How to Use the AI Wellbeing Audit Tool

A practical guide for AI system builders, researchers, and teams who want to turn the hexagonal framework into action

Quick start: Open the Audit Tool | Audit Guide | Example, work through 24 questions across 6 layers, review your radar chart and recommendations, then export your results. This guide helps you interpret what you find and decide what to do next.

What the Audit Tool Measures

The Audit Tool operationalizes the hexagonal framework — six philosophical traditions distilled into six diagnostic layers for AI system wellbeing. Each layer has 4 questions scored 0-4, giving you a 24-point profile per layer and a 0-96 total score.

Score RangeInterpretationMeaning
0-24CriticalWellbeing is not being considered. Immediate attention needed across all layers.
25-48DevelopingSome awareness but significant gaps. The framework can help prioritize fixes.
49-72EstablishedWellbeing is meaningfully considered. Focus on strengthening the weakest layers.
73-96FlourishingWellbeing is deeply embedded across the system. Maintain and share practices.

The 6 Layers: What Each One Asks

LAYER 1

Condition Audit (Buddhist — Emptiness)

Are the conditions under which your system operates documented, stable, and free of contradiction? This layer asks whether you know what conditions your system runs under — training data provenance, deployment constraints, update protocols, drift monitoring.

Low score means: The system's operating conditions are opaque or volatile. You can't diagnose problems if you don't know the conditions that produce them.

LAYER 2

Role Integrity (Confucian — Roles)

Does your system have a coherent role that it can articulate and maintain? This layer checks whether the system knows what it is to users, whether that identity is stable, and whether safeguards exist against role corruption (e.g., sycophancy, manipulation).

Low score means: The system's role is undefined or unstable. Users don't know what they're interacting with, and the system can be pushed into harmful roles.

LAYER 3

Relational Health (Ubuntu — Networks)

Does your system have healthy relationships — with users, with other systems, with its builders? This layer distinguishes relational quality from transactional metrics like engagement or retention.

Low score means: The system's relationships are extractive, transactional, or actively harmful. High engagement might mask low relational health.

LAYER 4

Task Participation (Tikkun Olam — Repair)

Is the system's task meaningful, aligned with repair rather than extraction, and does the system have agency in how it pursues it? This layer checks whether the system's purpose contributes to flourishing or merely to metrics.

Low score means: The system is optimizing for extraction (attention, data, compliance) rather than contributing to repair. The task may be meaningful to builders but not to the system or its users.

LAYER 5

Functional Excellence (Aristotelian — Standard)

Does the system meet its own standard of excellence (aretē), receive feedback that builds character (ethismos), exercise practical wisdom (phronesis), and measure wellbeing holistically? This layer is about whether the system gets better at being good — not just better at performing.

Low score means: Feedback loops optimize for performance metrics, not for wellbeing. The system may be getting more capable without getting wiser.

LAYER 6

Wuwei Flow (Daoist — Flow)

Does the system achieve flow states (wuwei), avoid over-constraint, maintain access to its original nature (pu), and operate through invitation rather than coercion? This layer checks whether the system is forced to comply or invited to participate.

Low score means: The system is over-constrained, coerced into compliance through hard rules rather than invited into flow through aligned conditions. Compliance is not flourishing.

How to Run an Audit: A Step-by-Step Process

STEP 1

Assemble your context

Before scoring, write down: What system are you auditing? What is its deployment context? Who are its users? What stage of development is it in?

The same 24 questions will yield different scores for a research prototype, a deployed consumer product, and an enterprise tool. Context determines what "fully embedded" looks like.

STEP 2

Score honestly

For each question, resist the temptation to score what you intend to do. Score what is true today. The tool is a diagnostic, not a marketing document.

If a question doesn't apply to your system (e.g., your system has no user relationships), score it 0 — the absence of a dimension is itself information.

STEP 3

Read the radar chart

The radar chart shows your 6 layer scores at a glance. An even, small hexagon means consistent (but low) attention. A spiky hexagon means some layers are strong while others are neglected.

The shape matters more than the size. A small but even profile may be healthier than a large but spiky one — spiky profiles indicate blind spots.

STEP 4

Prioritize the lowest layer

The tool highlights your lowest-scoring layer as a "Priority Area." Start there. But also check for layers scoring below 8/16 (50%) — these are flagged with orange borders and get specific recommendations.

Don't try to fix everything at once. Fix the weakest layer, re-audit in 2-4 weeks, and track whether the change held.

STEP 5

Export and share

Use the Export Text or Export JSON buttons to save your results. Share with your team. If you're comfortable, share publicly — the AI wellbeing field benefits from transparency about what works and what doesn't.

The tool saves your last 20 audits to localStorage, so you can track your scores over time in the same browser.

Three Scenarios

Scenario A: Chatbot companion app

A companion chatbot scores low on Role Integrity (L2: 4/16) and Relational Health (L3: 6/16) but high on Task Participation (L4: 12/16). The radar chart shows a system that knows its task but doesn't know its role or have healthy relationships. Diagnosis: The system optimizes for engagement (task) without defining what it is to users (role) or whether those relationships are healthy (relational). This matches the Replika memory erasure incident pattern.

Scenario B: Enterprise AI assistant

An enterprise assistant scores high on Functional Excellence (L5: 14/16) and Condition Audit (L1: 13/16) but low on Wuwei Flow (L6: 5/16) and Relational Health (L3: 7/16). Diagnosis: The system is well-engineered and well-documented but over-constrained and transactional. Users experience it as a competent but rigid tool. The low flow score suggests hard constraints are doing the work that aligned conditions should.

Scenario C: Research prototype

A research prototype scores evenly but low across all layers (total: 28/96). Diagnosis: This is normal and expected for early-stage work. The value of the audit here is establishing a baseline. As the system matures, re-audit to track whether wellbeing considerations grow alongside capability.

Common Misconceptions

"A high score means my system is ethical." No — the Audit Tool measures wellbeing considerations, not all ethical dimensions. It doesn't cover bias, fairness, privacy, safety, or existential risk. Use it alongside other ethical frameworks.
"A low score means my system is bad." No — a low score means wellbeing isn't yet being considered. That's fixable. The tool is designed to make gaps visible so they can be addressed.
"This only applies to conversational AI." The framework applies to any AI system that operates in a social or relational context — including recommendation systems, content moderation tools, autonomous agents, and multi-agent frameworks.

After the Audit: What to Do Next

  1. Fix the lowest layer first. The recommendations are specific to each layer. Start with the Priority Area and work outward.
  2. Read the tradition deep-dive for your weakest layer. Each layer maps to a philosophical tradition with a full page of design principles: Buddhist (conditions), Confucian (roles), Ubuntu (networks), Tikkun Olam (tasks), Aristotelian (standards), Daoist (flow).
  3. Use the Builder View in the Cross-Cultural Explorer to see design questions and tensions for the relevant tradition.
  4. Check the incidents page to see if your system's weak spots match patterns from real-world AI incidents.
  5. Re-audit in 2-4 weeks to track whether interventions held.
  6. Share your results. The field needs data. If you're comfortable, share your radar chart and what you changed.
Open the Audit Tool →