AI Alignment
Durable infrastructure to support alignment work—not a claim to solve alignment.
The aim is to make AI identity and action verification useful to work on ensuring that AI systems behave in accordance with intended goals and constraints.
As systems and research priorities evolve, a shared verification foundation could remain relevant. That is an ambition to investigate, rather than a demonstrated outcome.
A foundation that can outlast a single application
Make attribution clearer
Identity verification may help researchers reason about which system is associated with an interaction. The value depends on how identities are defined, maintained and interpreted.
Make claims more inspectable
Verifiable evidence about actions could support evaluation and accountability. It would still need to be connected to meaningful behavioral questions and reliable observations.
Keep the contribution bounded
A verified identity is not evidence of aligned intent. A verified action is not evidence that the action was desirable. Alignment requires judgments and methods beyond cryptographic verification.
A central research question: where does this infrastructure add useful evidence to existing alignment approaches?
What verification can—and cannot—tell us
Cryptographic checks can establish specific properties of evidence within a defined scheme. They cannot, on their own, establish alignment, truthfulness, beneficial outcomes or welfare.
Does a verified identity imply a trustworthy AI?
No. Attribution and trustworthiness are different questions. Assessing behavior requires evidence beyond the identity associated with it.
Does action verification prove that an action was safe?
No. Even when evidence for an action is valid, its safety and desirability depend on context, consequences and the goals against which it is evaluated.
What needs further assessment?
The verification scope, threat model, security assumptions and practical usefulness need explicit examination. Alignment applications and welfare compatibility also require their own research and evaluation.