Why Half Think AI’s Revolutionary & Half Thinks It’s Bullshit
Two people sit down in front of the exact same system. They spend roughly the same amount of time with it. They walk away with completely opposite conclusions about what they just experienced.
One thinks it’s revolutionary. The other thinks it’s a glorified search engine. Both are confident. Both have examples. And if you put them in the same room, they will each privately wonder how the other person could possibly be this wrong about something so obvious.
At this point the debate has hardened into two predictable camps. “AI is about to change everything.” “AI is overhyped autocomplete.” The arguments on both sides have calcified. People have picked their team. And both sides are missing something important, because the divide isn’t really about expectations or hype cycles or which industry analysts you follow. It’s about something more structural than that. It’s about how people interact with the system itself.
— — —
The Tool Assumption
Most software works like a machine, and we’ve spent forty years learning to interact with technology on exactly those terms. You give it a command. It executes a procedure. You get a result. Press button, get deterministic output. That model is so deeply wired into how we think about software that most people carry it into their first AI interaction without even noticing they’re doing it.
Large language models don’t work like that. They generate responses by navigating probability across enormous pattern spaces, which means the result you get depends not only on the question you ask, but on how you explore the system after that first answer appears. Ask once and move on, and the output often looks average. Probe, refine, iterate, reframe, and push back, and the system’s capabilities start expanding in ways that aren’t visible from the first interaction.
This is the thing that almost nobody talks about clearly: people think they’re using a tool. What they’re actually interacting with is something closer to a probabilistic reasoning partner. And that distinction matters enormously, because the mental model you bring to the interaction shapes almost everything about what you get out of it.
— — —
Interaction Fit
Once you see that, the divide starts making a lot more sense. Some users approach AI like traditional software: ask, evaluate, move on. They treat each response as a deliverable rather than a starting point. Others approach it like a collaborative thinking process, probe, refine, test, reframe, recurse. They’re not looking for the system to hand them an answer; they’re using the system to think.
The second pattern extracts dramatically more capability from the same system. Not necessarily because those users are smarter, but because they’re interacting with the system in ways that align with how it actually generates responses. The interface is identical. The interaction model is not. Which means the capability people experience isn’t uniform across users. It’s partly a function of cognitive interaction style, of whether the way you naturally think happens to map onto the way these systems work.
Which is where things get genuinely interesting.
— — —
The Edge That Reveals the Mechanism
You can see this dynamic most clearly at the extremes. Some users don’t just find AI useful, they experience it as unusually natural, almost disorienting in how well it fits. Talk to enough heavy users and a pattern starts surfacing: certain neurodivergent cognitive profiles, including ADHD, dyslexia, and some people on the autism spectrum, often report an unusually strong alignment with these systems. Not minor productivity gains. Something closer to cognitive amplification, a sense that the tool finally works the way their brain already does.
The language that shows up repeatedly is striking. “This is the first tool that works the way my brain works.” “It feels like an external processor for my thinking.” That framing, not “this helps me” but “this fits me”, has shown up often enough that researchers have started investigating it systematically.
Early human-computer interaction studies have found that autistic workers often prefer AI-mediated assistance for certain workplace communication tasks, suggesting the systems can reduce friction in areas where many autistic professionals traditionally face the most resistance.² Separate research into AI writing tools for adults with dyslexia found that large language models significantly helped participants organize ideas, structure writing, and overcome the blank-page problem that so often accompanies dyslexic writing challenges.³ Follow-up work using LLM-assisted readability tools demonstrated measurable improvements in reading performance, particularly for participants with more severe reading difficulties.⁴
None of this establishes a universal rule. It doesn’t mean neurodivergent users are inherently better at AI. But it does point at something real: certain cognitive interaction styles may align unusually well with how these systems operate. And understanding why helps explain the divide.
— — —
Why That Alignment Is Plausible
Decades of cognitive research describe characteristic tendencies associated with ADHD and dyslexia, divergent thinking, associative reasoning, pattern recognition, exploratory problem-solving, that can be significant liabilities in environments built around rigid procedural execution.⁵ ⁶ The same tendencies map surprisingly well onto how large language models actually reward interaction.
LLMs aren’t optimized for the person who asks once and accepts the first answer. They respond to iteration. Reframing. Exploration. Thinking out loud and letting the response reshape the next question. The cognitive style that struggles most in assembly-line knowledge work, the mind that resists staying in the lane, that keeps finding unexpected connections, that works by circling rather than by marching forward, may be the cognitive style that unlocks the most from these systems.
Which means the interaction advantage isn’t really about intelligence or technical sophistication. It’s about whether the way you already think happens to align with how the system generates responses. Some people get that alignment for free. Others have to work against their natural habits to find it. And most people, encountering the technology for the first time, have no idea which situation they’re in.
— — —
Edge Cases Make Mechanisms Visible
I’ve written about my own experience with this elsewhere, the last year produced an arc of work I would have considered flatly impossible before, and the story of how that happened is worth reading if you want the personal version of this argument. You can find it here. But the anecdote isn’t the point. The point is what edge cases do.
When a technology produces wildly uneven outcomes, when some users find it transformative and others find it useless, the edges reveal the mechanism. They show you what the system is actually responding to. And right now the edges of the AI divide are showing something important: this technology doesn’t just automate tasks. It interacts with cognition. It responds differently depending on how you think. And when the architecture of the system and the architecture of a particular mind happen to line up, the results can look almost absurd to anyone watching from the outside.
That’s not a flaw in the technology. It’s a structural feature of how it works. And it reframes the debate entirely.
— — —
The Wrong Question
The current standoff asks: is AI revolutionary or bullshit? But that’s a question about the technology in the abstract, as if there’s a single answer that applies uniformly across all users and all interaction styles. There isn’t. The better question is: which kinds of thinking does this technology amplify?
Because once you start looking at the divide through that lens, the standoff stops being mysterious. People aren’t disagreeing about the same experience. They’re describing genuinely different experiences produced by genuinely different interaction patterns with the same underlying system. Both sides are right about what they encountered. They’re just not encountering the same thing.
And if that’s true, if the divide is real and structural and not just a matter of learning curve, it forces a harder question than anything the current debate is asking.
Because the cognitive styles that align naturally with AI aren’t evenly distributed across professions. Some roles are built almost entirely around the kind of exploratory, iterative, pattern-connecting thinking these systems reward. Others are built around something else entirely, execution, consistency, process fidelity. The qualities that make someone excellent at those roles are largely orthogonal to what AI amplifies.
Which means the divide isn’t just between people who “get” AI and people who don’t. It runs straight through the architecture of work itself. And the roles on the wrong side of that line aren’t there because the people in them lack capability. They’re there because the work they do was designed around a different cognitive contract entirely.
That’s where this gets uncomfortable. Not which jobs AI takes, but which kinds of work it structurally deprioritizes. And what happens to the people, and the organizations, built around them.
— — —
Sources
1. Large Language Model Mechanism
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., … Amodei, D. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems (NeurIPS), 33. https://arxiv.org/abs/2005.14165
2. AI-Mediated Communication Support for Autistic Workers
Valentine, M. A., Retelny, D., & Rahmati, M. (2024). Generative AI and Neurodiversity: Opportunities and Risks in Workplace Communication Support. Human–Computer Interaction / ACM CHI. https://dl.acm.org/doi/10.1145/3544548.3581514
3. AI Writing Assistance for Adults with Dyslexia
Gajos, K. Z., Bigham, J. P., & colleagues. (2023). AI Writing Assistance and Accessibility: Supporting Adults with Dyslexia in Writing Tasks. ACM CHI. https://dl.acm.org/doi/10.1145/3544548.3581398
4. AI-Assisted Reading Support for Dyslexia
Morris, M. R., et al. (2023). Let AI Read First: Using Large Language Models to Improve Reading Accessibility for People with Dyslexia. ACM CHI. https://dl.acm.org/doi/10.1145/3544548.3580908
5. ADHD and Divergent Thinking
White, H. A., & Shah, P. (2011). Creative Style and Achievement in Adults With Attention-Deficit/Hyperactivity Disorder. The Journal of Creative Behavior, 45(2), 87–110. https://doi.org/10.1002/j.2162-6057.2011.tb01090.x
6. Dyslexia and Pattern Recognition Strengths
Schneps, M. H., Brockmole, J. R., Sonnert, G., & Pomplun, M. (2012). History of Reading Struggles Linked to Enhanced Learning in Low Spatial Frequency Scenes. PLoS ONE, 7(4), e35724. https://doi.org/10.1371/journal.pone.0035724
Schneps, M. H., Rose, L., & Fischer, K. W. (2007). Visual Talent in Dyslexia: Seeing Forests When Others Only See Trees. Scientific American Mind. https://www.scientificamerican.com/article/the-advantages-of-dyslexia/
— — —