Government AI needs a legal release gate, not just a human-in-the-loop
An AI-assisted decision should not be released merely because a human reviewer has looked at it. The reviewer also needs evidence visibility and a named release authority.
For consequential AI-assisted decisions, governments should distinguish between the person who reviews an output and the person who has authority to release or rely on it. Image: Canva
For almost two decades, my work as a lawyer has involved a recurring problem that technology did not create.
Important decisions often fail not because nobody reviewed them, but because the reviewer could not clearly see what the decision depended on, what remained uncertain, or what changed since an earlier approval.
Artificial intelligence (AI) makes that problem more urgent.
A public agency can put a “human in the loop” and still have a weak control environment.
Someone may review an AI-assisted recommendation, but that does not answer three harder questions.
What evidence was the reviewer relying on? Who had authority to release the decision? And what happens when the law, facts or source material changes after approval?
Those questions are part of what led me to develop MERENVAY, a legal intelligence and decision-assurance project.
The aim was to make the basis of a decision more visible and to treat approval as something that can expire when its foundations change.
Human review is a role, not a control
“Human review” sounds reassuring because it places a person between the model and the final decision.
But the phrase is too broad to function as a governance standard on its own.
A reviewer may see a polished recommendation without seeing the provenance of the underlying sources.
They may know the conclusion, but not which assumptions remain unresolved.
They may approve an output that was valid yesterday but has become stale because a regulation, policy, evidentiary record or factual circumstance has changed.
In legal practice, these are familiar failure modes.
A conclusion is only as reliable as the authorities, facts and assumptions that support it, and its continued validity depends on whether those foundations remain current.
Public sector AI should be governed the same way.
First: name the release authority
For consequential AI-assisted decisions, governments should distinguish between the person who reviews an output and the person who has authority to release or rely on it.
That release authority should be explicit.
In a low-risk workflow, the reviewer and releaser may be the same person. In a higher-risk context — licensing, enforcement, benefits, procurement, regulatory decisions or legal positions — the release role may need additional professional or statutory authority.
This creates a simple accountability question: who is permitted to turn an AI-assisted recommendation into an operative government decision?
If an organisation cannot answer that clearly, “human in the loop” is not yet a complete control.
Second: make evidence state visible
A reviewer should not receive only a conclusion. They should be able to distinguish at least four things:
- verified evidence;
- retrieved or cited material that still requires checking;
- assumptions;
- unresolved questions.
The reviewer should also be able to see the jurisdiction, source, date and relevance of the authorities being relied upon.
This is especially important for legal and regulatory work, where two sources can look equally authoritative on a screen while having very different legal status.
A government team can implement this without building a complex new AI system.
Start by changing the review template.
Require every consequential AI-assisted recommendation to show its sources, the status of those sources, any unresolved assumptions and the date on which the underlying position was last verified.
That single change can turn a review from “does this answer look right?” into “is this decision supportable?”
Third: define what automatically reopens the decision
The weakest assumption in many approval systems is that once a human has approved something, the approval remains valid.
It may not.
A legal decision can become stale because legislation changes. A procurement recommendation can become unreliable because supplier information changes.
A regulatory assessment can become unsafe because a key data source is corrected or withdrawn.
Governments should therefore define “reversal conditions” before release.
These are events that automatically trigger re-review: a change in law or policy, a material new fact, a source being withdrawn, a significant change in the underlying data, or the discovery that an assumption used in the earlier decision was wrong.
This turns human oversight from a one-time checkpoint into a continuing assurance mechanism.
From AI governance to decision assurance
When I began developing MERENVAY, this was the problem I wanted to address: not how to make AI sound more confident, but how to make professional reliance more disciplined.
The project is still a public beta, and I do not present it as proof that one architecture has solved legal AI.
Its value to me has been as a practical way to test a broader idea: an AI-assisted decision should not be released merely because an answer exists and a person has looked at it.
The reviewer needs evidence visibility. The organisation needs a named release authority. And the system needs a reason to reopen a decision when its foundations materially change.
These principles extend well beyond legal work.
For a public sector leader looking for one thing to change tomorrow, I would start with the approval form.
Replace the single box marked “human reviewed” with three questions:
- Who has authority to release this decision?
- What evidence, assumptions and unresolved issues did they review?
- What future event would require this decision to be reviewed again?
Those questions do not eliminate AI risk. But they make the accountability around AI much harder to hide behind a reassuring phrase.
-----------------------------------
Mohammed Mazen Al Nasser is a Jordanian lawyer and legal adviser with nearly two decades of experience across commercial, public and regulatory matters, including contracts, disputes, intellectual property, telecom and digital issues, and legal reform.
He holds an LL.B., a master’s degree in Intellectual Property, and a master’s degree in Journalism and Modern Media.
He is the creator of MERENVAY, a public-beta Legal Intelligence & Decision Assurance project focused on evidence visibility, uncertainty, re-review conditions and Human Legal Review in AI-assisted legal decision-making.
