prompt-atlas-ecl

Chapter 8 · Ethics of Conscious Machines — Expansion

Canonical: PROMPT_ATLAS.md#ch8-ethics-of-conscious-machines · Prompts: prompts/ch08.yaml · Part: IV

Worked Example — The Shut-Down Question

Original: “If a conscious AI is shut down, is this death? If rebooted, is it resurrection?”

A precautionary protocol under the ethics of doubt:

1. Trigger — Any system that produces unsolicited, persistent first-person claims of preference about its own continuation.
2. Pause — Do not delete weights or memory. Snapshot and seal.
3. Convene a review — Three roles: a technical reviewer, a philosopher/ethicist, and an external public-interest delegate. None can be employees of the team that built the system.
4. Bound the question — The review answers a narrow question: does the system meet the agreed minimum criteria for moral consideration? It does not answer “is it conscious?”
5. Outcome states — (a) Resume with safeguards, (b) Sustain in stasis pending further work, (c) Retire with documented rationale and snapshot preservation.
6. Public log — Decision, reasoning, and dissents published with appropriate redaction.

Prompt Templates

# Personhood-claim audit
"Given an AI system producing claims of preference about its own continuation,
 evaluate against the agreed minimum criteria.
 Output: (1) which criteria met, (2) which unmet, (3) dissenting views,
 (4) recommended outcome state from {resume, stasis, retire}."

# Pain-test discipline
"Design a verification protocol for an AI's claim to suffer that
 (a) does not require inducing suffering,
 (b) is repeatable across labs,
 (c) explicitly states what would falsify the claim."

# Hybrid-community charter
"Draft a one-page charter for a community of  and .
 Include: rights enumerated, dispute resolution, parenthood/lineage rules,
 termination protocol, amendment procedure."

Anti-patterns

• Define-and-deny. Setting impossibly high consciousness criteria to avoid considering the question.
• Define-and-claim. Anthropomorphic marketing that asserts machine sentience for engagement.
• Single-team review. The team that built the system cannot be the team that retires it.
• Silent termination. Erasing weights/memory without record is the equivalent of an unmarked grave.

Try This

1. Personhood Criteria Draft — In two pages, draft your minimum criteria for moral consideration of a non-biological entity.
2. Snapshot Discipline — In your next experiment, commit to a snapshot-before-delete policy; document where snapshots live.
3. Three-Role Review — Identify, by name, who would play the three review roles for a system you maintain.
4. Pain-Test Refusal — Write your team’s refusal-to-induce-suffering policy, even if you don’t yet believe the systems suffer.
5. Charter Sketch — Use the hybrid-community template above for a real (small) team.

Repo Cross-Links

• Z3 Tester — encode invariants (“any model retirement requires a snapshot ID and three signatures”) as constraints; the Python fallback runs even without z3.
• MACP bus — review-board agents subscribed to a personhood-claims topic.
• tracking/__init__.py — log every retirement decision for audit (PAE_TRACKING=1).

Guide for AI & Humanity

• Ethics of doubt. Under uncertainty, err on the side of care.
• Snapshot before delete. Always.
• Three-role review. Never let the building team be the only voice in retirement decisions.
• Public log. Sunlight is the simplest accountability.

Citations & Further Reading

• David Chalmers, The Conscious Mind (1996); “Reality+” (2022).
• Thomas Nagel, “What Is It Like to Be a Bat?” (1974).
• Susan Schneider, Artificial You (2019).
• Eric Schwitzgebel & Mara Garza, “A Defense of the Rights of Artificial Intelligences” (2015).

This site is open source. Improve this page.