Hong Kong’s banking supervisor now asks banks to pair artificial intelligence with “effective human oversight”, and the EU AI Act demands the same of high-risk systems. But what does a “reviewed by” entry in a file actually prove? Independent AI researcher Andreas Ehstand argues that an AI explanation describes the machine, not the person who clicked “Approve”, and proposes distinguishing four acts of human oversight so that records say what a human really vouched for.
Hong Kong’s banking supervisor now asks banks to pair artificial intelligence with “effective human oversight”, and the EU AI Act demands the same of high-risk systems. But what does a “reviewed by” entry in a file actually prove? Independent AI researcher Andreas Ehstand argues that an AI explanation describes the machine, not the person who clicked “Approve”, and proposes distinguishing four acts of human oversight so that records say what a human really vouched for.
On 15 October 2026 the Hong Kong Monetary Authority announced its “ART” Framework, built on accountability, resilience and trust, and issued new supervisory guidance on the responsible adoption of artificial intelligence by banks. One of its deputy chief executives, Arthur Yuen, spoke of pairing innovation with “effective human oversight”. Where such oversight is documented, someone who was not there may later have to read the record. This essay asks what such a record can show, and what it leaves open.
It is a reading aid, not legal advice. The case below is constructed and describes no particular authority, bank or product, and the essay does not replace an organisation’s own controls, legal duties or the validation of its own setup.
The caseworker’s click
A public authority uses an AI system to assess eligibility for a social benefit. Annex III, point 5(a), of the EU AI Act lists systems intended for that purpose. The system proposes a refusal and shows an explanation, such as the factors that weighed most. A caseworker opens the file and presses “Approve”. The record now reads “reviewed by”, followed by her name. Months later the applicant contests the refusal. What can anyone reading that record conclude about the caseworker?
Very little, and the explanation does not change that. The machine’s reasons and the human confirmation arrive together and feel like one assurance. In fact they are evidence about two different subjects.
Two subjects, four acts
An explanation describes the system: how, more or less faithfully, an output came about. “Reviewed by” describes a person, and it rarely says what she did. The same click can stand for four acts. Three describe what she did. The fourth adds the authority under which she did it.
She may only have been shown the output. That is an observation, and it establishes nothing about the output. She may have checked it, meaning she determined whether a stated property holds. One example is whether the income figure the system used matches the submitted payslips. This is a named property: it is fixed before the check and stated precisely enough that someone else can tell what a yes or a no would establish. She may have decided to issue the refusal, override it, or pass the file to a colleague. Or she may have released it, permitting the refusal to go out under authority her role gave her before the case arrived. A colleague without that authority can decide but cannot release.
The acts establish different things. A check of the income figure says nothing about household size. A release shows that the refusal was permitted, not that it was right. A record that names no act shows only that an interaction took place. I call such a record an undifferentiated approval.
The explanation does not close this gap, and it can widen it. It may make a check easier by pointing to the factors worth examining. But suppose the caseworker finds it plausible and adopts its conclusion without determining anything herself. Then she has observed a tool’s result and nothing more. Article 14(4)(b) of the AI Act names the tendency to rely automatically or excessively on output as automation bias.
What the AI Act describes
Article 14 sets a design requirement. High-risk systems are to be built, including with suitable human-machine interface tools, so that natural persons can effectively oversee them while in use. Article 14(4) lists what the persons assigned to oversight are to be enabled to do. The list runs from understanding the system’s limitations and monitoring its operation to overriding its output and stopping it.
Against the four acts, the list is mixed. Monitoring aims at detecting anomalies, but without a named property it can remain observation. Overriding and interrupting are decisions. The provision describes capabilities a system is to make possible. It does not say which act an overseer performed in a given case.
Article 14(5) comes closest to a check. It covers remote biometric identification under Annex III, point 1(a). No action or decision may rest on such an identification unless at least two persons with the necessary competence, training and authority have separately verified and confirmed it. An exception applies to law enforcement, migration, border control or asylum where Union or national law considers the requirement disproportionate. The object is named, but the features to be verified are not.
Two neighbouring provisions answer other questions. Article 12 requires the logs of those biometric systems to identify the persons involved in verifying results. Article 26(2) obliges deployers to assign oversight to people with the necessary competence, training, authority and support. Both concern who was involved, not what that person determined.
Where the gap matters
Article 22(1) GDPR gives a person the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects. Article 15(1)(h) grants access to meaningful information about the logic involved. That information explains the machine. Whether a decision was solely automated, however, depends on the human contribution, and an undifferentiated approval leaves that blank. The AI Act’s Article 86 is headed “right to explanation of individual decision-making”. However that right comes to be read, an account of the system is not a record of the human.
In SCHUFA (C-634/21), the Court of Justice held that a credit score on which a lender “draws strongly” is itself such a decision. Treating the score as merely preparatory, the Court reasoned, would risk circumventing Article 22. The case concerned scoring, not human review. Its method is still instructive, because the Court looked at the score’s actual influence rather than the formal allocation of roles. A reader of “reviewed by” has no material for that kind of scrutiny.
Ben Green, writing in Computer Law & Security Review (2022), surveyed 41 policies that prescribe human oversight of government algorithms. He argued that evidence suggests people cannot perform the desired oversight functions, so that such policies provide “a false sense of security”. He proposed that agencies support any proposed human oversight with empirical evidence, subject to democratic public review. Such evidence presupposes two things a click does not supply. One is a statement of which act a review step performs. The other is a way to find out whether the step works.
Testing the review step
Take a review screen that writes “checked” whenever it is closed. It produces the same record whether or not anyone read the file, and ordinary operation never reveals the difference. Only cases known to be defective show whether a defect is reported. One such case is a refusal built on a wrong income figure. Testing human inspectors with planted faults is long established, for example in aviation security screening. Such a test is meaningful only if planted cases cannot reach applicants, since a planted refusal that went out would be a defective decision. Results about named caseworkers are also assessments of their work and may fall under rules on employee monitoring. Reliability can be stated for a review step as staffed by a team, rather than for each person.
I have proposed this vocabulary in an individual Internet-Draft, draft-ehstand-oversight-acts, a work in progress.
What a fuller record would say
Return to the contested refusal. The present record shows how the system reasoned and that the caseworker handled the file. A fuller record would say that at 10:42 she checked the income figure against the payslips and found that it matched. It would say that she did not examine household size. It would also say that she then released the refusal under her role’s authority, at a step where planted defects had been caught in testing. Only such a record tells the applicant what a person actually vouched for, and which part of the decision remains open to challenge.
References
- https://www.ietf.org/archive/id/draft-ehstand-oversight-acts-00.txt
- https://artificialintelligenceact.eu/article/14/
- https://artificialintelligenceact.eu/article/12/
- https://artificialintelligenceact.eu/article/26/
- https://artificialintelligenceact.eu/annex/3/
- https://arxiv.org/abs/2109.05067
- https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62021CJ0634
- https://www.hkma.gov.hk/eng/news-and-media/press-releases/2026/10/20261015-3/
AI tools were used in researching and drafting this article; the author takes full responsibility for its content.
Read also
- Can AI explain itself?
- “AI is first and foremost applied mathematics, not the magic we are led to believe” – Axel Cypel (interview)
- Relational AI for Minors: Where the Evidence Supports Regulation – and Where It Does Not Yet
- “Yes to the AI Act, but with moderation” – Thierry Berthier (interview)
- Europe trapped in the Artificial Intelligence paradox