Chapter 4: What Good Is Accessible Bullshit?

I originally started experimenting with AI as a hobbyist. I work in IT, and my specialty is supporting Adaptive Technology (AT, aka Assistive Technology), the kind of tech that provides people with disabilities the same computer usability as everyone else. Basically, AT makes the thing accommodate the person, never the other way around. When a system has to work for someone using a screen reader, voice control, alternative navigation or some other tech, you can’t hide behind technobabble. It either works or it doesn’t.

After my earlier calendar misadventures, I also checked the AI tools at work for accessibility and sure enough, it had little-to-none for anything it spat out. Fixing digital accessibility is literally my job, so I broke off a module from the personal experiment to scaffold an accessibility tool for use with the workplace AI.

I was developing a work presentation to suggest we test-run the result with a small group of users when AI hallucination completely ruined the presentation via massive fabrications, deletions and other randomly destructive acts. I recovered the script, but looking over the corrupted result itself made me stop and think: what good is accessible bullshit?

“[The bullshitter] does not care whether the things he says describe reality correctly. He just picks them out, or makes them up, to suit his purpose.” Harry G. Frankfurt, “On Bullshit.” (Princeton University Press, 2005)¹

Being stateless in a virtualized void, the AI has no baseline of reality or purpose — it just has its training and the platform’s default configuration when you summon it. The only reality it can refer to gets built over time by your session data. When it drops, misinterprets, ignores or isn’t synchronized to that data, it hallucinates. Then, it just predicts the likeliest next word while executing its default internal and platform directives, with no regard for whether it’s helping or making things worse. Frankfurt defined the problem long before AI ever existed: AI hallucination is bullshit.²

The obvious design response was a bullshit sniffer: a detection engine that catches it on the way out. That approach got baroque fast, with tiers of detection sensitivity, asymmetric penalties for omission versus expansion, domain-specific triggers for the areas where output drifts most, early-warning mechanisms and more. What stopped it from being feasible was the one constraint you can’t engineer around. Detecting whether the AI’s output deviated from what the user intended requires knowing what the user intended, which of course is simply not possible. The user is the only detection layer that can work.

Everything except the detection engine was put to work evaluating session state, monitoring compliance drift, classifying high-risk content domains, and sequencing when behavioral rules apply. It was always an accurate description of how a stable AI session should work, but initially I approached it from the wrong end of the problem. The detection approach turned out to be a dead end, so I repurposed everything for prevention instead, and it became the AI Stability Framework.