A diagnostic practice for AI tool deployment

AI HARM DIAGNOSIS

A retrospective tool for analyzing an AI harm that has already occurred. Work through each step — your diagnosis builds on the right as you go.

A note on how this tool works: This interactive tool was developed to run entirely in your browser. It makes no API calls, uses no generative AI, and nothing you enter is sent or collected. Your progress is saved only on your device, so you can pick up where you left off. It works best on a desktop browser, not mobile.

Before you begin

Read first

This framework names systems of oppression — racism, racial capitalism, settler colonialism, patriarchy, transphobia, homophobia, ableism, and their intersections — as the conditions in which AI harms occur, not extra factors layered onto neutral deployments. When harm originates in existing inequities, technical fixes alone cannot repair it. Throughout, keep naming the structural and historical conditions the technology entered, encoded, and amplified.

Intended purposeWhat was the technology designed to do?

Actual outcomeWhat did it actually end up doing?

Harmed partiesIndividuals, communities, institutions, ecosystems — be specific

Harm typesSelect all that apply — some harms span several categories

Scope to the harm this particular deployment caused to the parties identified above. If an upstream/embedded AI harm (e.g.,scraped training data, energy and water costs, land theft, etc.) doesn't help explain the harm, it likely deserves its own diagnosis. Test: if that upstream harm hadn't occurred, would these people still have been harmed this way?

Severity is never averaged. Intensity and Equity set the level on their own; Scale and Consent set it under conditions; Reversibility raises it. The full profile travels with the overall level.

Technical system layer

Most AI harms at the Technical System Layer originate in choices about data, model, objective function, architecture, evaluation, and deployment context. Begin with “Harm by design” and require an affirmative test before assigning anything else.

Human & institutional layer

Most AI harms at the Human & Institutional Layer originate in choices about risk management, necessity, training, and policies. Begin with “Governance failure” and require an affirmative test before assigning anything else.

Responsible partiesSelect all that apply — a case usually implicates several

Basis for responsibility Optional — foreseeability, duty of care, structural position (intent is the weakest test)

6a · Repair to technology & governance — pick all that apply

6b · Repair owed to the people harmed

Affected parties define what repair requires — co-defined with the people harmed, not designed for them. Surface their analysis first, before any menu of options.

6c · Repair to structural conditions

History testDoes this use case sit inside histories of systemic injustice?

Subtraction testCould the same harm have happened without AI? If yes, the injustice was already operating.

Structural repair may include

Leverage heuristic ≈ proximity to harm × structural power × political will to act. The highest-leverage actor isn't always the one with the most formal authority.

Potential intervenersSelect all that could prevent, detect, stop, correct, or repair

AI HARM DIAGNOSIS

AI HARM DIAGNOSIS

Before you begin

What happened?

Who was harmed & how?

How severe was it?

What caused the harm?

Who is responsible?

What would repair require?

Who can intervene?