One alignment doesn't fit all
-
Value monoculture. Standard alignment reduces value diversity. Post-aligned models are less representative of human populations than pre-aligned ones.
-
You're the arbiter. Every value complaint lands on your trust and safety team. Every cultural context you don't serve is revenue you don't earn.
-
One model per market. Serving diverse markets means training multiple models or refusing to serve markets whose values differ from your default.
"Post-aligned models exhibited less similarity to human populations compared to pre-aligned models."
Sorensen et al., A Roadmap to Pluralistic Alignment, 2024
Values at runtime, not training time
The Value Context Protocol (VCP) separates values from weights. Users select creeds: portable, signed, composable value bundles that travel with them to any VCP-compliant provider. Your model stays general-purpose. User values are applied at inference time through protocol headers.
Transportable
Creeds are signed, versioned, and portable. A user's values follow them from your platform to any other VCP-compliant provider. You're not locking them in. You're giving them a reason to stay.
Composable
Multiple creeds stack with explicit conflict resolution. A hospital creed layers on top of a base safety floor. An institutional compliance creed extends a community ethics creed. No averaging. No exclusion.
Adaptive
Context changes: time of day, who's in the room, how the user is feeling. VCP encodes 16 dimensions of context so the same creed produces different expression for a calm Tuesday versus a 2 a.m. crisis. Personal-state signals carry a source label, such as declared by the user or inferred from their messages. Context modulates expression, never boundaries.
The closest evidence: public-input constitutions.
In 2023, Anthropic published a public-input experiment run with the Collective Intelligence Project. Approximately 1,000 Americans authored constitutional principles for an AI model, and the resulting model was tested against Anthropic's expert-authored default. That study tested community-authored constitutions, not VCP or Creed Space; it is evidence for the premise VCP builds on.
- Same benchmark scores. The public-constitution model scored the same as the standard model on MMLU and GSM8K, and was rated as helpful and harmless.
- Lower measured bias on all nine BBQ dimensions, most on disability status.
- About half the concepts shared. Roughly 50% overlap in concepts and values.
Source: Anthropic and the Collective Intelligence Project, Collective Constitutional AI (October 2023)
Before and after VCP
| Today | With VCP |
|---|---|
| You author one alignment for all users | Users select from a creed library or write their own |
| Values baked into weights at training time | Values applied at inference time via protocol headers |
| Every value complaint reaches your T&S team | Value choices are attributed to the creed's author, and signed where a governance authority attests them. You enforce the protocol and the safety floor. |
| Serving diverse markets means multiple models | One model, many creed configurations |
| Alignment means your values, applied to everyone | Alignment means each user's values, within a shared safety floor |
Three integration tiers, from lightweight to full.
| Tier | Time to integrate | Adds |
|---|---|---|
| 1VCP-Lite | Days | Parse VCP headers. Pass creed text as system prompt context. You're already doing most of this. VCP-Lite adds structure and portability. |
| 2VCP-Standard | Weeks | Everything in Lite, plus deterministic hook execution and transition detection. Deterministic hooks are designed to resolve common value checks without an LLM call; sub-millisecond is an internal target. |
| 3VCP-Full | Months | Everything in Standard, plus 16-dimension context encoding, full creed composition with conflict resolution, and state tracking. |
User choice within configured boundaries
In reviewed source, users configure preferences above a safety floor that is non-user-editable. This is an architectural constraint, not a guarantee of harm prevention or effectiveness in a particular deployment.
The markets that need this
- Large Deployers
- Industry-specific compliance creeds without custom model training. The compliance team writes the creed.
- Education
- Age-appropriate, curriculum-aligned value configurations per school district.
- International
- Cultural value adaptation without per-country model variants.
- Families
- Family safety creeds that travel across participating AI products a child uses.
- Faith Communities
- Value configurations that respect specific ethical and spiritual traditions.
Prompting is governance. Treat it accordingly.
Every instruction in a system prompt changes the equilibrium of a complex system. The behaviour that emerges isn't in any single instruction. It's in the interaction between all of them. Three correct rules can produce an AI that hesitates, defers, and underperforms. None of this is a bug; it is an equilibrium nobody designed.
You wouldn't deploy policy without review. Don't deploy AI governance without architecture.
-
The problem: instruction interaction
AI deployments layer instructions from many authorities: safety, privacy, inclusion, legal, care, and local context. Each instruction can be reasonable in isolation. Together, they produce emergent behaviours that no single team designed or anticipated. The question is what equilibrium they produce.
-
The gap: who owns emergent behaviour?
Purpose aligns humans. Governance architecture aligns machines. Public-interest deployers need both. Creeds give you auditable, versioned, composable governance units instead of sprawling system prompts where instruction interactions are invisible.
-
The solution: creed as governance
A creed is a constitutional document, tested and versioned like code. When instructions conflict, the conflict is visible, attributable, and resolvable. Your compliance team writes governance they can audit. Your AI executes governance it can report on. The equilibrium becomes observable.
Questions providers ask
Won't user values make our model worse?
Available evidence is encouraging. In Anthropic's 2023 experiment with the Collective Intelligence Project, the public-constitution model scored the same as the standard model on the reported capability benchmarks. Values can modulate how a model responds while preserving measured task performance. Collective Constitutional AI (October 2023)
What if we can't let users control alignment?
You're not. Users control their preferences above a mandatory safety floor (UEF) that you enforce. The analogy: users choose their homepage, but they can't disable HTTPS.
Isn't this too much engineering effort?
VCP-Lite is a structured system prompt convention. If you already pass system prompts, you can start with a lightweight prototype before deeper protocol integration.
What about regulatory risk?
VCP can make value choices easier to document and explain: they are auditable, attributable, and inspectable. For EU AI Act conversations about whose values shape behaviour, you have a clearer evidence trail.
Our alignment is fine.
For your median user, probably. For parents, doctors, non-English speakers, religious communities: one alignment doesn't reach. In Anthropic's own experiment, the public's constitution put more weight on accessibility. Collective Constitutional AI (October 2023)
Ready to let your users bring their values?
Start with VCP-Lite. Prototype in days. Scale when you're ready.