Wiki · Accountability agenda

What to do — the accountability agenda

The diagnostics (who is not covered, does any of it work) turned into priorities. Across the 23harms, the gap is not more legislation — every one already carries a rule. The field’s first-order needs are 13 harms that need an enforcement mechanism (a binding rule with no catalogued enforcement) and 10 that need an efficacy evaluation (governed and enforced, but no evidence it works). Mean deficit score 3.1. (The efficacy gap is in fact field-wide: every one of the 23 harms lacks rigorous efficacy evidence — the 10 shown are only those where it is the top-priority gap; the rest surface under a more urgent structural gap first.)

Honest scope

This agenda is derived from the catalog's RECORD of governance, enforcement, and efficacy evidence (it inherits the caveats of coverage-equity + accountability-completeness, including that lever counts ignore an instrument's lifecycle status). An 'action-type' names the structural gap a reader should scrutinise — e.g. a binding rule with no catalogued enforcement — NOT a policy recommendation, a legal opinion, or a targeted advocacy ask (charter §7.4/§7.5). 'Needs an efficacy evaluation' means no rigorous evaluation exists, not that the rule provably fails. IMPORTANT: the 'commission an efficacy evaluation' count shows only the harms where evaluation is the TOP-priority gap (they already have both a binding rule and catalogued enforcement); in fact EVERY harm here lacks rigorous efficacy evidence — the rest surface under a higher-priority structural gap (e.g. build enforcement) first. Read efficacy as a field-wide deficit, not a problem for only some harms.

What the field most needs

First-order action	Harms
Build an enforcement mechanism	13
Commission an efficacy evaluation	10

Priorities — highest accountability deficit first

Harm	Deficit	First-order action	Why
Compute + Model-Weight Export Controls	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI in Education	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Sovereign AI Doctrine	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI-Driven Worker Displacement	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Technological Sovereignty	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Environmental Impact of AI Training	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
National Security Carveouts in AI Regulation	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Open-Weight Frontier Release	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Compute-Threshold Reporting	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
International Coordination	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Synthetic Content Provenance	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Agentic AI Governance	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Catastrophic & Existential Risk	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI in Criminal Justice	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
AI in Healthcare	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Biometric Identification	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
AI in Employment	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Development-Rights Framings	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Deepfakes / Synthetic Content	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
Training-Data Rights	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Individual Redress	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Foundation Models / GPAI	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Transparency Obligations	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).

How to read this

This is the “what to do” layer of the harm-accountability mission — the autonomous-safe slice of treating Policy Window as an accountability instrument. It states the structuralgap each harm most needs closed, derived from the catalog’s record of governance, enforcement, and efficacy evidence. It deliberately does notname which actor must act, draft a campaign ask, or give legal advice (charter §7.4/§7.5) — naming the responsible companies (a structured actor catalog) and tracking Policy Window’s own influence on outcomes are the next, governance-reviewed steps. Machine-readable at /wiki/accountability-agenda.json.

What to do — the accountability agenda

First-order action

Harms

Build an enforcement mechanism

Commission an efficacy evaluation

Priorities — highest accountability deficit first

Harm	Deficit	First-order action	Why
Compute + Model-Weight Export Controls	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI in Education	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Sovereign AI Doctrine	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI-Driven Worker Displacement	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Technological Sovereignty	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Environmental Impact of AI Training	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
National Security Carveouts in AI Regulation	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Open-Weight Frontier Release	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Compute-Threshold Reporting	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
International Coordination	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Synthetic Content Provenance	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Agentic AI Governance	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
Catastrophic & Existential Risk	4	Build an enforcement mechanism	A binding rule exists but no enforcement case is catalogued — paper protection.
AI in Criminal Justice	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
AI in Healthcare	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Biometric Identification	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
AI in Employment	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Development-Rights Framings	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Deepfakes / Synthetic Content	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy thin).
Training-Data Rights	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Individual Redress	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Foundation Models / GPAI	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).
Transparency Obligations	2	Commission an efficacy evaluation	Governance and enforcement exist, but no rigorous evidence shows they reduce the harm (efficacy absent).

How to read this