A label for the most capable general-purpose AI models at any given time, the ones that match or exceed the abilities of the most advanced systems so far. It is a moving target by design: what counts as frontier shifts as the technology advances. The term is used mostly in policy, where it marks out the systems thought to carry the most serious and least understood risks.
Technical view
Popularised in 2023, notably by Anderljung and colleagues’ July paper on frontier AI regulation, which defined frontier models as highly capable foundation models that could have dangerous capabilities severe enough to threaten public safety, and by the Bletchley Declaration, which described highly capable general-purpose models matching or exceeding the most advanced of the day. Because ‘most advanced’ is relative, legal instruments substitute measurable proxies such as training compute, as in the European Union’s systemic-risk threshold. Critics argue the framing focuses on speculative harms from a few large developers at the expense of present harms from ordinary systems; supporters reply that capability best predicts novel risk.
Related
Part of AI Terminology Index, section: Governance, Law and Regulation.