Exploring Solid as a persistent personal authorisation layer for AI across health and social services

Hi all,

I’m currently developing a research proposal and would really appreciate feedback from the Solid community before I finalise the direction.

The broader research area is Human-Centred AI and personal control over the data context available to AI systems. My particular interest is health, social care and social protection, where relevant personal information may already be distributed across multiple authoritative systems.

Rather than moving or copying all of that sensitive data into a Solid Pod, I’m exploring a different role for the Pod:

a persistent, user-controlled authorisation layer for AI access, while the underlying authoritative data remains within the organisations that hold it.

For example, a person might interact over time with several AI-enabled services:

  • an income-support navigator;
  • a health-support assistant;
  • a social-care navigator.

Each agent might request access to different information, for a specific purpose and duration.

The Pod would maintain a personal record of those authorisations, for example:

Income Support Assistant
  Household information   ACTIVE
  Income                  ACTIVE
  Purpose: income support
  Expires: 14 days

Health Support Assistant
  Functional assessment   ACTIVE
  Purpose: support navigation
  Expires: 30 days

Social Care Assistant
  Functional assessment   EXPIRED

Conceptually:

User
  ↓
Solid Pod
(agent / data / purpose / expiry / revocation / access history)
  ↓
Access gateway
  ↓
AI services
  ↓
Health / Social Protection / Social Care systems

The external systems would remain the authoritative sources of data.

The access gateway would authenticate the requesting agent, check the current authorisation represented in the user’s Pod, and then permit or reject the request to the relevant external system.

I’m not assuming that Solid currently enforces access control over arbitrary external databases. The gateway connecting the personal authorisation state to external APIs would be an explicit part of the research prototype.

I’m also not claiming that Solid is uniquely capable of this pattern. UMA, OAuth-based approaches and centralised consent/permission registries can address related problems.

What I’m interested in is whether Solid provides a good foundation for a personal, portable and persistent locus of control across multiple services, rather than each service maintaining its own isolated permission state.

The Human-Centred AI study would then compare:

service-scoped permission management

with

a consolidated personal authorisation view

while keeping the information and controls available in both conditions equivalent.

The study would focus on whether people can understand and manage cumulative access over time, for example:

Which AI currently has access to this information?
Why was access granted?
Has it expired?
What has actually been accessed?
Can I revoke one agent without affecting another?

The broader idea I’m exploring is:

decentralising control over AI context without necessarily decentralising the authoritative data itself.

I’d particularly value feedback on:

  1. Does using a Solid Pod as a persistent personal authorisation registry/control layer fit reasonably with Solid’s architecture and philosophy?
  2. Are Access Needs, Access Authorizations, Access Grants and the Authorization Registry in the Solid Application Interoperability work appropriate building blocks for this kind of design?
  3. Is there existing Solid work addressing authorisation of access to data that remains outside the Solid ecosystem?
  4. If the external gateway described above is needed, where would you see the trust boundary and agent-authentication responsibility sitting?
  5. Would representing purpose, expiry, revocation and access history in the Pod be a reasonable pattern, perhaps using something such as DPV for purpose metadata?
  6. If current interoperability implementations are not mature enough for this, would a simpler RDF-based authorisation registry stored in a Pod be a sensible research prototype?
  7. What is the biggest technical or conceptual weakness you see in this architecture?

This is still at proposal stage, so critical feedback is very welcome. I’m particularly interested in identifying where I may be stretching Solid beyond its intended role before committing to the implementation.

Thanks in advance

You may be interested with ODRL V2.2 Implementation Best Practices

This should be implemented in lws specification https://www.w3.org/TR/lws10-core/. This specification shall be in someway’s complementary to Solid.

A CSS PR tries to propose the basis for an implementation of lws. The solid-lws should work with Solid.

An experimental CSS mashlib recipe with solid-lws can be tested locally GitHub - SolidOS/css-mashlib at feat/lws · GitHub. There are 2 config available for solid-lws

LWS is not finalised.

Hi, thanks for that. I wasn’t aware of that, and it does look very interesting.

I’m now reconsidering this idea overall. My initial thinking was just using SOLID pods for managing consent to use your data as context for AI, but having spoken to a few people on this, Im not sure people care enough about their data being used to bother.

I’m not sure people care enough about their data being used to bother

People don’t care enough to get vaccinated, doesn’t mean we should stop promoting vaccinations. People’s ignorance of the dangers of AI and data-hoarding in general is not an excuse to provide them with the means to shoot themselves in the foot.

1 Like

I’ve reshaped my reseach proposal as follows:

Scientific Abstract

AI assistants increasingly maintain persistent personal context to personalise advice, but that context can become outdated, inaccurate or fragmented across services. In health and social services, errors relating to employment, household circumstances, benefits or functional needs could cause relevant supports to be overlooked or inappropriate guidance to be provided.

This research investigates whether shared personal context can help users maintain more accurate and consistent information across multiple AI services. A lightweight prototype will use a SOLID Personal Data Pod to hold personal context together with its source and update history, while simulated authoritative agency records remain separate.

A controlled between-subjects study using fictional health and social-service scenarios will compare two information-equivalent conditions in which personal context is either maintained separately by each AI assistant or shared across assistants through the Pod. Participants will encounter outdated, disputed and AI-inferred information together with a scripted change in circumstances. The primary outcome will be the proportion of context problems correctly resolved across all affected assistants. Secondary measures will examine recommendation appropriateness, perceived agency, usability, effort and whether participants consider cross-service use of particular information contextually appropriate.

Research Questions

RQ1 – Architecture:
How can SOLID support a provenance-aware personal context layer that is reusable across AI services while preserving distinctions between user-provided information, AI-generated inferences and agency-sourced information?

RQ2 – Primary empirical question:
Compared with context maintained separately by individual AI services, does a shared personal context layer improve users’ ability to maintain accurate and consistent information across multiple AI assistants when circumstances change and errors may propagate between services?

RQ3 – Human-centred trade-offs:
How does shared personal context affect perceived agency, users’ mental models of context propagation, usability, maintenance effort and judgements about the appropriateness of cross-service information reuse?

I would very much welcome any feedback (good or bad) on the direction and/or mertic of this reseach -

Thank you

The point about people not caring is fair, but I’d be careful designing around it either way. If assistants are juggling benefits and functional needs, I’d want the study to separate ‘user corrected it’ from ‘AI inferred it’ clearly, otherwise the resolution numbers get murky. For what it’s worth, I read that Apple CEO Ternus pushes for faster, agentic Siri, and the direction toward assistants acting more on their own makes provenance feel less academic and more practical. I’m unsure whether participants will grasp a Pod in a short scenario. Anyone run pilot sessions before the main study?