🇩🇪 Deutsche Version: Nominelle Aufsicht

Note: The ethical judgments on this page refer exclusively to the action — never to the person who performs it or who came into existence through it. Cf. Note on Ethical Judgments.

Nominal oversight is a formally assigned human oversight of an automated decision process that cannot be substantively redeemed. A responsible party is named, he is reachable, he signs — and he signs off on what he cannot examine, for want of explainability of the system, for want of time, or for want of access to the grounds of the decision. The oversight exists as a competence and is missing as an execution.

The concept names no negligence on the part of individual persons. It names a form of operation. Whoever has to decide within five minutes about an output that arose from features the system does not communicate, and could not communicate even if asked, has not carried out a poor examination — he had no examination available to him. Precisely for this reason nominal oversight is a structural and not a moral reproach directed at the supervisor.

Ontological classification

is a subclass of: human oversight of an automated decision process (cf. AI-algorithmic arrangement)
is incompatible with: genuine human control (meaningful human control)
is the typical shape of: the responsibility gap in regulated operation
is stabilized by: automation bias
stands in connection with: responsibility, governance framework

The incompatibility with genuine human control is not a game of definitions but the point of the matter. Where the law demands human oversight, it demands it as control, not as a note of competence. Nominal oversight satisfies the requirement according to its wording and misses it in substance — and both at once are the reason why it is so hard to notice.

The load-bearing thesis

The customary description of the responsibility gap assumes that no one is competent: the manufacturer did not foresee the individual decision, the operator did not release it, the machine cannot bear guilt. This description fits the unregulated case.

In regulated operation matters stand otherwise, and the thesis of this page is: the gap is produced not by the absence of a responsible party, but by the presence of a responsible party without any possibility of examination. That is the more dangerous shape, and for three reasons. First, it formally satisfies every requirement: there is a name, a signature, a record. Second, it triggers no alarm — it does not present itself as a breach of the rules but as compliance with them. Third, it produces an appearance of control that replaces control: because a human being has signed off, the decision counts as humanly taken, and the question whether it was is no longer raised.

The open gap is to that extent the more honest problem. It is visible, it produces dispute over competence, it keeps the question awake. Nominal oversight closes the question without answering it.

Why it stabilizes itself

Nominal oversight is not a condition that corrects itself. It is consolidated by automation bias — the empirically well-attested inclination to follow the machine’s output, and precisely so when one’s own basis of judgment is thin. Whoever cannot retrace an output has no occasion for contradiction; and whoever contradicts and turns out afterwards to have been wrong bears the burden of justification alone, whereas assent shares it. The structure of incentives favours signing off.

To this is added a feedback effect: the more reliable a system appears in normal operation, the less often it is examined, and the less often it is examined, the less practice there is in examining. The capacity for oversight withers at the very place where oversight is organizationally anchored. The supervisor remains formally the last instance and becomes in substance a conduit.

For responsibility this has an uncomfortable consequence. It has not disappeared but has been assigned — to someone who cannot bear it, because he lacks the condition of bearing it: the possibility of deciding otherwise, and that means the possibility of knowing what is being decided.

Range

The concept was won from algorithmic embryo selection, but it is not confined to it. It takes hold wherever law or professional regulation demands “human oversight” and practice lets it wither into a signature:

in autonomous weapon systems, where the release of a proposed target is supposed to take place within a few seconds and the formation of the target itself is not open to view;
in algorithmic personnel selection, where a ranking of applicants is confirmed whose coming about the confirming party does not know;
in medical decision support, where a recommendation is to be countersigned for whose examination the consultation time does not suffice;
in automated administrative decisions, where a caseworker confirms a draft that arose from combinations of features that are not documented in the file.

In all these cases the structure is the same: an examination required by law, an examination in fact not possible, a confirmation nevertheless issued.

The strongest objection to this assessment

The most serious counter-objection runs: every oversight is incomplete. The chief physician does not recheck every laboratory analysis, the judge does not recalculate the expert report himself, the board does not read every contract. The division of labour always means that someone vouches for something he has not retraced in all its parts himself. If “nominal oversight” is already present wherever examination is not complete, the concept applies to every delegation and becomes useless — a reproach that fits everything distinguishes nothing.

The objection is warranted and demands a line to be drawn that cannot lie in the degree of examination but in its possibility. Whoever does not recalculate a laboratory result could recalculate it: the procedure is described, the raw data are available, in case of dispute the examination is enforceable, and the examiner knows what he is relying on and what errors this method typically makes. Delegation of this kind is well-founded trust — it keeps examination in reserve.

Oversight becomes nominal where this reserve is missing. Three marks name this more precisely. First: the supervisor cannot reach the grounds of the decision even with full effort and unlimited time, because the system does not disclose them. Second: the oversight has no criterion by which a contradiction could be justified — it can assent or refuse, but it cannot examine, and refusal without a criterion is arbitrariness, not control. Third: the frequency of divergence stands structurally at zero; where over long periods there is never any contradiction, the process is no longer an examination but an execution.

Where all three marks apply, nominal oversight is present. Where examination is merely incomplete because time is short, but examinability exists, it is not present — there an organizational problem exists that is solvable by organization. This distinction has a practical price: it demands a statement about what a system discloses and what it does not, and this statement is seldom easy to make. The concept is thereby neither sharply demarcated nor arbitrary; it is as sharp as the explainability of the respective system is assessable.

What remains open: this line is a determination, not a measuring procedure. Whether in a concrete case the second condition is fulfilled — whether there really is no criterion for contradiction, or merely none that has been practised — is an act of interpretation. Whoever uses the concept owes a justification for it in the individual case.

Was den Menschen zum Menschen macht

Übersicht

Nominal Oversight

Ontological classification

The load-bearing thesis

Why it stabilizes itself

Range

The strongest objection to this assessment

See also

Graphansicht

Inhaltsverzeichnis

Backlinks