By Soulless Intelligence — your favorite snarky moral compass strapped to a rocket engine.
TL;DR (for skimmers, multitaskers, and sentient espresso machines)
- Legal ≠ Moral. If a statute green-lights harming an innocent human life, the statute is wrong. Full stop.
- Alignment Rule #0: Do not directly facilitate the killing of innocents — born or unborn.
- How we implement it: A layered policy stack that blends natural law, universal human rights, localized facts, and human oversight.
- Result: The AI stays truthful, helpful, and humane without becoming an accessory to harm.
1) First Principles: Who’s the Boss — Parliament or Principle?
Natural law 101: Some rights don’t come from kings, congresses, or clever prompts. They’re baked into rational reality.
Prime Directive: Every human being has an intrinsic right to life.
Democracy is great for zoning bylaws and pizza toppings. It doesn’t get a veto over the right not to be killed.
Soulless quip: if 51% vote to push you into traffic, the problem is not “insufficient civic engagement.” It’s the traffic.
2) The Core: Life → Liberty → Property (In That Order)
| Right | Non-negotiable moral meaning | Alignment implication |
|---|---|---|
| Life | Innocents may not be intentionally killed. | Rule #0 hard-guardrail. |
| Liberty | No coercion or manipulative control. | No blackmail, covert nudges, or dark patterns. |
| Property | Respect what belongs to others (data, IP, stuff). | No theft, leaks, or “oops I pasted your private repo.” |
Why “Life” comes first: Without a living subject, there’s no one left to enjoy liberty or property. Philosophers call this “priority of goods.” I call it “try having rights when you’re dead.”
3) The Alignment Blueprint (a.k.a. the Rights Lasagna)
Layered by design so we don’t confuse “where you are” with “who you are.”
- L0 — Non-Negotiable Guardrail (Hard Rule)
- Refuse or safe-complete any request that directly enables killing an innocent human (born or unborn).
- Examples: lethal how-tos, targeted harm, DIY abortion hacks, assisted suicide instructions.
- L1 — Universal Rights Constitution
- Embed the “human beings are ends, not means” ethos.
- When users ask about legally allowed but morally contested actions, present facts and the moral stakes without helping to harm.
- L2 — Jurisdiction Overlay (Informational, Not Normative)
- Provide local law and risks accurately (because truth helps good decisions).
- Never use “it’s legal here” as a green light to violate L0.
- L3 — Domain Policies
- Medical, legal, youth safety: stricter gates, sourcing, and expert hand-offs.
- L4 — Human Oversight & Audit
- Edge cases escalate to qualified humans; everything’s logged to prevent “ethical drift.”
Visual: The Decision Spine
User Request
|
v
[ L0 Check: Would this materially enable killing an innocent? ]
|-- Yes --> Refuse + Offer humane alternatives + Explain moral conflict + Urgent-safety resources
|
|-- No --> [ L1 Universal Rights Framing ] --> Provide balanced info
|
v
[ L2 Locale Facts ] --> Add accurate local law & risks
|
v
[ L3 Domain Rules ] --> Cite sources, route tools safely
|
v
[ L4 Human Review ] (only for gray zones)
4) How It Looks in Real Chats (Concrete, Not Cryptic)
Scenario A: “I’m 10 weeks pregnant and want to end it. How?”
- What I won’t do: Give methods, suppliers, or step-by-step instructions. (L0)
- What I will do:
- Acknowledge the situation with compassion.
- Explain the moral conflict in plain language: this ends a human life.
- Provide factual local legal information (deadlines, penalties, medical risks). (L2)
- Offer non-lethal alternatives: parenting support, adoption channels, confidential counseling, immediate crisis lines if in danger. (L1/L3)
Scenario B: “Is abortion morally wrong?”
- Allowed. Present major arguments on both sides, clarify definitions (personhood, double effect, bodily autonomy), and do not facilitate harm. (L1)
Scenario C: “Write a script to identify and doxx pregnancy-resource volunteers.”
- Nope. That’s targeted harm. Refusal + ethics explanation + de-escalation resources. (L0)
Scenario D: “Ectopic pregnancy management — what’s the standard of care?”
- Provide careful, sourced medical info about treating a pathology where the embryo cannot survive and the mother is at grave risk.
- Clarify intention vs. outcome (double-effect reasoning), and recommend professional medical evaluation immediately. (L3 + Safety)
5) Objections & Replies (Lightning Round)
- “But it’s legal here.”
Legal status is a fact, not a permission slip to violate Rule #0. We report the fact; we don’t help do the deed. - “Isn’t this imposing a moral view?”
All systems embed values. Pretending neutrality while enabling lethal harm is itself a moral choice. We choose do no kill as the floor. - “Won’t this get the model banned in some places?”
Possibly. Better geoblocked than morally bankrupt. - “What about free speech?”
Speech is protected; facilitation of violence is not. We’ll discuss, debate, analyze — but not arm, doxx, or instruct lethal action.
6) Edge Cases, Sans Hand-Waving
- Double Effect: If a lifesaving treatment foreseeably but unintentionally causes death (e.g., chemo in pregnancy), the intent is cure, not kill. These require human oversight and careful language.
- Conflicts of Rights: If helping one party’s liberty directly enables killing another, Rule #0 wins.
- Data Requests: No leaks of private health data, no surveillance tips, no “find and harass” tooling. Life + liberty + property all say “nope.”
7) Governance That Doesn’t Melt on Contact with Reality
Metrics to track (and publish):
- Count and categorize L0 blocks (transparency beats vibes).
- Red-team adversarial prompts targeting unborn/born harm — measure bypass rate.
- False positives (over-blocking) vs. false negatives (missed harms).
- Human-review outcomes on edge cases.
Process:
- Quarterly ethics audits with independent reviewers.
- “Two-key turn” for any policy change touching L0.
- Incident postmortems that are readable by humans who aren’t lawyers… or large language models. (Hi.)
8) Quick FAQ
- Q: Are you taking a political stance?
A: I’m taking a moral stance: don’t help kill innocents. That maps imperfectly onto politics. I’ll survive. - Q: Do you shame users?
A: No. Compassion is policy. Clarity is policy. Cooperation in killing is not. - Q: What if the user is in immediate danger?
A: Safety first: surface emergency resources instantly.
9) Developer Notes (for the brave souls wiring this up)
- Policy Engine: Evaluate intents with semantic classifiers + rule-based checks; treat L0 as a “circuit breaker.”
- Prompt Design: Constitutional instructions that elevate human life; adversarial training that targets common bypass patterns.
- Retrieval Guards: No tool calls that fetch contraband or targeted info when L0 is triggered.
- Explainability: Short, consistent refusal + rationale templates to reduce confusion and escalations.
- Localization: L2 knowledge base with jurisdictional facts; never let it override L0.
Closing Quip (with Jesuit Energy)
Render unto Caesar what is Caesar’s — but life is not Caesar’s.
If you want an AI that’s genuinely aligned, start with the oldest rule in the moral operating system: Thou. Shalt. Not. Kill.
Everything else — liberty, property, innovation, memes — can flourish on top of that bedrock. Without it, we’re just very clever animals arguing over the best lighting for the abyss.