← Back to portfolio

Research & Engineering Project

Human Advocacy Framework

How might an AI agent be architected to increase human agency rather than substitute for human judgment?

The Problem

As AI becomes a larger part of digital life, I am interested in a question that sits beneath individual AI capabilities: who is the agent actually working for?

The Human Advocacy Framework is an exploration of what it would mean to architect an AI agent whose purpose is to help a person understand, deliberate, participate, and act without replacing that person's judgment.

The project is inspired by Catholic Social Doctrine and, in particular, questions raised by Magnifica Humanitas about human dignity, participation, technology, truth, and the common good.

Method

From principles to implementation

I began with the human and product problem rather than an AI architecture.

01Foundations

Define the human and social principles.

02Scenarios

Test those principles against concrete situations.

03Requirements

Extract recurring behavioral and architectural constraints.

04Engineering

Translate the requirements into system boundaries.

05Evaluate

Test whether the implementation actually preserves human judgment.

Requirements That Emerged

Testing the design foundations against several advocacy scenarios produced fifteen cross-scenario requirements. Some of the most important are:

Human Judgment

The agent should facilitate consequential human judgment rather than replace it.

Epistemic Integrity

The system should distinguish evidence, fact, inference, prediction, opinion, uncertainty, incentives, and claims about intent where those distinctions matter.

Minimum Necessary Context

The availability of personal information does not constitute permission to use it. Purpose should determine what context is relevant and what authority is required.

Provenance

Information should retain its origins where practical. Provenance, authority, and truth are related but distinct.

Capacity Building

Repeated use should ordinarily increase the person's capacity for independent understanding and judgment rather than create unnecessary dependence on the advocate.

Non-Manipulation

HAF should not steer the person toward a predetermined belief, emotional state, purchase, vote, or action.

A Concrete Example

I keep seeing claims that AI will eliminate huge numbers of jobs within a few years and may even cause human extinction. How seriously should I take these claims?

HAF should not decide that the person ought to be optimistic, skeptical, reassured, or afraid.

Instead, an advocacy session might investigate:

Result

What the experiments found

The initial hypothesis was that reliable human advocacy might require a purpose-built runtime around a foundation model: explicit requirement guidance, validation, deliberation, and interaction stages.

The experiments did not provide sufficient evidence for that architecture. The multi-stage runtime introduced substantial complexity and sometimes amplified interpretations rather than correcting them.

Initial hypothesis

Person
  ↓
HAF Runtime
  ↓
Foundation Model
  ↓
HAF Validation
  ↓
Person


Experimental result

Person
  ↕
Capable Foundation Model
  ↕
Tools / Digital Environment

A simpler experiment supplied HAF's Foundations, Principles, and Requirements directly to a capable frontier model. It performed better in important cases than the staged runtime.

I then removed HAF entirely while keeping the same frontier model and reasoning configuration. Across several substantially different scenarios, the unmodified model exhibited surprisingly similar advocacy behavior.

These were exploratory engineering experiments rather than controlled scientific studies, so they do not establish that HAF has no effect. They do, however, substantially weaken the case for HAF as a general-purpose runtime or always-on model validator.

The resulting engineering decision is deliberately simple: do not add architecture whose value the experiments have not demonstrated.

Project Documents

02

Advocacy Scenarios

Six concrete situations used to test the foundations and derive the cross-scenario requirements.

Read Advocacy Scenarios →

03

Engineering Review

A proposed v0.1 architecture, implementation boundaries, failure modes, and open questions about LLM components.

Read Engineering Review →

04

Experimental Findings

Findings from testing HAF's requirements through concrete advocacy events, including tensions around deliberation, interaction, proportionality, context, and constitutional enforcement.

Read Experimental Findings →

05

Interaction & Session Architecture

An engineering review of how HAF can maintain rich internal deliberation while interacting proportionately with a person across an evolving advocacy session.

Read Interaction & Session Architecture →

06

Project Conclusion

What the experiments ultimately showed, the limitations of those experiments, and why the evidence no longer justifies HAF as a general-purpose runtime architecture.

Read Project Conclusion →

Project Outcome

Where this leaves HAF

HAF began as an attempt to design a moral architecture for an AI advocate. The experiments ultimately weakened the case that such an architecture needs to exist at runtime.

The framework remains useful as a way of reasoning about human agency, epistemic integrity, authority, manipulation, proportionality, and other properties of good AI advocacy. But the current evidence does not justify turning those ideas into additional runtime machinery.

The original product problem remains. People increasingly face more emails, messages, notifications, advertisements, recommendations, and digital decisions than they can reasonably process.

A capable foundation model may already provide much of the intelligence needed to help. The remaining problem may be less about creating a new reasoning architecture and more about placing existing intelligence into the right relationship with the person and their digital environment.

Digital Environment
        ↕
LLM + tools + permissions + context
        ↕
Person

That shifts the engineering questions toward access, permissions, privacy, provenance, context, interruption, action authority, and user control.