Written and maintained by CASRAI Editorial Board
Last updated
What the Frontier Model Forum is
The Frontier Model Forum (FMF) is an industry-supported non-profit whose stated purpose is to help ensure that the most advanced (“frontier”) AI systems remain safe and secure. It was founded on 26 July 2023 by Anthropic, Google, Microsoft and OpenAI. Amazon and Meta have since joined; as of this writing the Forum lists six member companies: Amazon, Anthropic, Google, Meta, Microsoft and OpenAI.
The FMF is not an evaluator, a standards body with enforcement power, or a government agency. It is a coordinating body run and funded by the frontier AI labs themselves — a structural point that matters when comparing it to the independent and government evaluators covered elsewhere in this cluster.
What the Forum says it does
The FMF organises its work around three mandates, stated on its own site: identifying best practices and supporting standards development for frontier AI safety and security; advancing independent AI safety research, including through the FMF-managed AI Safety Fund; and facilitating information-sharing on AI safety and risk across companies, governments, academia and civil society. One of its own stated objectives is “advancing AI safety research to promote responsible development of frontier models, minimize risks, and enable independent, standardized evaluations of capabilities and safety.”
In practice this has included a first-of-its-kind information-sharing agreement among member companies on AI safety and risk, announced by the Forum, and ongoing workstreams organised around specific risk areas: AI-bio, AI-cyber, AI-nuclear, AI security, and safety frameworks.
Frontier capability assessments
A recurring piece of FMF work is what it calls frontier capability assessments: procedures intended to identify whether an advanced AI model has capabilities that could meaningfully increase risk in areas such as chemical, biological, radiological or nuclear (CBRN) weapons development, advanced cyber threats, or autonomous behaviour. These assessments are produced and used within the Forum’s member-company workstreams rather than published as independent third-party evaluations of a given model.
Who can join
The FMF publishes three qualification standards for membership. A candidate must demonstrate frontier AI capability — the expertise and resources to develop or deploy frontier AI models at scale. It must show a safety track record: publicly acknowledging risks, disclosing risk-mitigation guidelines, having processes to assess model safety through the model lifecycle, and supporting independent AI safety research and evaluations. And it must commit to contribute — participating in FMF workshops and activities and financially supporting the Forum’s safety efforts for a minimum of three years.
How the Forum differs from independent and government evaluators
The Frontier Model Forum sits in a different part of the AI safety ecosystem from the third-party evaluators and government testing institutes covered elsewhere on this site. The distinction is about who funds and governs the work, not just who performs it.
- Independent evaluators, e.g. METR. METR is a nonprofit that states it has not accepted funding from AI companies, and treats that independence as a condition of being able to evaluate those companies’ models credibly. The FMF is the opposite funding structure: it is financed and governed by the same labs whose models its members build and deploy.
- Government testing institutes, e.g. UK AISI and US CAISI. These are state bodies that can secure pre-deployment access to models under statutory or agreed arrangements. The FMF has no equivalent statutory authority; membership and information-sharing are voluntary commitments among participating companies.
- Cross-cutting evaluator standards. CASRAI’s overview of third-party evaluator standards maps independence, access, validity, publication rights and retaliation protection across evaluator types generally. The FMF is not itself an evaluator against that framework — it is one of the industry bodies whose member labs are the subject of, and sometime funders of, that evaluation activity.
This funding and governance distinction is also the reason the Forum sits differently against NIKOLAI’s N8 evaluator-independence element, which sets out the conflict-of-interest standards a reviewer needs to meet to be treated as an impartial third-party evaluator. An industry-funded, industry-governed coordinating body such as the FMF is not attempting to meet that bar, and CASRAI’s NIKOLAI dictionary does not treat it as an independent evaluator for that reason.
Frequently asked questions
Is the Frontier Model Forum an AI evaluator?
No. It is an industry-led coordinating and standards body funded and governed by its member AI labs, not an independent evaluation organisation such as METR or a government testing institute such as UK AISI or US CAISI.
Who funds the Frontier Model Forum?
Its member companies. Membership requires committing to financially support the Forum’s safety efforts for a minimum of three years, alongside participating in its workshops and activities.
When was the Frontier Model Forum founded, and by whom?
It was founded on 26 July 2023 by Anthropic, Google, Microsoft and OpenAI. Amazon and Meta have joined since founding, bringing current membership to six companies.
What is a frontier capability assessment?
A procedure the FMF uses to check whether an advanced model has capabilities that could increase risk in areas such as CBRN weapons development, advanced cyber threats, or autonomous behaviour.
How is the Frontier Model Forum different from UK AISI or US CAISI?
UK AISI and US CAISI are government bodies that can obtain pre-deployment access to models under statutory or agreed arrangements. The FMF has no such authority: it is a voluntary, industry-funded body with no government mandate.
