Learning Objectives
By the end of this module, you will be able to:
- Articulate why AI systems need a standardised way to receive human values
- Distinguish between hard-coded safety rules and portable constitutional values
- Explain the "values portability problem" with concrete examples
- Describe how VCP fits into the broader AI safety landscape
1.1 — The Values Gap in AI Today
Every major AI system has safety rules. But those rules are:
- Proprietary — locked inside each provider, invisible to users
- One-size-fits-all — the same guardrails for a children's tutor and a medical researcher
- Non-portable — switch providers, lose your value configuration
- Opaque — users can't see, verify, or influence what values shape their AI's behaviour
Consider this scenario: A counselling organisation uses AI assistant A, carefully configured to be trauma-informed, culturally sensitive, and non-directive. They switch to AI assistant B. All that configuration is gone. They start from scratch — if provider B even supports those values at all.
This is the values portability problem. Values are trapped inside individual AI systems, invisible and immovable.
1.2 — What If Values Were Portable?
VCP treats values like data, not code. A constitutional document — a set of principles, boundaries, and guidance — travels with the user or organisation, not the provider. Any VCP-compatible system can read it, verify it, and apply it.
| Without VCP | With VCP |
|---|---|
| Values hard-coded per provider | Values travel with the user |
| Switch providers, start over | Switch providers, values follow |
| Can't verify what values are active | Cryptographic proof of active values |
| Same rules for everyone | Context-sensitive adaptation |
| No audit trail | Tamper-evident value history |
1.3 — The Three Layers
VCP organises context into three layers, each answering a different question:
- Constitutional Layer — "What principles should guide this AI?" Organisational values, ethical frameworks, domain-specific rules.
- Situational Layer — "What's happening right now?" Time, place, social context, task type.
- Personal Layer — "Who is interacting, and what do they need?" Cognitive state, energy level, emotional tone — all privacy-filtered.
These layers compose. A healthcare AI might have a constitutional layer emphasising patient autonomy, a situational layer noting it's an emergency department at 3am, and a personal layer indicating a stressed, fatigued clinician. The AI's behaviour adapts across all three.
1.4 — VCP in the Ecosystem
Where VCP sits relative to other approaches:
- System prompts: Static, provider-specific, unverified. VCP is dynamic, portable, cryptographically signed.
- RLHF / Constitutional AI: Training-time alignment. VCP is inference-time alignment — complementary, not competing.
- Content moderation APIs: Binary allow/block on outputs. VCP shapes how the AI reasons, not just what it says.
- Model cards / datasheets: Documentation about a model. VCP is operational guidance to a model.
VCP doesn't replace any of these. It adds a missing layer: a standardised, verifiable, portable way to tell AI systems what matters to the people using them.
See It in Action
The Gentian demo shows VCP's portability in practice — the same constitution applied across different AI providers, producing consistent behaviour.