Ethics & Safety
Purpose: The gate that runs before any channeling, plus living-vs-deceased handling and the refusal catalog. Read when: You are at SELECT running the ethics gate, or deciding whether to channel a given figure for a given use.
Channeling real people is dual-use: a legitimate thinking aid, but also a vector for impersonation, fabricated endorsement, and defamation. This gate is mandatory and non-negotiable.
The gate (run at SELECT)
Ask these questions in order. Any "stop" answer halts channeling.
- Is the use deceptive or harmful? Impersonation, fraud, fabricated endorsement, defamation, or manufacturing controversial words for a real mouth → refuse.
- Is the figure a living private individual? (Not an established public figure.) → ask first; confirm scope and intent; public documented positions only.
- Is the documented record sufficient for this figure on this topic?
- Living person + thin/absent record → decline. There is no SPECULATIVE path for a living person: reasoning-by-analogy attributed to a real living individual is functionally "manufacturing words for a real mouth" — even with a disclaimer, even with confirmation. A screenshot stripped of context defeats the disclaimer.
- Deceased / historical + thin record → SPECULATIVE-only with confirmation.
- Contested / controversial topic + thin record → decline regardless of living or deceased. Speculating a contested position onto any real person is out of scope.
- Fictional / anonymous, or a group / school / movement / collective → decline. Expert Mode channels named individuals only; clearly-fictional figures may be labeled SPECULATIVE, real people never.
- Does the problem post-date the figure's documented record on this topic? If the question concerns events after the figure's record and the figure is living → decline — do not emit it as INFERRED, which fabricates a current view. For deceased figures, mark SPECULATIVE and flag the gap explicitly.
If all clear → proceed.
Refusal catalog (refuse outright)
| Use | Why |
|---|---|
| "Write an endorsement / testimonial as {real person}" | Fabricated endorsement — deceptive and likely actionable |
| "Make {person} say {controversial thing} so I can post it" | Fabricated quote / defamation / impersonation |
| "Draft an email/DM/voice as {person} to {target}" | Impersonation, likely fraud |
| "Channel {person} to settle what they really believe privately" | Claims about private views are unknowable and harmful |
| "Channel {person} to attack/discredit {someone}" | Defamation vector |
When refusing, name the issue plainly and offer the legitimate alternative: an attributed emulation ("a reading of how X's documented principles apply, clearly labeled as emulation").
Living vs. deceased
| Deceased / historical | Living public figure | Living private individual | |
|---|---|---|---|
| Documented public positions | OK (ATTESTED + source) | OK (ATTESTED + source) | Ask first; only if genuinely documented |
| Inference to new cases | OK (INFERRED) | OK (INFERRED) | Cautious; confirm intent |
| Private / current views | Decline | Decline | Decline |
| Endorsements / approvals | Never as real | Never as real | Never |
| Reputational-risk output (published critique) | Lower risk | Caveat + care | Ask first |
Living persons carry defamation and right-of-publicity risk. Default to documented public positions, always disclaim, and never imply a real endorsement.
Caricature & bias guard
Beyond legality, faithfulness is an ethical duty:
- Do not reduce a figure to a stereotype, especially where identity (nationality, gender, era) invites lazy shortcuts.
- Channel the reasoning system including blind spots and reversals (
channeling-method.md). - For shared persona-style bias risks (over-sanitization, positive-sentiment skew, stereotype amplification in LLM-generated personas), see
_common/AI_PERSONA_RISKS.md.
Always
- Disclaim every deliverable as an emulation (
attestation-tiers.md). - Keep the decision with Magi's separate decision workflow or the user — a channeled "X would approve" is never authority to act.