What the next wave complicates
Four developments are arriving together. Each is useful. Each moves the trust question somewhere new.
Mesh AI: agents talking to many other agents
What it promises: tasks handed from one AI agent to another across different suppliers' systems, with people stepping in only where it matters. Open protocols for this now exist under neutral stewardship: the Linux Foundation's Agent2Agent project, launched in June 2025, describes agents that find one another, exchange information securely and work together across systems.
What it complicates: when agents act on each other's instructions, a mistake, a hallucination or a planted instruction can travel the mesh before any person sees it. Trust can no longer rest on one agent being well behaved. It has to rest on each action being checked against a rule someone wrote down, and recorded. The plain question for any community adopting agents: which agent may do what, who decided, and where is that written?
Federated and small language models
What it promises: instead of one large model in one place, many small models run close to the data they serve, in a school, a council, a rūnanga. Less data leaves the building, and a model can be tuned to local language and context. Federated approaches go further: models learn from data that never leaves its home.
What it complicates: more models means more things to vet. Who trained this one, on what, and who can change it? A small model can be wrong in locally confident ways, and nobody outside has tested it. The trust question moves from “where is my data?” to “who checked this model, and can I see the record?” A community adopting one should ask for the model's provenance the same way it would ask for a supplier's.
Sustainable AI: smaller, regional data centres
What it promises: data centres designed for low water and energy use, sited in regions and serving a district's schools, clinics and councils.
What it complicates: a building in your region is not the same as a provider under your law. A regional centre owned and operated from abroad is reachable from abroad, and a sustainability claim is easy to make and hard to check. The questions are the same seven as for any sovereign claim, plus one more: show us the water and energy figures, measured rather than modelled.
Post-quantum cryptography
What it is: the encryption and digital signatures that protect almost everything online today rest on mathematical problems that a large enough quantum computer could solve. Post-quantum cryptography uses different problems that are believed to resist such a machine. In August 2024 NIST, the US standards body, published the first three finished standards: FIPS 203 (ML-KEM, for encryption keys), FIPS 204 (ML-DSA, for digital signatures) and FIPS 205 (SLH-DSA, a second signature method built on different mathematics as a backup).
Why “a few years” is the horizon: NIST noted that some experts expect a capable machine within a decade and urged administrators to start bringing in the new standards straight away. The UK's National Cyber Security Centre sets milestones of 2028 for a plan, 2031 for the highest-priority systems and 2035 for completion. The reason to start before the machine exists is simple: anything encrypted today can be stored and read later, so the protection a record needs is measured by how long it must stay private, not by when the threat arrives.
What it complicates: every signature, every certificate and every time-stamp will eventually have to be redone with new algorithms, and a record that cannot be re-signed is a record with a shelf life. Nobody can promise safety from a machine that does not yet exist. What a community can do is list the records that must last, ask every provider for its plan, and prefer records that can be checked without the provider.
For our part: a second signature using the post-quantum algorithm ML-DSA (FIPS 204) is built into several Village record types and turned on per account, and we plan to add it to Gateway receipts. The implementation has not been independently audited. It is a step, not a destination.
Sources
Every outside fact on this page rests on one of these sources.