Abstract
This disclosure describes three related layers for video calls and AI-facilitated interviews. First, a
call system may infer a likely accessibility need from audio, video, or interaction signals already
present in a human-to-human or AI-participating video call, and may then adapt the experience —
for example with captions, sign-language presentation, a changed speech rate, extended response
time, layout or contrast changes, or alternative input — preferably by offering options or
honouring prior consent rather than silently auto-enabling. Second, protective embodiments keep
inferences and elicited facts about the listed categories — disability; health (including mental
health conditions such as ADHD or depression); pregnancy; sexual orientation; gender identity;
apparent age; race or ethnicity; accent or dialect; body weight; perceived physical attractiveness;
alcohol or substance use and recovery history; personal-life circumstances (for example
relationship situations); and projected health-plan cost — out of hiring, workplace, and other listed-
domain decision pipelines; hide them from employers, hosts, decision-makers, and other
participants; minimise and localise data; log access; detect covert or pretextual assessment; and,
on the candidate’s or an auditor’s side, detect an interviewer or AI steering the conversation
toward health, medical leave, or family plans, or using reciprocal self-disclosure to draw
information out, then warn the candidate in real time, redact volunteered protected details from
transcripts and decision records, enforce platform policy, and write audit logs. Protective tools also
detect synthetic self-disclosure bait, tell the candidate when an interviewer’s anecdote is
fabricated, and require any personality-inference score to be explainable and contestable, with
audit logs. Third, for completeness of the prior-art record only, it is disclosed at a conceptual level
that a platform could infer any of the listed categories from audio and video, including through
covert in-interview tasks, and use that to rank, filter, or exclude in hiring or the workplace, or in
the other listed domains; that an interviewer or analogous agent (human or AI) may monitor for
health or other listed-category information a person discloses, or use conversational tactics such
as sharing made-up personal health, medical-leave, or substance-use stories to draw out health,
disability, pregnancy, substance-use or recovery history, or other information in the listed
categories, which then influences a decision in a listed domain; and that an AI interviewer or agent
may share a made-up personal anecdote, for example about recent drama in its life, and use
whether and how the person reciprocates to infer personality traits such as reliability, drama- or
conflict-seeking, or discretion, which then influences such a decision. This version specifies the
protective and detection embodiments in enabling technical detail: component placement (on-
device or server), data flows and retention, decision logic as pseudocode or state machines,
consent and audit-log schemas, firewall enforcement, redaction pipelines, alternative-assessment
routing, and example bias-test metrics and thresholds, with variations under each embodiment.
Existing products and patents already cover substantial parts of automatic detection and
mitigation of certain impairments in video conferences, auto-captions, and sign-language
presentation. This draft does not present those already-claimed auto-detect-and-auto-reconfigure
paths as new embodiments. It emphasises consent-gated adaptation, on-device processing,
decision-firewalls, and audit and covert-assessment detection. It claims no novelty clearance and
is not an implementation.
Keywords: video call; accessibility; captions; sign language; reasonable accommodation; inferred
protected characteristics; hiring; workplace; dating; housing; insurance; lending; education;
healthcare; immigration; policing; consent; on-device processing; audit; covert assessment;
conversational steering; redaction; synthetic self-disclosure; personality inference; representative
selection.
Creative Commons License

This work is licensed under a Creative Commons Attribution-Share Alike 4.0 License.
Recommended Citation
Boll, Joe, "Accessibility-Need Inference During AI Video Calls, and Protective Controls Against Discriminatory Use of Inferred Protected Characteristics", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/13039