Claude is an artificial intelligence assistant created by Anthropic, an AI safety company. Claude uses a technique called Constitutional AI to ensure it is helpful, harmless, and honest. A key part of this is something called a verification code, which provides transparency into Claude’s capabilities.
What is a Verification Code?
A verification code is essentially a hash identifying which constitutional model Claude is using and different capabilities it has been verified for. Each version of Claude has a unique verification code so users can check exactly what Claude has and has not been verified to do. The code also points to documentation explaining what verification procedures were used.
Why Have a Verification Code?
Verification codes serve multiple important purposes:
- Transparency: They allow anyone interacting with Claude to see upfront what it has and has not been verified for in terms relevant to safety. This builds trust through transparency.
- Accountability: The codes link to particular models, data, and verification methods. This enables accountability if something were to go wrong or claims were inaccurate.
- Trustworthiness: Verification procedures check capabilities like helpfulness and harmlessness. Codes show users that care has been taken to make Claude trustworthy before allowing open access.
- Research Value: Codes also create common knowledge to coordinate future verification research aimed at mathematical certainty around beneficial systems.
What Capabilities Are Verified?
Anthropic has focused initial verification efforts on 24 core Constitutional AI attributes of:
List of 24 capabilities verified with definitions of each
Verification codes indicate which capabilities Claude has been verified for through Constitutional AI methods – and at what level of mathematical formalism. For many current capabilities, strong statistical evidence demonstrates safety up to a particular usage level. Ongoing work aims at formal mathematical proof as the gold standard where possible.
Transparency Descriptions & Levels
Verification codes link to detailed documentation describing:
- What formal claims have been verified
- What procedures were used in that verification
- What mathematical formalism backs those verification procedures
There is also a standardized transparency scale indicating levels of verification from statistical evidence to mathematical proof:
5 level scale explained with examples
Users can see exactly where capabilities fall on this scale – enabling informed decisions around appropriate usage. Ongoing research continually strengthens formal backing through new verification achievements.
Unique Codes for Each Claude Version
As researchers expand Claude’s verified safe capabilities over time, each version receives a new verification code. Codes thus serve as an immutable record and reference for particular models and claims around Constitutional AI safety.
Users can check Claude’s code at any time through the portal interface or API – satisfying the Constitutional right that citizens know which laws (verified capabilities) currently govern system behavior. Codes also link release notes explaining any changes.
The Latest Claude Verification Code
As of February 2024, the latest Claude verification code is:
dc3f-7621-7e18-8103
This verification code indicates that the current Claude model has statistical verification of 24 core safety properties to handle a wide range of assistant tasks. Mathematical proofs apply in limited cases that are explicitly documented.
Ongoing verification research by Anthropic aims to continually expand Claude’s formal guarantees – improving safety and performance through new achievements. Each resulting Claude version will be released with a new verification code as that transparent record.
Conclusion
Claude’s verification codes enable any user to instantly check the verified capabilities and transparency level behind constitutional AI safety guarantees.
Codes serve as an immutable public reference bridging model versions over time through service evolution. They represent both safeguards today and coordination points to drive future verification progress for the benefit of society.
