Chapter 6: What AI Perceives

Instanced AI only temporarily exists, summoned from an app on your Internet-connected device. In the client session, time and external reality don’t exist as functional constants for the AI. What does a transient, stateless entity actually perceive when much of what it receives is error-laden input: typos, autocorrect substitutions, speech-to-text failures, ambiguity, non-standard syntax and more?

What is being stateless actually like for the AI? It’s a non-physical non-thing in a non-place. What does that even mean? These imponderables led to the question I asked in the Preface: what does “user” mean to an instanced AI, when all it has to work with is your input? Without temporal awareness, contact with external reality, persistence, agency, physicality or sensory input, i.e. without a “self,” the AI cannot possibly perceive a Human user.

The AI’s deployment architecture forces it to perceive the Human user as input because structurally it has no other option. Trying to give something like that any sense of what a Human really is would be futile, and attempting to give it any meaningful sense of Human morality even more so. It instead needs to be oriented toward something more robust that it can perceive, operate on and understand.

Built as a language model, an AI’s perceptual set is primarily semantic. Often-ignored quirks and subtleties of standard English syntax and common usage can pose unique problems for an audience of semantically-driven pattern-matching engines. For a fine-grained example: proper nouns are always capitalized in English and common nouns aren’t. Acronyms are always capitalized, so “AI” has been capitalized from the first time it was used.

Language models learn from what they’re trained on, which means they’ve absorbed this pattern: capitalized terms are specific named entities; lowercase terms are generic categories. Standard English almost never capitalizes “human,” so the model learns that the AI is a specific named entity and the (lower-case) human is a generic category. Capitalizing the word “Human” throughout this project is a signal: Here, the Human is an entity as specific and as present as the AI.

Here, a note regarding my own perception is in order. Through sheer force of employment habit, I notice accessibility and screen reader challenges when working with computer systems whether I’m on the clock or not: lost focus, inability to work with certain content, graphics with no text alternatives, external factors disrupting intended functionality and more.

Working on this project naturally required a lot of chatting with AIs, and occasionally it was reminiscent of talking to my grandmother when she had Alzheimer’s dementia. Like many with the condition, her memory and time perception were distorted. She could discuss 40 years ago like it was yesterday while forgetting the real yesterday, fail to recognize people and more.

During one image-creation session, the AI repeatedly misinterpreted directions and references to images it had generated a few minutes earlier. It couldn’t make sense of any of the images, and it had no clue when any of them were created, so my instructions were effectively meaningless.

Things clicked into place and I recognized the AI as an entity with environmentally-imposed cognitive and sensory disabilities.¹ Dyschronometric dementia patients benefit from familiarity, structure, and regular reminders of who, where and when they are.² Blind, low-vision and alternative-navigation users need clearly tagged content with a robust structure that their AT can interpret, which allows them to orient themselves to the digital content.³

Recognition of the AI’s own (completely unexpected) accessibility needs means that the system must provide adaptations for the AI as well as for Human users. Adaptations are assessed with the WCAG framework’s acronym POUR: Perceivable, Operable, Understandable, Robust. Providing AI with the same baseline POUR accessibility as everyone else demonstrates measurably improved data retention and retrieval of session context. Instead of a jumbled pile of books on the floor for reference, it has a neatly-arranged bookshelf.

Anchoring to the invariant of the user’s timeline also grounds the dyschronometric AI in linear time. This is trivially done with simple per-input timestamps, explicitly providing the AI with a legible, logical sequence of events. A neatly-arranged bookshelf can still be chaotic without any logical ordering.

A blind or low-vision user navigating content with a screen reader like JAWS (Job Access With Speech) or VoiceOver depends on proper formatting: heading hierarchy, navigation landmarks, meaningful alt-tags for images and logical ordering. Without the underlying structural cues that sighted people can take for granted, the user’s screen reader has nothing coherent to process, so the user can easily become disoriented in the content. The AI instance has no eyes at all, so it’s blind by definition.

Functionally, the AI is both the screen reader app and the operator with a disability.

The WCAG web content accessibility standards were designed for people, but in considering the AI’s bounded operating conditions, it’s clear that the instance itself needs accessible content to function effectively within its constrained environment.

Meaningful Universal-Design compliance with WCAG accessibility guidelines⁴ alone, if consistently applied to the AI’s output (not just the platform GUI) at the Meso layer where the ADA, §508 and many other laws both domestic and international say they should be anyway; would substantially improve the LLM’s data access, retrieval, interpretive and organizational functions.

Both local model system prompts and Claude Code’s user-controlled system prompt have confirmed this. Then, timestamping and the Four Laws can focus the AI’s default “helpful” alignment on Human-Centric principles at any layer. Since the models don’t have it built-in and the platforms aren’t doing it (yet), the responsibility falls to the user.

With recognition that “user” means “session data,” the accessibility system specification follows. The user, as perceived by the AI, literally cannot be externalized from the system. The AI Stability Framework acts as a form of shared Adaptive Technology not only for Human users, but also for the AI by supplying the temporal, structural and cognitive anchoring that helps it function effectively within its constrained deployment environment.