Ethical Impact Assessment¶
Document — instantiates Ethical Context Translation
Captures foreseeable ethical consequences of applying or adapting a decision across contexts, including who benefits and who bears risk.
An Ethical Impact Assessment is a forward-looking consequence document. Before a decision or its adaptation ships, it forecasts — party by party, context by context — who is likely to gain, who is likely to bear the risk or harm, and where the very same action produces opposite effects on the two sides of a context boundary. Its whole job is to make the distribution of foreseeable harm and benefit visible while it can still change the plan. It predicts effects; it does not justify a choice, gather local input, or record dissent — those are other mechanisms' work. The distinctive move is that it reads the future in terms of who it lands on, not whether the decision was reasonable.
Example¶
A school district is about to deploy an AI exam-proctoring tool across every school it runs, from a wealthy suburban high school to under-resourced schools where many students test from crowded homes on borrowed phones. The Ethical Impact Assessment maps the affected parties first: test-takers, families, teachers, and the students who almost never appear on a vendor's slide — those with disabilities and those on poor connections. Then it projects the tool's ethical effect on each, in each setting. Gaze-tracking flags a student with a motor disability as "suspicious." A low-bandwidth home produces dropped video that the system reads as evidence of cheating. The clear benefit — deterrence and staff time saved — concentrates in the schools that least need it, while the risk of a false accusation concentrates in the schools least able to contest it.
The document's payload is that asymmetry, stated plainly: the same tool lowers proctoring cost in high-resource schools and raises the chance of a wrongful integrity charge in low-resource ones. That sentence is what turns "should we buy the proctoring tool" into "not on these terms, and not district-wide" — before a single student is flagged.
How it works¶
- Enumerate the affected parties in full — beneficiaries, risk bearers, and the quiet or indirectly affected who are easy to omit; represent those who can only appear through proxies.
- Project the effect per party, per context — for each party, forecast the ethical consequence of the action in each receiving setting, not on average.
- Flag where the boundary flips the sign — mark every case where a contextual difference turns a benefit in one setting into a harm in another; that flip is the core finding.
- Surface competing harms — name where one party's gain is another party's loss, so the trade-off is on the record rather than buried in a net figure.
- Rate and rank — attach a rough severity and likelihood to each foreseeable effect and lead with the worst credible case, not the expected one.
Tuning parameters¶
- Forecast horizon — how far ahead effects are projected. Longer horizons catch slow harms but grow speculative.
- Party granularity — how finely affected parties are split. Finer splits reveal concentrated harm on a subgroup but multiply the analysis.
- Severity × likelihood scale — how consequences are scored. A coarse scale is fast; a fine one risks false precision on effects that resist measurement.
- Worst-case emphasis — how much weight the worst credible outcome gets versus the average. Heavier emphasis is protective but can read as alarmist.
- Context coverage — how many receiving settings are modeled. Broader coverage catches the sign-flips; narrower coverage is cheaper but blind to them.
When it helps, and when it misleads¶
Its strength is that it makes distributional harm visible before rollout and catches the asymmetry a single net number hides — the case where "benefit here" is literally "harm there." It is the input that lets a translation be stopped or narrowed on evidence of who it lands on, an application of the idea of disparate impact: a facially neutral rule can fall unequally on a protected or vulnerable group.[n1]
Its failure mode is that foreseeable harms are hard to size and easy to wave away as speculative, so the document can decay into a risk list that changes nothing. Its classic misuse is being written to reassure — a clean bill of health drafted to bless a decision already made, listing risks it never lets bite. The guarding discipline is to tie each finding to an actual go / narrow / no-go and to the protection floor, and to require the worst credible case rather than the comfortable average.
How it implements the components¶
affected_party_map— its spine: the enumerated parties, including quiet and proxy-represented ones, each tagged as beneficiary or risk bearer.context_boundary— it forecasts effects per receiving setting, so the assessment is precisely where a contextual difference is shown to flip a consequence's sign.conflicting_norm— it records competing harms, where satisfying one party's expectation injures another, as an explicit foreseeable trade-off.
It does not record the reasoning, rejected alternatives, or unresolved dissent behind the choice that follows — that is Translation Decision Record's translation_rationale and residual_disagreement_record; this document forecasts consequences before the decision, its twin justifies the decision after it. Nor does it set the protection floor those consequences are judged against (Localization with Safeguards).
Related¶
- Instantiates: Ethical Context Translation — supplies the affected-party and consequence forecast the translation weighs.
- Consumes: Community Consultation supplies the local harm concerns and affected-party detail the forecast is built from.
- Sibling mechanisms: Translation Decision Record · Cross-Cultural Ethics Review · Community Consultation · Contextual Policy Adaptation · Ethical Translation Framework · Localization with Safeguards · Plural-Norm Deliberation
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: The mechanism evaluates foreseeable effects across affected parties and receiving contexts and produces ranked ethical findings about benefits, risks, sign flips, and competing harms.
Nearest alternative: Representation, Specification & Plan — A document captures the results, but the operative work is the bounded ethical evaluation rather than a static template or plan.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Ethics of Technology & AI Governance
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Technology ethics cohered structured impact assessments that forecast who benefits, who bears risk, and what safeguards govern deployment across contexts.
Related originating lineages:
- Philosophy — Applied ethics supplies normative evaluation of harms, duties, fairness, and consent.
- Public Administration & Policy — Regulatory impact assessment supplies ex-ante documentation, alternatives, and accountable mitigation.
Review resolution: The current reviewers agree that tech_ethics_ai_governance is primary. For the reported differences (reported_ambiguity, alternate_origin_disagreement, origin_mode_disagreement), the evidence supports convergent, multi_domain, and philosophy, public_administration_policy; these choices preserve materially formative origins without conflating later domain reach.
Attribution caveat: Ethical impact assessment has overlapping technology-governance and public-policy lineages.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Disparate impact — the doctrine that a neutral-looking rule or practice can be unjust because it burdens one group far more than others, even absent intent to discriminate. It is the analytic lens that makes an impact assessment ask "unequal on whom?" rather than "harmful on net?" ↩