Start from the failure mode, not the model
Every public-service problem Mosaic maps has a failure mode: a document that can’t be verified in time, a record that can’t follow a person, a standard that doesn’t exist yet. AI is useful exactly where it reduces one of those specific failure modes — and a distraction everywhere else. This framework is a filter, not a pitch.
Three legitimate roles for AI in a DPI stack
- Matching and triage — pointing a case, a document, or an anomaly to the right human faster. Useful in fraud detection on a shared exchange layer, or triaging which farmer-registry records need manual review.
- Drafting, not deciding — generating a first draft of a standard, a translation, or a summary that a person still signs off on. The credential or the decision itself is never machine-authored without review.
- Pattern-finding across published evidence — once pilots publish their results in the open (see our activities), AI can help surface patterns across them faster than a person reading every whitepaper manually.
Where it doesn’t belong, by default
- Issuing a Verifiable Credential. A credential’s trust comes from a named, accountable issuer. An AI system is not an issuer.
- Making an eligibility or benefits decision alone. Automating a decision that affects someone’s access to a service without an appeal path reproduces the exact fragility DPI is meant to fix.
- Standing in for consent. Consent has to come from the person or institution it concerns — a model inferring “likely consent” is not consent.
The test we apply to a proposed AI use
Before Mosaic treats an AI application as part of a use case rather than a side experiment, we ask:
- Can a person understand and challenge what the system did?
- Does it fail safely — does a wrong answer degrade gracefully, or cascade?
- Is the training or reference data itself governed the same way the rest of the use case’s data is (consent, provenance, minimum necessary)?
If the answer to any of these is no, the use case isn’t ready for that AI component yet — not because AI is disallowed, but because the governance underneath it isn’t there yet. That’s consistent with how Mosaic treats every other shared capability: proof comes before policy, and the boring infrastructure (consent, identity, audit trails) comes before the model.