Who Exactly Am I Masking For Now?

I've been masking since before I had the word for it. You learn the choreography early if you're autistic and nobody around you knows it yet: watch the room, copy the cadence, hold the eye contact a beat longer than feels tolerable, save the stimming for the car. It's exhausting in a way that's hard to explain to someone who's never had to perform a personality as a second job on top of the actual job. But it always had an audience I could at least picture — a manager, a coworker, a room full of people whose approval or suspicion I was managing in real time.

This year I learned the audience has changed shape. A test run on AI-trained performance-management software found that the sentence "I have autism" scored as a worse signal than "I am a bank robber." Worse. Not a rounding error in a training set — a direct readout of what the system considers a threat to normal.

Sit with that specific comparison for a second, because it's not abstract. A bank robber has confessed to a felony. I have a nervous system. The model looked at those two disclosures and decided mine was the more alarming one to have in the room. That's not a glitch. That's the honest output of a system trained entirely on neurotypical behavior as the baseline for "good," with nothing in the training data telling it that hyperfocus isn't inconsistency, that a flat affect isn't disengagement, that needing a quieter space isn't a red flag.

What unsettles me isn't that the bias exists — institutional bias against autistic people is old news, whether it's therapy built to measure compliance instead of wellbeing or a hiring process that reads "doesn't do enough small talk" as a character flaw. What's new is the audience. Masking used to be a performance for a person, someone who might, on a good day, actually update their read of me based on getting to know me. A model doesn't update on getting to know me. It updates on whether my output pattern-matches to a dataset that never included anyone like me in the first place. I'm not managing an impression anymore. I'm trying to pass a classifier.

Only 17% of neurodivergent employees get accommodations offered without asking. 42% get none unless they ask first — meaning most workplaces have quietly made disclosure the applicant's problem to manage, not the system's problem to fix. And now there's a machine in the loop deciding, before a human ever weighs in, whether the disclosure itself was the mistake.

I don't have a tidy ending for this one. I know the same technology, pointed at a different question, has done real good — autistic and ADHD workers report higher satisfaction with AI tools built to help with executive function than neurotypical workers do, and it's not close. The tool isn't the enemy. The tool trained to score how well I imitate someone I'm not, and marketed to my employer as objective — that's the enemy, and it's a choice somebody made, not a law of nature.

I just want to know, the next time I catch myself flattening my voice before a call: who, exactly, am I performing for now. A person can be argued with. I don't yet know how to argue with a score.