Neurodivergent folks are facing a new barrier in education and work, and we need to talk about it.
The issue
I have now come across several instances where neuro-atypical individuals, especially but not exclusively autistic ones, have been told that their material is AI-generated, even though they were the unaided authors of their work.
This happened on job applications, course work, online contents and more.
So why is this happening?
There is an overlap between AI and autistic communication styles:
- Both use accurate language: to a neurotypical reader or listener, this can seem too literal, too formal, even pedantic;
- Both focus on ‘report’, not ‘rapport’: texts are highly structured, devoid of fillers and social fluff, and focussed on information sharing;
- In speech, both rely less on voice inflection to express the author’s emotion or intent, making it ‘sound robotic’ to neurotypical listeners.
Now let’s have a look at how AI detectors work:
- They perform stylometry, the analysis of a writer’s linguistic fingerprint;
- In particular, they measure perplexity (the unpredictability of a text) and burstiness (the variation of sentence in length and structure);
- And they look for specific signs of humanness, like typos, stylistic choices and emotional nuance.
The problem is that they have only been trained to contrast AI output with neurotypical text, without considering the neurodiversity of human communication.
Ergo, to an AI detector, the combination of predictable grammar, formal vocabulary register, defined structure, being a stickler for spelling, and a rigorous style without emotional coding, is not human.
Yet, to neuro-spicy writers, these can simply reflect their natural inclination towards accuracy, elegance, rigour, patterns and rhythm.
Why it matters
Aside of the questions this raises around the policing of expression styles and the de-humanisation of anything that sounds vaguely different, the ‘computer says so’ approach to AI detection has major immediate impacts on neuro-atypical people.
- Students saw their contributions dismissed, and they were stigmatised as cheats (and as liars when they attempted to defend themselves);
- Job seekers were deprived of interview opportunities when their applications were wrongly flagged as AI-generated.
Autistic and other neurodivergent individuals already face a long list of challenges in education and the workplace; this latest, pervasive AI-related addition is compounding the problem.
What can we do about it?
If you are serious about neurodiversity inclusion in education and work, it is essential that you keep this issue in mind and that you take pre-emptive steps to correct this form of discrimination.
In recruitment, do not rely solely on AI to parse applications. And work closely with hiring agencies to understand how their tools can impact outcomes for neurodivergent folks (don’t be surprised to find that they are not aware of the issue, or not actively taking steps to address it).
In education, get to know your students, their speech patterns and writing styles. Once this baseline is clear, you’ll be more able to spot actual AI fraud. Also, if your AI detector systematically flag a genuine student writer as AI, note this in their academic file, so that they have historical evidence to support their appeal when their examination pieces fail for the same reason.
Finally, if you are an AI developer, please actively work at making AI more inclusive. While it is a wonderful tool, it is only as good as the parameters we program it with. Please teach it about the beautiful neurodiversity of the world out there.
#DEI #EDI #inclusion #autism #neurodiversity #AI
