The Inclusion Floor: A Governance Framework for Groups That Enterprise AI Fairness Assessment Cannot See, Measure, or Reach
Year: 2026
Type: preprint
Abstract: Enterprise AI fairness assessment assumes that every group it must protect is present in the data, large enough to measure, and able to reach the system being assessed. In regulated enterprises, none of these assumptions can be taken for granted. Characteristics that define disadvantaged groups, such as disability, are often unrecorded; groups that are recorded are often too small for a fairness metric to detect a disparity at the threshold the organisation has set; and people who cannot use the channel through which a system is delivered never enter the population it assesses. In each case, the assessment reports no breach for the group, and current practice treats that silence as a pass. This paper identifies the conditions of the Inclusion Floor Problem and argues that it is a governance design problem, not a statistical one. The Inclusion Floor Problem has three components. First, enumeration failure: a group whose defining characteristic is not recorded cannot be disaggregated, so no fairness metric can be computed for it. Second, measurement indeterminacy: below a group size that depends on the metric, the threshold and the outcome rate, a disparity at the threshold is more likely to be missed than detected, a property established in the statistical literature on fairness estimation (Lum et al., 2022; Ferrara et al., 2025). Third, access exclusion: people who cannot use a system’s channel are absent from its decisions and data, reproducing the first two failures in the next assessment cycle. Statistical methods can narrow the second component. None of them determines what an organisation must do about a group it cannot assess. This paper introduces the Inclusion Floor Framework: three governance obligations that ensure every member of a system’s intended population is either covered by its fairness assurance or served by an alternative that does not depend on it. The Group Enumeration obligation requires the deployment authority to identify, before assessment, the groups whose outcomes must be assured and whether each is recorded, inferable, or unrecorded in the data. The Measurability Floor obligation sets, for each group and metric, the minimum count below which an assessment result is recorded as Not Assessable rather than as a pass, with floors derived from the fairness thresholds of the system’s risk tier. The Alternative Pathway obligation requires that every unrecorded, unassessable, or channel-excluded group be offered an accessible route to a decision that does not rely on the unassured output, with a remediation plan and disclosure of the groups the assurance does not cover. The paper analyses eight major governance frameworks and regulatory regimes, finding that several require representative data and one names small groups as a consideration, but none specifies a measurability floor, treats an unassessable result as distinct from a pass, or requires an alternative pathway for the groups the assessment cannot reach.
Author: Hossain, M Maruf
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236160223
