Mission
node&norm examines what makes human control of AI-assisted decisions possible, and what evidence would allow others to verify it.
The mission is to make claims of human oversight open to examination. The work asks when an institution has reason to slow down, who has authority to intervene, and whether an intervention changes the result for the people affected.
The longer-term purpose is to inform AGI preparedness by examining how people and institutions can retain practical control as AI capabilities advance.
How the work is examined
The testing perspective starts with explicit conditions: what is being claimed, what evidence could support it and what would reveal a failure. The approach translates research questions into testable requirements, evaluation methods and candidate operational controls. Claims about effectiveness require evidence from the relevant operating context.
Different claims require different methods. Documentary studies examine records; executable checks test behavior under specified inputs; conceptual and normative arguments examine responsibilities and reasons for action. Conclusions must remain within what each method and its evidence can establish.
A confirmation that permission was withdrawn is one piece of evidence. Whether a pending action still happened is a separate question. The example below shows why the distinction matters.
Example: did withdrawing permission stop access?
Fictional workflow · How evidence changes the conclusion
A confirmation is only part of the story.
An AI-assisted workflow is preparing temporary administrator access. A security lead withdraws permission before that access becomes active.
Withdrawal takes effect
The authority change is durably recorded in the same ordering as the protected changes.
A confirmation appears
The interface acknowledges the withdrawal. This message alone does not show what happened to the pending action.
Examine what the system actually did
Check whether access became active afterward, which permission was used, and why any request was rejected.
Three possible readings of the evidence
- Supports the control in this case
- No new access was granted, and the record shows rejection because permission had been withdrawn.
- Shows a failure in this case
- New access was granted after withdrawal took effect, using the withdrawn permission.
- Leaves the outcome unresolved
- Only the confirmation is retained; the records needed to establish what happened to access are missing.
The public research repositories preserve sources, assessment rules, reference implementations and version histories. A correction should record what changed, why it changed and which claims need revision. The research-integrity commitments explain how sources, disclosure and correction are treated.
This emphasis on testable conditions comes from professional experience. Questions about judgment and human control also draw on the intellectual influences behind the work.
The researcher behind the work
Mark Julius Banasihan examines evidence, human authority and accountability in AI-assisted decisions. A 15-year background spanning engineering, software quality, customer operations and technical leadership grounds this research in how systems are built, tested and operated.
Ongoing study at Harvard University complements professional learning through the Massachusetts Institute of Technology, across social sciences, computing and AI.