Context
Understand the stakes.
Document intended uses, foreseeable misuse, affected groups, data sources, and alternatives. Ask whose needs are missing from the problem definition.
Governance & society
Responsibility is part of the work.
An AI system sits within a human system: people make decisions, carry risk, and live with the consequences. Our proposed approach makes those responsibilities explicit.
Turn principles into decisions.
A practical assessment should explain what a system is for, who it affects, what could go wrong, and who has the authority to act. The depth of review should match the context and potential harm.
Oversight
Keep judgment meaningful.
Specify which decisions need review, what reviewers can see, and how they can challenge, override, or stop the system. A person in the loop needs time and authority.
Accountability
Name the responsibility.
Identify a decision owner, a route for reporting concerns, review triggers, and a process for responding to incidents. Keep a record of decisions and unresolved risks.
What an assessment could cover.
These are proposed areas for collaboration, subject to an agreed scope and suitable expertise. They are not a certification or a guarantee of compliance.
Evidence & limitations
Examine representative use cases, failure modes, performance differences, and uncertainty. Record what the evaluation did not test and where its conclusions should not be generalized.
Data & privacy
Map what information enters the system, why it is needed, who can access it, and when it should be removed. Minimize collection and consider the people represented in the data.
Deployment & review
Agree on release conditions, monitoring, escalation, and rollback. Revisit the assessment when the model, data, users, or purpose changes.
Our own research principles
Make claims inspectable.
- Distinguish proposed work, original outputs, and external reading.
- Describe methods, sources, assumptions, and limitations alongside findings.
- Credit contributors and disclose relevant interests and dependencies.
- Correct errors and explain material changes to published work.
- Consider who could be harmed by a result or its release before publishing.
Build / evaluate conflict
Be clear about who did the work.
If Zero One Labs helps build a system and later evaluates it, that involvement creates a conflict of interest. We would disclose our role, the scope of the evaluation, and any commercial relationship alongside the findings.
Work we evaluate ourselves must be described as an internal assessment. It must not be presented as an independent audit. Where independent assurance is needed, a separate qualified reviewer should be commissioned, with access to the relevant evidence and freedom to report concerns.