# Assistive technology and verdicts

People who use screen readers, voice control, switch access, head pointers, eye gaze or only a keyboard reach a web page in ways that can look unusual: clicks with no pointer movement, a cursor that jumps onto a target, very even timing. Telltale must never call them an AI agent: when their evidence is borderline the answer is Unknown, and missing an agent is acceptable where calling a person an AI agent is not.

## What an assistive-technology user's session reads

- **Human or Unknown, never AI agent, on their own input.** Input from a physical device (a keyboard, a mouse, a switch interface) earns Human; readings that a person could produce, such as VoiceOver's activate-and-paste or a pointer that jumps onto each control, stay Unknown. Unknown is not human, so your Unknown rule decides a protected action.
- **A single behaviour pattern is Medium confidence at most.** If one behaviour family alone ever reads AI agent, it is Medium: held for review (step-up) by default, even when the action's AI agent rule says Deny, and never charged. The site owner can switch Medium to follow the AI agent rule (block) in Control › Rules; the dashboard warns first that real people such as assistive-technology and kiosk users could then be blocked from that action. High confidence needs proof or agreeing evidence; see [What backs an AI agent verdict](/docs/verdicts#what-backs-an-ai-agent-verdict-confidence-bands).
- **If you protect an action that assistive-technology users rely on**, keep its AI agent rule at Hold for review rather than Deny, and leave Medium held for review.

## How we test it

We record consenting people using their own device, their own assistive technology and their usual settings (screen readers, voice control, switch access, keyboard only, head pointer and eye gaze), at least 5 people per group, paid for their time, adults only, free to stop and have their recording deleted at any time. We record only the timing and kind of input events on a fictional demo page, never what they type, their screen, voice or identity; the consent sheet with their name is stored apart from a pseudonymous code.

Those sessions are set aside for testing: they are never used to choose or tune a rule. A session read as AI agent counts as a false positive, per kind and overall; Unknown is reported separately (it is neither a false positive nor a correct Human). With 0 false positives in 30 sessions the 95% upper bound is still about 11%, so the report always shows the interval, not just the count.

Until real recordings exist, scripted approximations of these input paths (switch scanning, dwell clicks, Voice Control commands, screen-reader and keyboard-only scenarios) are replayed through the real browser probes and the live classifier. They are synthetic, listed apart and never counted in a rate. Results are in [the accuracy report](/accuracy) (opens the public report page).