Skip to content

Toolkit — Abolition check

Gate: Tier 0. Category: refusal artefact.

What problem it solves

The technochauvinism check asks should a computer decide this? The abolition check asks the prior question: should this decision be made at all? Some engagements propose to improve, automate, or scale a process whose existence is the harm. No efficiency gain on such a process is a benefit — it is an amplification. The abolition check forces one explicit pass at is the thing itself the problem? before G1 assumes the thing is worth supporting.

How it is used

A 20-minute Tier 0 conversation, after or alongside the technochauvinism check. The chair names the process the engagement wants to improve and walks three prompts: who is this for? / what would its abolition look like? / what argument would its abolition have to overcome? If the abolition case is stronger than the engagement case, the engagement is refused at Tier 0. If the process survives scrutiny, the engagement proceeds with the abolition case recorded as a dissenting view the engagement has to keep answering.

Inputs

  • The engagement brief, especially the description of what process the engagement touches.
  • At least one participant outside the process's sponsoring function — someone who has been on the receiving end of the process, a critic, or a stakeholder who has previously proposed its removal.
  • The engagement's stated outcomes.

Outputs

  • If refused: a Tier 0 abolition memo — what was proposed, what the abolition case is, why the abolition case wins.
  • If cleared: a dissenting-view record attached to the G1 framed problem — the abolition argument is kept on file and re-examined at G3 and the G5 sunset review.

Visualisation

Abolition check — three prompts and two outputs; process exists and process does not The process as proposed Prompt 1 Who is this process for? Prompt 2 What does abolition look like? Prompt 3 What must abolition overcome? REFUSE Abolition memo CLEAR to G1 Dissent recorded, re-examined at G3, G5 abolition wins process stands

The process is probed by three prompts; the chair's read produces either a crossed-out (refuse) or cleared (proceed, dissent recorded) output.

Anatomy

Prompt 1 — Who is this process for? Often the first answer is for the organisation. The second-pass answer names specific people; the gap between them is where abolition cases live. A process that exists for the organisation but costs its subjects is a candidate for abolition, not improvement.

Prompt 2 — What would abolition look like? A concrete description, not an argument. "We stop doing tenant screenings and accept applicants on rent-payment history alone" is a description; "abolition is radical" is not. Descriptions invite evaluation; labels block it.

Prompt 3 — What would abolition have to overcome? The strongest legitimate arguments for keeping the process. Often these are real (regulatory, safety, fiduciary). Sometimes they collapse under scrutiny. The chair's job is to hear the arguments and assess whether they are load-bearing or ceremonial.

Chair's read. If abolition is stronger, the engagement is refused and the memo explains why. If the process survives, the dissenting view is recorded — not dismissed — and returned to at G3 (routing) and G5 (sunset review).

Example

Paper trail — refusing an AI-for-shoplifter-prediction engagement

Tier 0 session, 25 minutes. Sponsor: a retailer's loss-prevention lead. Outside participant: a retail-worker representative. Chair: Ada.

T+0 — proposal. Sponsor: "We want an AI system that predicts shoplifting risk at the door, so staff can approach likely shoplifters."

T+5 — Prompt 1. Ada: "who is this process for?" Sponsor: "for the store — it reduces shrinkage." Worker rep: "it's not really for the store. Shrinkage is 2% of revenue; staff-customer conflict on false positives is the cost. It's for the loss-prevention function's quarterly numbers." The gap is open.

T+10 — Prompt 2. Ada: "what would abolition look like?" Worker rep: "we stop profiling at the door. Shoplifting gets handled the way it was before — item-by-item at checkout, investigated after the fact." Sponsor: "shrinkage goes up." Worker rep: "by an amount you already budget for."

T+15 — Prompt 3. Ada: "what would abolition have to overcome?" Sponsor: "shareholder pressure on shrinkage numbers." Worker rep: "that's a shareholder-conversation argument, not an AI argument." Sponsor acknowledges.

T+20 — chair's read. Ada: "the AI is for the quarterly number. The cost — profiling false positives, staff-customer conflict, legal exposure under disparate-impact law — is borne by customers and frontline workers. The abolition case is stronger."

T+25 — memo drafted. One page: proposal refused at Tier 0. The engagement would amplify an existing process whose distributional cost outweighs its stated benefit. Recommended next step: shareholder-facing review of shrinkage target, not AI engagement.

Six months later. Sponsor returns with a narrower engagement — shrinkage analytics to inform store-layout changes, no per-customer prediction. Tier 0 is re-run; the narrower framing clears. The first refusal's memo is attached to the new engagement's provenance.

Pitfalls

Strawman abolition. Stating abolition in its weakest form and knocking it down. Prompt 2 requires the strongest description of abolition, not the sponsor's preferred version.

Abolition as universal. The check is not should anything like this ever exist? It is should this engagement do this now? Abolition recommendations scale to the engagement, not to the world.

No outside voice. Running the check with only the sponsoring function in the room produces a foregone conclusion. The outside voice is the check's anchor. If none is available, the check is postponed, not performed alone.

Treating dissent as resolved. When the engagement clears, the abolition argument goes on file. If G3 or G5 never re-examines it, the check has functioned as ceremony. The dissenting view must be live — re-read, not filed.

When not to use

  • Engagements whose process is regulatorily required and whose abolition is foreclosed in law. The check is thin; the dissenting view may still be worth recording for the law's direction of change.
  • Engagements that are themselves abolition projects (the engagement is removing a process). The check's work is already done by the engagement's framing.

Provenance

Abolition as a methodological stance — not simply a political one — is treated in Davis's Are Prisons Obsolete? [1] and formalised in Gilmore's writing on abolition geography [2]. Within AI specifically, Benjamin's Race After Technology [3] argues for a default abolition stance toward systems whose upstream processes encode bias. The pattern of refusing engagements on process-existence grounds — rather than improving the process — is recurring in the ethics-of-AI literature.

  • Technochauvinism check. The paired Tier 0 artefact — not should this decision be made but should a computer decide it?
  • Framed problem paragraph (G1). The G1 artefact inherits the dissenting view.

Verification

[1] Davis AY. Are Prisons Obsolete? Seven Stories Press; 2003. [verified] Treats abolition as a methodological stance distinct from reform.

[2] Gilmore RW. Abolition Geography: Essays Towards Liberation. Verso; 2022. [verified] Formalises abolition as an analytic frame for institutional review.

[3] Benjamin R. Race After Technology: Abolitionist Tools for the New Jim Code. Polity; 2019. [verified] Argues for default-abolition review of AI systems whose upstream processes encode bias.