prohibits federal agencies from procuring LLMs with political or ideological biases. It mandates adherence to Unbiased AI Principles for truth-seeking and nonpartisan AI. In force since July 23, 2025.
How does the UK Generative AI Framework treat LLMs?
It identifies LLMs as a prominent subclass of generative AI with risks like hallucinations. The framework provides guidance on safe procurement and deployment for civil servants and public bodies.
What risks are highlighted for LLMs in Norwegian health services?
The Joint AI plan treats LLMs as needing risk assessments for hallucinations, privacy, and misuse. It calls for dedicated knowledge resources in health and care services until superseded.
Related Terms
Large/fundamental models
AI models (≥1 billion parameters) serving as bases for many tasks....
General-Purpose AI (Foundation Model)
Large models designed for a wide range of downstream tasks....
National Foundational Model
A large-scale Arabic/Egyptian base model for domain adaptation....
Large Language Model (LLM)
Class of generative models producing human‑like text....
Language model
A model that predicts or generates natural-language text....