A system should not become powerful faster than it becomes answerable.
The Laboratory examines what an AI system claims, what authorises those claims, who remains responsible, and whether affected people can contest the result.
“The right answer is not enough without the right conditions for giving it.”
Build a preliminary risk profile.
Commission a full responsibility assessment before scaling or formal institutional reliance.
Begin with the question the system would prefer to avoid.
Each lens exposes a different form of authority: authority to claim knowledge, to interpret meaning, or to decide what happens to a person.
When should an AI-generated claim be treated as a reason to believe something?
Model outputs can be fluent, useful, and still poorly justified. This lens traces the conditions under which an answer deserves confidence rather than merely attention.
A system that sounds certain without warranted reliability can distort human judgment at scale.
What process produced the claim, and what evidence can it be traced back to?
How does performance change under novelty, ambiguity, or adversarial framing?
Where should uncertainty remain visible instead of being compressed into confidence?
Concepts are instruments, not decoration.
The map shows how epistemic and normative concepts depend on one another. Select a node to inspect its role in the method.
From technical description to accountable judgment.
Frame the decision
Name the decision, affected people, institutional setting, and consequences of a wrong output.
Move from a public heuristic to a versioned, evidence-based assessment.
Upload documentation, run all four modules, preserve findings, compare versions, and export a structured report for internal review.