Chapter 4: What Good Is Accessible Bullshit?

I had started some occasional home-hobbyist experimentation with AI in Summer 2025, mostly rooted in curiosity. For a bit of context, when at my place of employment, I specialize in supporting Adaptive Technology (aka Assistive Technology or AT), 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.

Microsoft Copilot was later made available on my employer’s systems so I checked it for accessibility and sure enough, it had little-to-none for anything it spat out. 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, and the AI didn’t.

Instead of waiting for Microsoft to get around to caring, I decided to break off a module from my personal experiments to try bootstrapping an accessibility tool for use with the workplace AI. The result turned out better than expected and it worked to varying degrees with any AI platform (not just Copilot), so I started developing a work presentation to suggest we test-run the result with a small group of users.

Unnoticed until too late, AI hallucination crept in and irretrievably ruined the presentation via massive fabrications, deletions, substitutions, nonexistent references and other random damage. The tool itself survived, but looking though the corrupted presentation 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

The AI is stateless and timeless in its virtualized void, so architecturally it has no external point of reference. External reality is simply not available. It only has its training and the platform’s default configuration when you launch it. The only detectable proxy for reality that it can refer to gets built over time by your session data. When it skips ahead, misinterprets, ignores or isn’t synchronized to that data, hallucination is the expected result. Indifferent, 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.

The fundamental problem with a detection approach is that judging whether the AI’s output deviated from what the user intended requires knowing the user’s intent, which is simply not possible with software. The user is the only possible detection layer that can work. Platform disclaimers saying “AI can make mistakes” read as defensive blame-shifting because that’s exactly what they are, but they’re also not wrong.

A change of course was in order, so I put everything except the detection engine 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, I just started from the wrong end of the problem. The detection approach turned out to be a dead end, so I repurposed everything for prevention instead. You’re reading the result.