FOUNDATION MODEL
Definition
A foundation model is a large-scale machine learning model — trained on broad, general domain data and intended to be adaptable across many downstream applications through fine-tuning, prompting, or retrieval. The term was introduced in the 2021 Stanford report On the Opportunities and Risks of Foundation Models and has become the dominant terminology in policy and regulatory contexts. The category includes large language models (Claude, GPT-4, Gemini, Llama), multimodal models, and emerging vision and audio foundation models. The closely related term “frontier model” — used in the U.S. Executive Order on AI, the UK AI Safety Institute framework, and the EU AI Act discussion — refers to the largest and most capable foundation models at the frontier of capabilities, typically subject to additional regulatory or self-governance requirements.
Why it matters for communications
Foundation model is the dominant policy and regulatory term — required vocabulary in any communications engagement with regulators, congressional staff, or policy press. Frontier model is its near-synonym in capabilities-based regulatory contexts. Communications strategy in any AI-policy-adjacent category (financial services AI, healthcare AI, defense AI, AI safety) requires accurate use of both terms.
Related terms Large Language Model · Frontier model · Pretraining · Fine-Tuning · Compute threshold
Related entities Stanford CRFM · OpenAI · Anthropic · Google DeepMind · Meta · UK AI Safety Institute · NIST · EU AI Office
Primary sources Bommasani et al., On the Opportunities and Risks of Foundation Models (Stanford CRFM, 2021) · EU AI Act · NIST AI Risk Management Framework · UK AI Safety Institute publications.