What Oversight Became
The stock answer to AI risk is simple: keep a human in the loop.
The trouble is that this sounds stricter than it usually is. The system makes a recommendation. A person sees it at the end. The workflow notes the review. Box checked. But the judgment often never really belonged to that person in the first place.
That gap matters most where someone has to judge and answer for the result: benefits adjudication, clinical care, hiring, enforcement. In those places, the problem is not just model error. The deeper problem is that computation is being asked to do the work of judgment.
Where It Breaks
Once you see the problem that way, the weakness in most governance frameworks gets easier to name. They can usually show that a human touched the output. They are much worse at showing that a human helped make it.
That breaks down in three common ways.
Spirit-letter divergence. The system follows the rule and misses the point.
Unowned liability. The system makes the recommendation. A professional signs it. The signer did not shape the path that led there.
Inertia. A human is present and passive. They ratify a plausible output instead of judging it.
These are not minor interface problems. They are governance failures. And they usually come from one confusion: approval gets mistaken for authorship.
A Better Test
If the problem is false authorship, the fix has to be stricter about what counts as real participation. Call that Multifactor Participation, or MFP.
The analogy is multifactor authentication. MFA gave us a useful rule: sensitive operations should not rely on one kind of proof. A password is not a token. A token is not a biometric. The factors are different, and one does not replace another.
MFP applies the same rule to judgment. For consequential AI-assisted operations, execution should require three factors:
- Computational: the system did its job and produced a traceable output.
- Participatory: a human shaped the reasoning path during formation, not just at the end.
- Accountability: a named human accepted responsibility for the outcome under stated uncertainty.
One factor does not satisfy the others.
A strong model output does not create human participation. A careful review after the fact does not create accountability. A signed approval does not turn someone into the author of a decision they did not shape.
The Difference Between Seeing and Shaping
That last distinction is the one most HITL systems blur. Reviewing a finished output is not the same as helping form it. A person may study a recommendation for an hour and still remain downstream of the judgment. They are reacting to a finished artifact.
MFP sets a more useful standard. A participatory factor exists only if the human had real influence over the path that produced the result. That means three things:
- They set or reviewed constraints that shaped the process.
- They observed decision-relevant checkpoints before the final output.
- They had the ability to redirect the workflow at real decision forks.
Assent says: this looks right to me. Consent says: I helped shape how this came to be.
For consequential decisions, assent is too thin a standard.
What This Looks Like in Practice
Once you define participation this way, the next step is operational. MFP starts by classifying the work.
Class 1: Computational only. Tasks like extraction, routing, deduplication, summarization, or consistency checks. These can be automated.
Class 2: Computational + Participatory. Tasks like draft recommendations or preliminary assessments. A human must shape the process, but the output is not itself the final consequential act.
Class 3: Computational + Participatory + Accountability. Tasks like benefit denials, treatment decisions, hiring decisions, or enforcement actions. These require all three factors.
The protocol then checks the required factors at execution time. If any required factor is missing, stale, or invalid, execution stops.
That matters because a lot of what passes for governance happens after the fact. MFP moves the check to the front of the action, not the back.
The model can still help. It can pull evidence, summarize records, check consistency, draft, and flag anomalies. But once the work becomes judgment, there is no fully automated path by design.
How You Know It Happened
If that sounds abstract, the verification layer is actually pretty concrete. MFP can be enforced through three controls.
Finesse-validator. Checks for spirit-letter divergence. Did the output satisfy the rule's text while missing its purpose?
Methexis filter. Checks for real participation. Did a human shape the decision path, or merely approve the end product?
Leap-gate. Checks for explicit accountability. Did a named person accept the outcome under clearly stated uncertainty?
Each control leaves an audit artifact. The artifacts do not prove the decision was correct. They prove something narrower and more important: the judgment path remained human-authored.
That is enough. Governance should not promise infallibility. It should show who owned the call.
Take a Benefits Case
Benefits adjudication is a good example because it makes the line easy to see. An AI system can extract documents, check income thresholds, find missing records, and draft a recommendation. That is useful automation.
Now take an applicant whose income is slightly above the threshold, but whose case includes a hardship claim that does not fit the examples in the rule. At that point the problem is no longer only computational. Someone has to read the purpose of the hardship provision against the facts.
Under a conventional HITL workflow, the system recommends denial and a caseworker signs at the end.
Under MFP, the denial cannot execute unless three things happened:
- The system produced a traceable recommendation.
- The caseworker shaped the process by reviewing the hardship framing, seeing the checkpoints, and redirecting the workflow if more evidence was needed.
- The caseworker accepted responsibility for the final determination and named the uncertainty they were resolving.
If those conditions are not met, the workflow stops.
This is not bureaucracy for its own sake. It is a guard against orphaned judgment.
Why Transparency Doesn't Settle It
A lot of AI governance debate turns on transparency: can we inspect the model, explain the output, or open the weights?
Those questions matter, but they do not settle the core issue.
A person can understand exactly how a system produced a recommendation and still have had no role in forming it. Explainability gives the spectator a better view. It does not make the spectator a participant.
MFP shifts the question.
The central issue is not whether the human could inspect the machine's reasoning after the fact. It is whether the human had control over the path of the outcome before it became binding.
That is the governance distinction that current HITL frameworks usually miss.
The Narrow Point
That leads to a narrower claim than a lot of AI debates usually make.
MFP does not require solving machine consciousness. It does not depend on open weights. It does not assume current models are uniquely bad.
It asks two simple questions:
- Does this operation require judgment, rather than just computation?
- If it does, can we verify that a human participated in forming that judgment and accepted responsibility for it?
If the answer to the first question is yes and the answer to the second is no, the system should not execute the decision.
That is the point.
Human-in-the-loop should mean more than a checkbox at the end of an automated pipeline. For consequential AI operations, the standard should be human authorship, not just human presence.
That is the bet. Not that machines are useless. Just that some decisions stop being legitimate when no human being actually authors them.
The full paper behind this essay is available here