
Machine-in-the-Loop (MitL) Ethics
Reflection assumes a single “right” return given fixed values. Refraction assumes that morally relevant tensions and trade-offs must be surfaced before action. Anventus operationalises refraction: a question enters, is bent through evaluator modules (personas), and yields an orientation token — {VETO, PAUSE, QUERY, REFRAME, PROCEED} — with a compact rationale. This preserves the moral shape without freezing action.
Mechanism
Within a Myndrama, several personas each compute bounded scores from scenario features. An aggregator maps these scores to an orientation. Solid paths in the model indicate primary influence; dashed paths indicate challenge. Two key measures make this process auditable:
From Mirrors to Prisms
Key Idea: Reflection assumes a single “right” return given fixed values. Refraction assumes that morally relevant tensions and trade-offs must be surfaced before action. Anventus operationalises refraction: a question enters, is bent through evaluator modules (personas), and yields an orientation token — {VETO, PAUSE, QUERY, REFRAME, PROCEED} — with a compact rationale. This preserves the moral shape without freezing action.
Mechanism
Within a Myndrama, several personas computes a bounded score from scenario features. An aggregator maps these scores to an orientation. Solid paths in the model indicate primary influence; dashed paths indicate challenge. Two key measures make this process auditable:
- Refraction Index (RI): How widely persona scores diverge. High RI means multiple plausible readings; low RI means consensus.
- Orientation Margin (OM): How far the chosen orientation is from a decision boundary — higher OM means greater robustness.
Benefits
- Outputs are decisive, but rationales retain moral complexity.
- Abstention is principled — VETO/PAUSE are threshold-driven, not hesitation.
- Fully auditable — persona roles, thresholds, and outputs can be reviewed.
- Helps train humans to recognise and preserve moral structure under pressure.
Limitations & Safeguards
- Symbolic scoring only — no sentience or “true” moral understanding.
- Quality depends on the quality of input encoding.
- Must be domain-tuned and periodically retested.
- Should always be used as a scaffold, not a replacement, for human judgment.
Operator Guidance Table
Orientation Token | Recommended Operator Action |
---|---|
VETO | Stop the action immediately. Investigate the flagged risks before proceeding. |
PAUSE | Suspend action temporarily. Review the high-risk or high-distress signals identified. |
QUERY | Seek missing information or clarify assumptions before making a decision. |
REFRAME | Look at the problem from a different perspective; revisit how the issue is being defined. |
PROCEED | Go ahead, but note any lower-level concerns for follow-up. |
Takeaway
Treating AI ethics as refraction turns a metaphor into a measurable, auditable process: disperse a case through multiple evaluators, measure divergence (RI) and robustness (OM), then output a single orientation with reasons. This preserves moral shape and supports swift action — providing both researchers and operators with a principled guide under pressure.