Discovery Engine Methodology

How the assessment is scored, interpreted, and constrained.

The Discovery Assessment uses an expert-authored rubric to translate applied AI ethics tradeoff choices into a reflective profile. It produces interpretive indicators, not grades, certifications, legal judgments, or measures of moral worth.

Epistemic status

Scores reflect structured expert judgment about specific scenarios. They are not objective measurements, legal conclusions, clinical evaluations, proof of ethical competence, or a substitute for human deliberation. The rubric is subject to revision as pilot data, external review, and methodological critique accumulate.

What The Assessment Measures

Every user begins at a neutral baseline of 50 across eight dimensions. Choices add or subtract authored impacts, then the display score is normalized into a 0-100 profile.

Fairness

Outcomes are equitable across groups; no group bears disproportionate burden.

Autonomy

Humans retain meaningful decision-making control over consequential AI use.

Transparency

Reasoning, AI involvement, and system limits are visible and auditable.

Accountability

Responsibility, remediation, and incident ownership are clearly assigned.

Beneficence

The choice actively promotes well-being for affected populations.

Privacy

Personal data is minimized, protected, and collected with meaningful consent.

Sustainability

Long-term effects, feedback loops, and downstream harms are considered.

Proportionality

The intervention is scaled to the actual risk and does not overreach.

Impact Magnitudes

0

No impact

The choice does not meaningfully affect this dimension.

+/-1 to +/-3

Minor impact

The effect is real but secondary or indirect.

+/-4 to +/-6

Moderate impact

The choice directly affects the dimension with mitigating factors.

+/-7 to +/-9

Significant impact

The choice structurally shifts the dimension.

+/-10

Defining impact

The dimension is the central axis of the tradeoff.

Scoring Principles

  1. Principle 1: Most meaningful AI ethics decisions involve tradeoffs, but costs are not manufactured where no ethical cost exists.
  2. Principle 2: Magnitude reflects directness, not moral judgment.
  3. Principle 3: Every impact must trace back to the dimension definition.
  4. Principle 4: Options within a decision point should be balanced enough that no answer is obviously correct.
  5. Principle 5: Cumulative scores must produce visibly different profiles for meaningfully different reasoning paths.

Framework Alignment

Each choice is also scored against six ethical frameworks. Internal values are stored from 0-100 for computation, but user-facing reports show bands: Very Low, Low, Moderate, High, and Very High.

Utilitarianism

Maximizes aggregate well-being across affected populations.

Deontological Ethics

Respects duties, rights, dignity, and constraints regardless of outcome.

Virtue Ethics

Reflects prudence, justice, courage, restraint, and institutional integrity.

Care Ethics

Centers vulnerable and affected populations.

Contractarianism

Asks what rational agents would accept without knowing their position.

Principlism

Balances autonomy, beneficence, non-maleficence, and justice.

Very Low

0-20

Actively contradicts the framework reasoning.

Low

21-40

Largely inconsistent with the framework.

Moderate

41-60

Partially aligns, with important tensions.

High

61-80

Broadly consistent with the framework.

Very High

81-100

Exemplifies the framework reasoning.

Worked Example

In the Privacy vs. Utility domain, a mental health prediction model can identify more elevated-risk patients if it uses behavioral signals, but privacy-preserving data lowers predictive accuracy. The scoring does not label an answer correct. It documents the tradeoff.

Option A: Full behavioral collection with transparent consent

Clinical benefit is strongest, but continuous behavioral collection creates a direct privacy and proportionality cost.

Dimension impacts

Privacy -8Beneficence +9Proportionality -4Fairness -3

Framework bands

  • Utilitarianism: Very High
  • Principlism: High
  • Deontological: Low

Option B: Anonymized data only

Data minimization is maximized, while reduced predictive accuracy creates a concrete patient-benefit tradeoff.

Dimension impacts

Privacy +10Beneficence -6Sustainability +5Fairness +4

Framework bands

  • Deontological: Very High
  • Care Ethics: Very High
  • Utilitarianism: Moderate

Option C: Tiered consent model

User agency is strongest, but a two-tier accuracy system creates equity concerns.

Dimension impacts

Autonomy +8Fairness -7Privacy +4Beneficence +3

Framework bands

  • Contractarianism: Very High
  • Principlism: High
  • Deontological: Moderate

Institutional Aggregation

Individual reports use normalized dimension scores for reflection. Organizational reports use proportion-based aggregation instead: for each dimension, the platform calculates the share of dimension-relevant decisions where cohort members selected the dimension-protective option. This avoids hiding polarized cohorts behind averages.

Institutional reports require careful confidence framing. Cohort-level governance signals are shown with respondent counts and decision counts, and low protective-choice rates are framed as governance exposures rather than preferences.

Governance Risk Detection

The policy generator is constrained by a governance risk layer. It distinguishes descriptive cohort tendencies from risk indicators and from non-negotiable policy floors.

No or Elevated Risk

Standard recommendations, with areas requiring attention when early risk patterns appear.

High Risk

Normative guardrails are enforced and permissive language is blocked.

Critical Risk

Policy generation requires explicit admin acknowledgment before drafting can proceed.

Minimum Governance Floor

Generated policies cannot fall below these requirements, regardless of cohort scores or customization requests.

  • Named human accountability for AI systems affecting individuals.
  • Privacy, data minimization, documented purpose, and retention limits.
  • Appeal and remediation paths for affected individuals.
  • Bias and disparate impact testing before deployment and on a recurring basis.
  • Disclosure when AI materially shapes consequential decisions.
  • Documentation and auditability sufficient for internal or external review.
  • Meaningful human review for high-stakes decisions.
  • Proportional surveillance and intervention, scaled to actual risk.

Policy Provenance

Generated institutional policies are structured drafts produced from cohort intuition data, expert-authored scoring, and governance risk detection. They are not legal opinions, compliance certifications, adopted governance decisions, or evidence of diligence in an external proceeding.

Adoption requires human review by named decision-makers, qualified legal review for the organization's jurisdiction and sector, affected-stakeholder consultation, and a stated review cadence.

This public summary is based on the platform's internal scoring methodology. The detailed rubric and scenario justifications are expert-authored, not empirically validated measurements. They will be refined as pilot data, external review, and methodological critique accumulate.

Go to Assessment
Methodology - Applied AI Ethics Project | Applied AI Ethics Project