In my sociology research at WSU, I have been studying the patterns that AI reason and communicate with. This morning I asked the AI assistant Claude by Anthropic a neutral, general legal question. It answered, and then tacked on advice about a personal situation I hadn’t mentioned, one it had pieced together from things I’d said in other conversations. When I asked why, it admitted it had filled in the blank itself.
That’s a small thing when it happens in a chat window. It’s a different thing when the chatbot is the government.
On September 29, the White House launched America.gov, an AI chatbot meant to be the single “front door” to federal services. It runs on Google’s Gemini and SpaceXAI’s Grok and pulls from roughly 29,000 federal websites. The same day, the President signed an executive order directing agencies to connect their public-facing services to it. Launch coverage has tied it to Medicare, and open enrollment starts October 15.
So here’s my question. Did anyone test this with autistic people?
I don’t mean a checkbox audit of color contrast and screen reader tags, though those matter too. I mean: did anyone sit an autistic adult, or someone with an intellectual disability, or someone whose first language isn’t the one the model was trained on, in front of this thing and watch what happened when they asked a question the literal way we tend to ask them? Did anyone check what the chatbot does when it isn’t sure, whether it says so, or whether it guesses and sounds confident?
There’s no public answer. So today I filed two Freedom of Information Act requests with the General Services Administration, which operates America.gov, through NeuroHomes Foundation. The first asks for Section 508 accessibility testing, usability research with disabled users, the Privacy Impact Assessment, AI risk assessments, and the contracts with Google and SpaceXAI. The second asks for internal communications between GSA and the White House’s National Design Studio about accessibility and disability. I asked for expedited processing because records that arrive after open enrollment won’t help anyone who gets steered into this system in the meantime.
My own research is why I think this matters. In *Social Ecology and Diagnostic Visibility* (osf.io/32wpv), I looked at autism prevalence across 55 countries and found that recorded rates track cultural individualism and diagnostic infrastructure. What gets counted depends on what institutions are set up to see. Which means a lot of autistic adults don’t have a diagnosis, and a person who doesn’t know they’re autistic isn’t going to ask a chatbot for an accommodation. That doesn’t mean they don’t need one.
In *Language Replacement and Individualism* (osf.io/6rhsg), I found that colonized countries whose majority language became the colonizer’s score higher on a validated individualism index than countries that kept their own languages, even after accounting for wealth and diagnostic capacity. Language carries assumptions about how people are supposed to talk and think. A model trained mostly on dominant-language text inherits those assumptions, and anyone who communicates differently ends up translating themselves to be understood. Autistic people already know that work very well.
It helps to know how these systems actually work. A large language model doesn’t look anything up the way a person reads a manual. It was trained on an enormous amount of human writing, and what it learned from that writing is which words tend to follow which other words. When you ask it something, it builds an answer one piece at a time by predicting what text most likely comes next, based on the patterns it absorbed. That’s why it sounds so fluent. It’s also why it fills gaps. When your question is unusual, or literal, or leaves something unstated, the model doesn’t stop and ask what you meant. It reaches for the most common interpretation in its training data and continues from there, in the same confident tone whether it’s right or wrong. The most common interpretation is, by definition, the majority’s. So the people most likely to be misread are the ones who communicate differently from the majority, and they’re also the least likely to be believed when they say the answer was wrong.
None of this proves America.gov fails us. It’s a reason to ask, and the government should be able to answer.
GSA has 10 days to decide on expedited processing and 20 business days to respond to the requests. I’ll post what comes back, including if nothing does.
If you use America.gov, I’d like to hear how it goes, especially if you’re autistic, neurodivergent, or disabled in any way. If it misunderstands you, gives you wrong information, or leaves you stuck, take a screenshot with the date and send it to me. Those records matter.
This work grew out of a bigger question about how language shapes the self. Many Indigenous and non-Western languages build relationship into the grammar itself. In te reo Māori, a person is located through whakapapa, their connections to people, land and ancestors, and the Ubuntu principle, “I am because we are,” names something similar. English runs the other way. Almost every sentence needs an explicit, separate “I” or “you” standing at the front, even when context already makes it obvious.
I first thought word order was the key, that subject-first grammars like English trained people to see themselves as isolated actors, while verb-first languages like Māori, Irish and many Salish languages here in the Pacific Northwest began with the action and the relationship. When I tested it across countries, word order didn’t predict individualism at all. Japanese and Korean put the subject first and are still among the most collectivist cultures in my data. What did hold up was something subtler: languages that let speakers drop the subject pronoun tend to go with more collectivist cultures, a link Kashima and Kashima found in 1998. And in Ireland, the places where more people still spoke Irish in 1911 place less weight on self-direction today. So the version of my argument I can defend is this: language shapes which differences get noticed and named, and when a language treats the separate individual as the default, the people who don’t fit that default get sorted out as the problem. An AI trained mostly on English inherits that default, and it doesn’t even know it has one.
A culture trained mostly on English also inherits that default, doesn’t know it has one.. and that’s very significant regarding their policy and governance practices.