The Frontier Model Forum (FMF) is an industry consortium founded in July 2023 by Anthropic, Google, Microsoft, and OpenAI to advance the safe and responsible development of frontier AI models — the most capable general-purpose AI systems at the frontier of performance. It operates as a collaborative, pre-competitive body focused on technical AI safety research, shared safety evaluations, red-teaming methodologies, and engagement with policymakers and civil society. The Forum functions as a bridge between frontier AI developers and regulatory bodies, aiming to establish shared technical standards and best practices without constituting a formal standards organisation or regulatory authority.
Overview
- The Frontier Model Forum emerged from growing concern that the most capable AI Systems — those with broad general-purpose capabilities across language, reasoning, code generation, and multimodal tasks — present qualitatively distinct risks compared to narrow AI applications.
- Founding context
- Announced 26 July 2023 with four founding members: Anthropic, Google, Microsoft, and OpenAI
- Formed in part to demonstrate voluntary industry leadership ahead of anticipated legislative intervention, including the EU AI Act and proposed US AI regulation
- Joined subsequently by additional frontier AI developers and cloud providers, broadening membership beyond the initial four
- Scope of focus
- Specifically targets the frontier of AI capability — models that are large-scale, general-purpose, and at or near state-of-the-art on a wide range of tasks
- Does not govern narrow AI systems, domain-specific models, or legacy AI applications
- Concentrated on AI Safety risks including misuse potential (e.g., support for Weapons of Mass Destruction design), misalignment risks, and societal harms from deployment at scale
- Organisational character
- A voluntary, non-binding industry body — not a standards organisation in the ISO/IEEE sense
- Member companies retain independent safety programmes and Responsible Scaling Policy commitments
- Operates via working groups on technical safety, evaluation, and policy engagement
Key Workstreams and Mechanisms
- Technical AI Safety Research
- Pooling resources and expertise to advance AI Safety Research in areas where cross-company collaboration yields faster progress than individual efforts
- Focus areas include Interpretability, robustness testing, and mechanistic understanding of model behaviour
- Shared Safety Evaluations
- A flagship goal of the Forum is developing a shared library of Safety Evaluation frameworks and Capability Evaluation tools
- Evaluations target dangerous capability thresholds: e.g., uplift for Biosecurity (CBRN — chemical, biological, radiological, nuclear threats), Cybersecurity exploitation assistance, and persuasion/manipulation at scale
- Enables consistent assessment methodology across different frontier models from different organisations
- Red Teaming Coordination
- Promoting and sharing methodologies for structured Red Teaming of frontier models before deployment
- Facilitating cross-company sharing of red-team findings on safety-relevant model behaviours
- Connects to government-affiliated red-team exercises organised through bodies like the AI Safety Institute
- Policy Engagement
- Engaging with national governments, international organisations, and regulatory bodies on AI Policy and AI Regulation frameworks
- Providing technical input to legislative processes (e.g., EU AI Act, US Executive Orders on AI, UK AI Safety Summit outputs)
- Interfacing with the OECD AI Policy Observatory, the Global Partnership on AI, the G7 Hiroshima AI Process, and the UN AI Advisory Body
- Knowledge Sharing on Best Practices
- Publishing guidance on Responsible AI deployment practices, model cards, system cards, and safety documentation standards
- Supporting development of shared vocabulary and taxonomies for AI risk levels — aligning with frameworks like the NIST AI Risk Management Framework
- Information Sharing on Safety-Relevant Incidents
- Facilitating confidential sharing of safety-relevant model behaviour findings between members, analogous to incident-sharing mechanisms in other critical industries
Membership and Governance
- Founding members: Anthropic, Google, Microsoft, OpenAI
- Subsequent members have included other major AI labs and cloud platform providers
- Governed by a steering committee with representation from member organisations
- Working groups operate on specific technical or policy topics, with defined deliverables
- Open to frontier AI developers meeting criteria related to capability threshold and commitment to safety principles
- Does not include open-source model developers as primary members by initial design, reflecting its focus on proprietary frontier development — a point of criticism from parts of the Open Source AI community
Relationships to Other Governance Initiatives
- The FMF occupies a distinct niche within the broader AI Governance ecosystem:
Relations (Semantic)
Standards and Context
- The FMF does not itself produce formal standards but actively engages with standards-producing bodies:
- NIST AI Risk Management Framework — FMF member commitments align with NIST AI RMF tiers for frontier model risk management
- SC 42 — international AI standards committee; FMF technical work feeds into SC 42 discussions
- EU AI Act — FMF member companies contributed technical expertise to the drafting of frontier model (“GPAI model”) provisions, particularly Article 51 requirements for systemic risk models
- UK AI Safety Summit (Bletchley) — November 2023 Bletchley Declaration co-signed by FMF member governments; FMF technical workstreams directly informed summit outputs
- Seoul AI Summit — May 2024 follow-on to Bletchley; FMF contributions to frontier AI safety commitments
- The FMF’s evaluation library concept aligns with the Model Evaluation requirements in the EU AI Act for GPAI models with systemic risk
Criticisms and Limitations
- Self-regulatory concern: as a voluntary industry body, the FMF lacks enforcement mechanisms; critics argue it may function more as a reputational shield than a genuine safety mechanism
- Membership gatekeeping: focus on proprietary frontier developers excludes Open Source AI organisations and academic safety researchers from governance
- Competitive tension: even pre-competitive collaboration faces limits from competition law and commercial sensitivity, constraining depth of information sharing
- Geopolitical gaps: dominated by US companies; Chinese frontier AI developers (e.g., Baidu, Alibaba, DeepSeek) are not members, limiting global coverage
- Resource asymmetry: smaller AI developers and civil society voices have limited influence relative to the largest member organisations