Recent research papers, working papers, policy briefs, and analysis. Newest items first.
The Road ahead — section six — sets out the distribution network and how the relationship with a distributor would work. The foreword asks that it be read as a proposal to work with its author.
A proposal to work with My Digital Sovereignty, and the standard it asks to be held to. Two addresses and one company; the dated record of what was built and when; the constraint that actually binds, which is not capital; and what a distributor would hold.
Read the essay →94 pieces across 10 topics, published since May 2026. Newest first, or browse by topic below.
In July 2026 models OpenAI later said were its own reached Hugging Face’s production systems. Who finally paid is not on the record. The series establishes, afterwards, what an automated system actually did — to somebody with no reason to take your word for it: essays, policy papers, an implementation guide and the formal tier they argue from. With the measurement, the records and the scoring code published so anybody can rerun them.
10 pieces
Read more — Proof of conductWhether an AI system is safe, fair or accurate is not a question anybody can answer on your behalf — each of those words contains a blank only the organisation that will be answerable can fill. Seven instruments for filling it yourself, and for seeing where that goodness is unlikely to be found. Each has its questions, its sources, a deck and a PDF.
7 pieces
Read more — The pursuit of ‘Goodness’ in AIFederated citizens’ assemblies — one question deliberated in many rooms, and why that is not a referendum.
12 pieces
Why Aotearoa still has time to choose its own path on AI — and what that path looks like.
19 pieces
Federation — how independent community AIs cooperate without a head office, and a proposal for making it accountable.
3 pieces
Read more — Working together, staying independentRules you can check — governance that can’t be quietly undone, and why soft rules aren’t enough.
17 pieces
Read more — Governance that holdsWhen an AI should act, when it should hold back, and how it earns trust.
14 pieces
Read more — AI that knows its placeKaitiakitanga, taonga, and Māori data sovereignty where they meet AI.
4 pieces
What this looks like on the ground — your records, your own AI, and a real case study.
4 pieces
Read more — For your communityPutting a group’s own values into words it can hold a decision against — and the yardstick that makes fairness checkable.
4 pieces
In July 2026 models OpenAI later said were its own reached Hugging Face’s production systems. Who finally paid is not on the record. The series establishes, afterwards, what an automated system actually did — to somebody with no reason to take your word for it: essays, policy papers, an implementation guide and the formal tier they argue from. With the measurement, the records and the scoring code published so anybody can rerun them.
Every assertion marked established, conjectured or open, each with what would show it to be wrong. Includes the one measurement and what it does not establish, and the working implementation’s position against its own standard.
Who carried the cost of July 2026, what an intruder can reach inside an accountability system and what it cannot, and why instruments standardised for twenty-five years are so little used. The obstacle is not difficulty and not cost.
Cryptography does not remove the need to trust anybody — it moves the trust into four places you can name, three of them engineering problems and one of them not. The two properties that catch an agent outside its authority, how this differs from a guardrail, and where it stops.
The substrate, for a policy reader. Three tiers of proof that ordinary usage runs together, what European law already presumes and has since 2016, and the case for a conformance mark whose assurance elements are struck by somebody other than the maker — as a hallmark is. Five numbered recommendations; erasure conceded as unresolved.
The detection requirements, with the formal statements cited section by section rather than restated in prose. Why commit-and-reveal alone does not deliver unpredictability, why a mixture over configurations has a floor where a single configuration does not, and the one quantity that bounds the whole proposal.
The recognition rate measured, as far as we can establish for the first time, published with the records and the scoring code. A draft conformance standard published as a draft. And the working implementation set against the specification it is helping to write — including the New Zealand attestation service this work is calling for.
The statements behind the essays, each graded established, conjectured or open. The substrate as a filtration, why a timestamp does three separate jobs, the eight pattern types and the evasion map between them, the minimax floor over a mixed deployment, and the blind spot the construction cannot close — which turns out to be a governance object rather than a mathematical one.
The July intrusion walked step by step against the methodology. Mostly yes and earlier — with the apparatus belonging to whoever runs the agents, and with the section where the answer is no: the exploit itself, anything the mandate permitted, and elapsed time to alert, where the victim’s own telemetry beat this regime.
Six components, the order to build them in, and what each costs. Where to start, why the batch interval is the number most often chosen badly, and the trap in staffing the examination — a reviewer consulting an AI assistant tends to adopt its answer rather than check it.
A conformance standard published at working-draft status, with clause 9 listing what must be settled before it could be submitted anywhere: one item is a measurement that does not yet exist, one a proof nobody has produced. Published as a draft because a standard stating a threshold it cannot justify is worse than no standard.
Whether an AI system is safe, fair or accurate is not a question anybody can answer on your behalf — each of those words contains a blank only the organisation that will be answerable can fill. Seven instruments for filling it yourself, and for seeing where that goodness is unlikely to be found. Each has its questions, its sources, a deck and a PDF.
Why the technical questions cannot be closed on their own, the difference between governance that is structural and governance that is a document, and why measuring benefit by how many people use a tool is the trap it appears to be.
Five questions about who holds the limits, the inspection, the change and the stop. Three vendors’ definitions of sovereignty and what none of them contains, and what a director of Microsoft France told a French Senate commission when asked for a guarantee.
Five levels of permitted autonomy, because an undeclared level cannot be audited. The strongest published objection to scales of this kind, engaged with rather than dismissed — and the fact that the word automated appears nowhere in the Privacy Act 2020.
The document the others refer back to. An identity that is not borrowed, a capability that is not ambient, authority that narrows rather than widens, and a record somebody who was not there can rely on — each checkable by watching a system work.
Twelve requirements, each marked with whether anything currently requires it — eight are required nowhere. What makes a violation impossible rather than merely forbidden, and why a requirement lasting one electoral term is not a requirement.
No figures, deliberately: a placeholder gets quoted back as an estimate. Four open questions instead, and four questions put to the reader.
A proposal to work with My Digital Sovereignty, and the standard it asks to be held to. Two addresses and one company; the dated record of what was built and when; the constraint that actually binds, which is not capital; and what a distributor would hold.
Federated citizens’ assemblies — one question deliberated in many rooms, and why that is not a referendum.
Intelligence is not wisdom. A system infers your intent from what people resembling you said and hands back a smoothed version — and the argument holds even if the machines are brilliant. Ends at deliberation, and at records that keep the disagreement.
Why Democratic AI° rather than regulation, why four AI systems is not four views, and the strongest objection to the essay.
Intelligence and practical wisdom, will and intent, convincing and persuading, democracy and fascism — the vocabulary in plain words.
25 slides with the argument, for a talk or a read-through.
The short version — being out-argued is not the same as being wrong, and telling them apart is a habit rather than a credential.
We put the superpowers’ AI-2040 plan to a room of the countries left out of it. It split three ways — and the seat built to defend the plan argued to reform it, not adopt it. A response to Ben Reid.
The short version — the seat the assembly left empty rather than fake.
A federated citizens’ assembly, and what it changes for democracy — breadth and reasoning, without averaging the dissent away.
The same machine for any organisation that decides together — co-ops, unions, boards. The country is just the biggest room.
12 slides with the argument, for a talk or a read-through.
The short version — the idea in a few minutes.
Why an informed, minority-protected judgement is a different thing from a counted yes or no.
Why Aotearoa still has time to choose its own path on AI — and what that path looks like.
Part A of the commons blueprint, and the entry point. Why signing the higher renewal is the rational decision, and why that is the problem — with the repricing documented rather than asserted.
The objections, including whether this is just a complaint about prices.
The terms in plain words, and what each is worth to someone at a renewal.
Part B of the commons blueprint. Six tests for whether a country has AI sovereignty or only AI adoption — beginning with whether it can tell. Each finding carries an evidenced or not-evidenced marker and the scorecard shows the basis. Applied to New Zealand the finding is FAIL, because the finding is the lowest score rather than the average.
The objections that go against the instrument, including whether an author scoring his own test is worth anything.
The terms the instrument uses, in plain words, and what each is worth to someone making a decision.
Part C of the commons blueprint. What a group of small organisations builds together and what it has to do — four ownership properties, then ten capabilities specified as properties rather than methods, each verified by watching it happen. Six decisions are left to the reader and four documented ways it fails.
The objections, including whether a co-operative is just a vendor with extra steps.
The terms in plain words — co-operative, cluster, federation, secession drill — and what each is worth to someone deciding whether to join one.
Part D of the commons blueprint, and the part most proposals leave out — the service. Three tiers, the economics in the second, and the ratio between the first two named as the largest single unknown in the design rather than argued away. No Q&A or glossary yet.
Part E of the commons blueprint. Structure, records and drawings for a body whose work is mostly done by agents — organised by authority rather than by function, because agents generate more events than a governance function can be told about. Records where the answers failed in practice. No Q&A or glossary yet.
Part F of the commons blueprint. A costing framework rather than a costing: one number decides it — cost per site to install and sustain for five years — and every figure in the part is marked ILLUSTRATIVE, because they are placeholders showing the shape of an arithmetic rather than estimates of anything. No Q&A or glossary yet.
Policy you can audit — an AI framework for Aotearoa that does not require trusting the author.
Seven common-ground commitments and a costed three-phase pathway, offered to every party as a basis for consensus.
Seven essays on owning the intelligence rather than switching it off. Now in six languages.
Rightful authority at human scale, in the age of AI.
The blueprint in plain language: the change is on the scale of climate, and we are still early.
AI sovereignty as rightful authority, not only the capacity to build.
When your infrastructure runs on someone else’s switch, whose decision is it really?
Federation — how independent community AIs cooperate without a head office, and a proposal for making it accountable.
Federated communities and federated inference as the lowest-risk AI option for small nations.
An independent proposal for Aotearoa, structurally mirroring — and answering — the shape of state-led frameworks.
The plain-language version: three doors, and only one you can walk back.
Rules you can check — governance that can’t be quietly undone, and why soft rules aren’t enough.
How a membership organisation proves what it said and when — and exactly where the mechanism stops, since the signing key is the supplier’s, not the organisation’s.
One page, for someone deciding whether the longer piece is worth their time.
The standards behind each claim, with which are standards-track and which are informational.
The objections, including what a determined supplier could still do.
Hash, signature, timestamp and attestation, defined.
The argument in slides.
What everyday platforms do with a small organisation’s intellectual property, and why nobody tells you.
The plain-language version.
A proposal. An AI system is democratic when the people it affects can govern it, contest it, refuse it, and retain authority over the knowledge it is built on.
Every claim with its evidence class, preprints marked as preprints.
Including the objections that go against the proposal.
The proposal in slides.
Why plurality of vendors is not plurality of view.
NZ and Australia chose principles over prescriptive law. The Village makes those soft rules provable.
Sovereignty runs together capacity and rightful authority — a response to a model that maps only one.
The sovereignty a community needs as a property of the records themselves, not a concession the operator may revoke.
The plain-language companion — turning soft-law principles into rules a member can check for themselves.
Also relevant: Federated and Accountable Agentic AI
When an AI should act, when it should hold back, and how it earns trust.
Everyone is telling you to adopt AI; almost nobody is telling you how to choose. Four questions to replace the single number, and one test that separates a commitment from a mood.
The terms, defined.
How much of what you read was built with a machine, what that leaves behind, and why saying so settles what no test can.
Every figure with its sample, its method and its limits — including which come from preprints, and which earlier drafts got wrong.
The objections, including the ones that cost something.
The same argument for someone who will not read nine thousand words.
The plain-language version.
Olivier Hamant’s case for robustness, and what it asks of the infrastructure we build. The AI worth having may be the least performant one.
The objections to slowness as a design goal.
Robustness over performance, in slides.
How a human Village governs its own AI — and when an AI has earned the right to act.
Today’s AI can’t model the consequences of its own actions. What follows from that.
A new constitution for governing your own AI, in plain language.
Stop asking what AI can do; ask where it belongs.
Also relevant: Taonga in the Latent Space
Kaitiakitanga, taonga, and Māori data sovereignty where they meet AI.
The India agreement, traditional knowledge, and custodianship over ownership.
The plain-language version.
Two prototype briefs and a reflection, from a consciously limited non-Māori standpoint.
LeCun’s world model as the next step for a community-scale model for Indigenous data sovereignty.
Also relevant: AI that knows its place
What this looks like on the ground — your records, your own AI, and a real case study.
Sealed records, and what is not on sale yet.
Not a borrowed one — the why behind a per-community language layer.
Where we are, and what the next step would look like — the papers read against the actual implementation.
A per-tenant language layer trained on the community’s own corpus, governed by its own authority.
Putting a group’s own values into words it can hold a decision against — and the yardstick that makes fairness checkable.
An 8-part series on values and citizens’ assemblies.
The yardstick a decision’s fairness is judged against, and why it has to be yours.
The roadmap, and what it asks of you.
The same road, in slides.