PUBLISHED 2026-06-14 to LinkedIn Post
AI is becoming infrastructure — the thing the world learns, decides, and creates with. So who controls it? Days ago, we got an uncomfortable answer: Anthropic disabled worldwide access to two frontier models — Claude Fable 5 and Mythos 5 — after a government order.
We argue about who can turn AI off. But can we believe it when it's on? Should every country have to build or back its own AI just to control and trust it?
There's a better way: AI built like the internet — open, plural, and resilient enough that no single, unaccountable hand controls it — and honest enough that its material claims can be traced, tested, and challenged.
I've written a short, first-draft manifesto for what that could look like: five public obligations, and a first practical test for evidence, correction, and contestability. It's a draft, and I'd genuinely like it torn into — especially by the people who build and study these systems. Is it right, or naïve? Where does it break? What's missing?
If you build AI, fund it, study it, govern it, teach it, or simply rely on it — you belong here. 👇 Let me know your thoughts and ideas.
#AI #TrustworthyAI #AIGovernance #OpenSource #PublicAI
The Anthropic access event this post cites is attributed to Anthropic's public statement.
Trustworthy AI, Accountable to People¶
What just happened¶
Last week, a government order forced Anthropic to cut global access to its two newest models. Anthropic complied and publicly objected — calling the process non-transparent and technically ungrounded. It didn't matter.
The precedent is set. The models may return. The precedent won't.
And there's a second problem alongside it: whether we can trust AI when it's on. Today's systems state falsehoods with the confidence of fact, cite sources that were never written, and seldom signal their own doubt. What the world now thinks with has become both fragile — switchable by a few hands — and unaccountable for what it tells us.
That's the urgency. It won't fix itself.
The world to build instead¶
A world where AI is a shared, trustworthy foundation for all of humanity — open, plural, and answerable to people — that no single company or government can switch off or bend, and whose material claims can be traced, tested, and challenged, with limits and unresolved risks visible.
The obligations¶
Trustworthy AI — and those who build and run it — must meet obligations.
-
Purpose-bound. A system must say what it is for, what it must not be used for, who may be affected, what risks are known, and when its claim to trust needs to be checked again.
-
Answerable to people. A named human institution must answer for the system's claims, harms, incidents, data use, and consequential decisions. People need notice, review, challenge, and remedy where the system affects them.
-
Safe, secure, private, and resilient. A system must protect people, data, infrastructure, and public-interest reliance: safety controls, cybersecurity, privacy, misuse handling, continuity, and exit paths cannot be afterthoughts.
-
Fair in practice. Where an AI system may materially affect people, rights, opportunities, or access to services, it must be tested and monitored for materially uneven or discriminatory performance across relevant groups, languages, regions, and contexts. Known limits must be disclosed, credible harm signals escalated, and harms mitigated, corrected, constrained, or withdrawn where necessary.
-
Open to evidence and correction. For material claims, recommendations, decisions, and risk-relevant system behavior, the system and its operators must preserve enough evidence for appropriate review: outputs, claims, sources where used, uncertainty signals, known limits, incidents, material changes, drift signals, complaints, corrections, withdrawals, and unresolved risks must be inspectable by responsible parties, triaged, and acted on proportionately.
It means contestable control: lawful recall, misuse safeguards, and named operators who answer for the system publicly — never secret kill-switches.
The plain test: can one party withdraw it for everyone, in secret, with no appeal? And when it makes a factual claim, can you trace, test, and challenge the basis?
How you'd know it's real¶
Obligations mean nothing if anyone can claim them. The endpoint is independent verification — not a new central authority (no one should hold that power), but a separation of roles so no one marks their own homework:
Someone writes the standard → independent assessors audit against it → an accreditor licenses and polices those assessors → a peer body keeps the accreditors honest.
This is the same chain behind ISO certification, organic and forestry labels, Fairtrade, and journalism's own trust badges. What makes a mark trustworthy is not that an authority issued it, but that no single actor in the chain checks their own work.
It can start light and ride accreditation infrastructure that already exists. A direction, not a finished scheme — but the difference between a slogan and a standard.
Who could drive it¶
Every working scheme has a credible initiator and a neutral host. An NGO stewards it — Reporters Without Borders drove the Journalism Trust Initiative; Poynter runs the IFCN. A neutral body hosts it — the Linux Foundation, OASIS, CEN. A coalition runs it — the CA/Browser Forum, the AI Alliance.
The model: a small founding coalition convenes, the standard is hosted inside an existing body, and groups already doing the work — MLCommons, the AI Alliance, Current AI — anchor it.
How it would be funded¶
No single funder holds the leash. Public-interest seed grants start Phase 1 — cheap to launch, light to run. Capped fees from assessed organizations (with hardship waivers, so no one is priced out) fund ongoing operations. A broad, geographically diverse donor base prevents capture. Revenue from the mark itself scales with adoption.
Phase 1 needs a small group of founding funders willing to back the standard before the mark has value. That's the real ask.
What I'm asking¶
This cannot be built by one person or one organization.
The builders already exist: open-model labs, public compute efforts, research groups, standards bodies, and public-interest coalitions across regions. What is missing is the connective tissue: shared evidence records, evaluation methods, accountability routes, continuity plans, and correction processes.
Making AI outputs evidence-backed, challengeable, and open to correction is the harder frontier. It needs deliberate investment, not optimistic assumptions.
If you build AI: publish who can switch your system off, and on what grounds. Work toward these obligations.
If you buy or depend on AI: demand traceable, contestable outputs. Penalise unverified claims. Shift spending toward systems being built to pass the test.
If you shape policy or study this: tell me where it breaks. Help place it.
The world has already declared what AI must respect — the Universal Declaration of Human Rights, UNESCO's AI ethics recommendation, the Council of Europe's AI treaty. This is the part that turns those declarations into something you can insist on.
It is already beginning — Apertus and Sarvam, Latam-GPT and Lelapa, Mistral and Cohere. The pieces exist. What's missing is us, deciding to connect them on purpose.
Trustworthy AI isn't a product we're sold. It's infrastructure we hold in common — dependable, and answerable for what it tells us.
Built by many. Accountable to all.
If you build AI, fund it, study it, govern it, teach it, or rely on it — you belong in this conversation. Bring what you have.
Working draft, updated 19 June 2026. Source and work in progress: robertschaub/our-ai-charter.
Accompanying comment — who would actually certify this?¶
Published as a LinkedIn comment on the article, 15 June 2026.
A fair question I keep getting: who would actually certify this?
Preferably not by building a new authority. Instead, extend the independent, revocable trust-mark systems that already work: the way fact-checking runs its trust badges, or the way ISO-style accreditation certifies everything from organic food to data security.
The path: keep the standard open; start light (a published test, a few independent assessors, a public registry, and the power to revoke); have the auditing done by accredited, independent third parties, never by whoever wrote the standard; host it inside a body that already exists, and grow into formal accreditation.
Build the test, borrow the machinery.