Artificial General Intelligence is only a few short years away. This independent layer exists so the people closest to the technology can advance the science of evaluation, the norms, and the standards a real Frontier AI Standards Body will need.
Not the official body. An independent layer for evaluation science, norms, and standards.
Vision
“When we look back on this time in the decades to come, I think we will realise we were standing in the foothills of the singularity - nothing less than the dawning of a new age for humanity.”
Artificial General Intelligence is only a few short years away. The magnitude of its impact may be an order of magnitude greater than the Industrial Revolution, at an order of magnitude greater speed. We have found a way to make sand think. The window to shape how this technology enters the world is open now - and it will not stay open indefinitely.
We are locked in an intense commercial and geopolitical race. Competitive dynamics are accelerating progress and unlocking real upsides, but they are also outpacing our understanding of the technology and the development of shared safeguards. Nobody knows for certain what happens next. When uncertainty is this high and the stakes are this large, cautious optimism and deliberate coordination are the only responsible strategy.
Demis Hassabis proposed a U.S.-led Frontier AI Standards Body: a technically rigorous public-private partnership, modelled in part on self-regulatory organisations like FINRA - with independent technical experts and open-source representation on its board. Its core functions would include:
The framework is designed to be adaptive, technically focused, and eventually international. It aims to buy time for responsible development without freezing progress.
The real Standards Body will be built by governments and labs. But the people who see the consequences first-hand: researchers, engineers, safety teams, and operators inside frontier organisations - currently lack a clean, high-signal public channel to contribute ideas, surface risks, propose evaluation criteria, and refine the technical foundations before decisions are locked in.
Standardsbody.ai exists to provide that channel. An independent space where the people closest to the technology can advance the science of evaluation, the norms, and the standards a real Frontier AI Standards Body will need - evaluation ideas, open questions, concrete proposals, and first-hand signal - free of corporate filters and political noise.
Approach
DeepMind and other frontier labs describe responsibility and safety as governance, research, and impact. This independent layer focuses on a narrower public good: the evaluation science, norms, and institutional ingredients a Standards Body will require.
Surface and pressure-test ideas without corporate filters or political capture. First-hand signal from researchers, engineers, safety teams, and operators.
Distill contributions into living foundations, open questions, and concrete evaluation or institutional proposals that can inform a real Standards Body.
Keep a visible trail of what has been established so serious input does not disappear. Attribution only with consent. Versioned where it matters.
Foundations · Living document
Version 0.1 · Last updated July 2026
Demis’s proposal is specific. A Frontier AI Standards Body will only work if certain technical and institutional ingredients are present. These eight are the core of the evaluation science, norms, and standards layer.
Benchmarks that are updated frequently (quarterly at first), with saturated or outdated tests deprecated. Static evals will be gamed; the standards body must keep pace with capability progress.
Eventually the body must develop its own private evaluations that frontier labs cannot train against. Without this, overfitting to public benchmarks becomes inevitable.
Cybersecurity, biological risks, nuclear-adjacent capabilities, agentic behaviour, deception, and self-improvement. These are the areas where catastrophic failure modes first appear.
The board and technical staff cannot be captured by any single lab or government. Independent researchers and open-source representatives are structural requirements, not optional.
The body itself cannot evaluate every model forever. It must cultivate and oversee a network of competent third-party auditors who can scale with the number of frontier systems.
Start with voluntary pre-release review (e.g. up to 30 days). Once the protocol is proven effective, formalise it so that frontier models must pass to be deployed in the U.S. market.
“Frontier Lab” designation should carry real weight. Organisations should want to be inside the process. Prestige and access to the technical community are more powerful than pure coercion at the beginning.
A U.S.-initiated body is the right starting point, but the design must anticipate shared international standards. Frontier models will not respect borders; the evaluation regime eventually cannot either.
This independent layer exists to help surface, pressure-test, and refine exactly these ingredients before the real Standards Body is formalised.
Directions
We are early. The shape of this independent layer is not fixed. These are the directions currently under consideration. Each one is a hypothesis about how to be most useful to a future Standards Body. Your input on priority and design is wanted.
A moderated forum where researchers, engineers, and operators can post precise proposals, critiques, and first-hand observations. Quality over volume. Pseudonymous options available. Permanent public archive so good ideas do not disappear into private chats.
Long-form essays, open letters, technical proposals, and independent analyses. Versioned, citation-ready, and curated so that the best work becomes part of a durable knowledge base rather than a fleeting feed.
A channel for people who see actual capabilities, risks, and failure modes every day. Attributed or anonymous. The goal is first-hand technical signal that corporate communications and public papers cannot or will not carry.
Focused, time-bounded groups that attempt to produce concrete outputs: draft evaluation criteria, open questions lists, threat models, or institutional design sketches. Less “community,” more temporary research teams.
Optional and only after proper legal structure. A transparent pass-through mechanism for directing resources to established AI safety and alignment organisations that share the goal of rigorous, independent standards work. Not a priority until the intellectual layer is real.
The most valuable path may not be on this list. If you see a higher-leverage way for an independent layer to strengthen a future Standards Body, that input is more useful than any of the options above.
Tell us what we are missing →Open Questions · Living document
Version 0.1 · Last updated July 2026 · Seeded from Demis Hassabis’s proposal
The science of evaluation and the design of norms and standards depend on questions that are still open. This list is living. High-signal refinements and new questions are welcome.
This list is intentionally incomplete. The most valuable contributions will be questions we have not yet thought to ask, or sharper formulations of the ones above.
Proposed questions and refinements are reviewed and, when high-signal, folded into this living document (with attribution only if requested).
Propose a question or refinement →Public record
Version 0.1 · Updated July 2026 · Living
High-signal contributions that are ready for public view appear here. This section starts sparse on purpose. It will grow only when the quality threshold is met.
Initial research-grade questions spanning evaluation, institutional design, incentives, and harder horizons, seeded from Demis Hassabis’s July 2026 proposal and independent analysis. This list is the first public artifact of the independent layer.
Distilled technical and institutional requirements drawn directly from the original proposal: dynamic protocols, held-out tests, high-stakes domain coverage, independent experts, third-party auditors, voluntary→mandatory path, prestige incentives, and international interoperability.
When high-signal submissions clear review, they will appear here as public signals: synthesised questions, short proposals, or attributed observations. Attribution only with consent.
This record exists so contributors can see that serious input does not disappear. Submit a contribution →
Contribute
Contributions to the science of evaluation, norms, and standards - open questions, technical proposals, first-hand observations, or offers to help.
How contributions are used
Private by default. Every submission goes into a private review queue. Nothing is published automatically.
High-signal work is synthesised. Strong ideas, questions, and proposals are used to improve the public Open Questions list, Foundations notes, and future synthesis documents. Attribution is only applied if you explicitly allow it.
Nothing disappears into a void. Even if a contribution is not published, it informs internal prioritisation and the shape of this independent layer. You will not receive automated replies; if direct follow-up is warranted, we will contact you.
Email is never shared or sold. We do not publish under your name without consent.
Received
Thank you. Your submission is now in the private review queue. High-signal contributions will be considered for the public Open Questions list, Foundations notes, or future synthesis, with attribution only if you indicated that is welcome.
There is no automated reply. If follow-up is warranted, we will reach out directly.
The real Standards Body will be built by governments and labs. This independent layer exists so that the people who see the technology most clearly: researchers, engineers, safety teams, and operators - can help advance evaluation science, norms, standards, and open questions before those decisions are locked in.
standardsbody.ai · thestandardsbody.ai