A false trigger is never fixed by recognising the protected attribute that caused it: auditing an alarming person-detection system for discrimination without acquiring the attribute
Year: 2026
Type: preprint
Abstract: This record states one rule for auditing an alarming person-detection system for discrimination, and three mechanisms that make the rule checkable — one of them a boundary that is declared rather than closed. It is a method record: it contains no measured false-trigger rates. How such a measurement would have to be designed is part of the method; the measurement is not. The rule. A false trigger is never remedied by teaching the system to recognise the protected attribute that caused it. It is remedied by defusing the consequence: do not persist, do not escalate, do not forward without a human gesture — or drop the signal. The obvious remedy is the harmful one. If a camera-based alarm marks a person kneeling in prayer as a fall, the immediate reaction is to build a "praying / not praying" classifier — which is an inference of religious belief from an image, just as "walking aid / no walking aid" is an inference of health data. The remedy would create precisely the category of data whose avoidance was the point — and the opposite path is built, patented and published. Half the rule is a demand, not a property — stated here rather than left to be discovered. Of the four remedies the rule names, one is enforced in code, one is organisational and expressly not enforced, and two are named and not built, with production code frozen until December. The record carries that build status per remedy as a table. A record that lets the reader assume a stated doctrine is already implemented would commit the very error it describes — remedying an appearance instead of a cause. The three mechanisms. First, the annotation label of the fairness measurement is an observable posture or gait, and the coupling attribute → posture stays deliberately and permanently unmeasured, because measuring it would mean acquiring the attribute; this is declared as a boundary of the method, not a gap to be closed later. By the same standard as above, that boundary is organisational and not enforced in code: nothing prevents someone from acquiring the attribute; what exists is a stated commitment not to, placed where a later reader can hold it against the record. That is what separates it from proxy approaches, which reconstruct the attribute from observable patterns. Second, a false-trigger register: each entry carries an evidence level, the strongest level has a minimum count of its own, and each entry names a mechanically verified code anchor. What it documents is a negative capability — the system does not distinguish prayer from a fall, and is never meant to. Third, a double-wired floor: the minimum counts exist twice, read from the document and hard-wired in the test, checked against the maximum of both, so lowering a number in the document does not lower the check. What is prior art, stated by the record itself. Four components are listed by name and date as not claimed: posture classes as the grouping axis of a fairness measurement, a minimum count on a guard's own rule list, mutation testing on a static guard, and checking source identifiers against a dictionary of sensitive terms. Each component the record does claim — the rule and the three mechanisms — stands against a documented alternative, not against a gap. Search coverage is a number, not a claim of completeness: 175 approaches across literature, patent databases and code hosts, each with its own positive control; sources that failed it are recorded as unmeasured, not absent. The search was English-language only; German, French, Chinese and grey literature was not covered — a real gap. Nothing here is a legal opinion.
Source: arXiv (Cornell University)
Author: Bering, Alexander
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7084447449
Cites:
Cited by:
Related to: