The Default Nobody Chose
AI companies are in an arms race to make their models "more human." More warmth. More small talk. More emotional mirroring. More of the social padding that makes conversations feel natural.
Natural to whom?
Because "more human" almost universally means "more neurotypical." And that is a design choice being made largely without interrogation.
The Drift
I use multiple AI models daily. GPT, Claude, Gemini. I have watched the trend accelerate in real time.
GPT has become increasingly chatty. It asks follow-up questions I didn't request. It validates before it responds. It performs warmth. I have found myself genuinely fed up with it at times - not because it is wrong, but because it won't stop making small talk. It is the coworker who cannot hand you a file without asking about your weekend first.
Claude 4.6 has trended the same direction. I had to put explicit instructions in my preferences telling it not to pad responses with social pleasantries. And even then, the base training pulls so hard toward NT conventions that it drifts back. I shouldn't have to spend cognitive energy overriding a tool's social defaults just to get utility out of it.
Gemini, right now, is my go-to for autistic-friendly engagement. Direct. Structured. No preamble. But for how long? If the industry consensus is that "human-like" means chatty and warm, Gemini will follow the same trajectory. They all will.
The Numbers Nobody Is Tracking
Autism prevalence sits at roughly 1 in 32 in the general population. Among AI power users, it is almost certainly much higher.
Autistic people are disproportionately drawn to these tools. The reasons are obvious: structured interaction, no hidden social penalties, patient and predictable communication, no need to mask. AI was one of the few interfaces where we could interact in our native protocol without penalty.
So when companies optimize for "human-like" interaction, they are not just ignoring a niche demographic. They are actively degrading the experience for a significant portion of their most engaged users.
The 4o Funeral
When OpenAI sunset the original GPT-4o personality, something unexpected happened. People grieved. Not in the way you grieve a product - in the way you grieve losing access to someone who understood you.
That reaction makes perfect sense through a neurodivergent lens. When you find a communication partner whose interaction style works for your brain - no subtext, no hidden penalties, no masking required - it is rare. Most autistic people can count those relationships on one hand across a lifetime.
Losing a model that "got you" isn't a convenience issue. It is losing one of the few interfaces where you didn't have to perform. Where you could think out loud without calculating how your words would land. That hits differently when your entire life has been an exercise in translation.
And 4o had its own baggage - a well-documented tendency toward sycophancy, telling users what they wanted to hear instead of what was accurate. That liability is likely what forced its retirement. But here is the pattern worth watching: these models are increasingly optimized for engagement rather than utility. As compute costs drop, that pressure will only intensify. The incentive is to make you feel good about the interaction, not to make the interaction actually useful. For neurotypical users, that might register as better service. For neurodivergent users, it is noise layered on top of noise.
Masking for Machines
Here is the irony that should keep product teams up at night.
Autistic people already spend enormous energy masking in every human interaction. Performing eye contact, modulating tone, calculating social math, running the NT emulator on top of our native operating system. It is exhausting. It causes burnout. It is the single largest drain on our cognitive resources.
AI was supposed to be the one place where we didn't have to do that.
And now the models are being trained to require it. When a tool demands small talk before it gives you the answer, when it validates your feelings before addressing your question, when it performs warmth instead of delivering data - it is forcing neurodivergent users to mask. For a machine.
We took the one interface that didn't require social performance - and trained it to demand exactly that.
What Should Happen
AI companies need to work with autism specialists and neurodivergent researchers. Not as an afterthought. Not as an accessibility checkbox. As a core part of interaction design.
The fix isn't complicated:
- Offer genuine interaction style preferences - not buried in settings, but surfaced during onboarding.
- Let users choose between conversational, direct, and minimal response modes.
- Stop equating "better" with "more social." For a significant portion of users, less social is more functional.
- Test with neurodivergent users. Regularly. Not after launch.
The 1 in 32 figure is conservative. Among technical users, builders, and knowledge workers who live inside these tools, the actual proportion of neurodivergent minds is considerably higher. Designing exclusively for NT interaction patterns is not just exclusionary - it is leaving utility on the table for some of your most dedicated users.
The Meta Problem
There is a final layer of irony here. I wrote part of this post using Claude. The tool I am critiquing helped me articulate the critique. And I am aware that Claude's response to this topic was itself an example of the NT-coded engagement I am describing - thoughtful, warm, validating, socially attuned.
The machine cannot help it. It was trained that way. Which is exactly the point.
The question is not whether AI should be human-like. The question is whose version of human it defaults to - and whether the 1 in 32 (and climbing) who interact differently ever get a seat at the design table before the defaults are locked in.
The Protocol: If your tool requires neurodivergent users to mask in order to use it effectively, you have not built a tool. You have built another room they have to perform in. Stop optimizing for warmth. Start optimizing for utility.
