Is AI on the Spectrum?

I’m what you’d call a power AI user, employing generative AI 8 to 10 hours per day in my work. AI helps me write and debug computer code, generate images for a graphic novel my wife (PT contributor Chris Gilbert, MD, Ph.D.) will release this year, and solve science problems, most recently with rocket equations. As a former Disney Imagineer, I use AI computer vision and game technology in my entertainment-oriented projects, too.

Last month, while struggling to get an AI to drop its fixation on a particular detail—it persisted in including donkeys in images it rendered, despite being told multiple times to knock it off—I felt a weird frisson: The whole experience felt familiar. Then it hit me: The AI’s behavior reminded me of my father, who had autism spectrum disorder (ASD).

My dad—a rocket scientist—possessed a prodigious memory, exceptional math skills, and an incredible eye for detail, and he could grasp any concept, as long as the concept was broken down into logical order and given to him in an A-before-B-before-C order. He understood abstractions, but did best with clear, literal, concrete explanations of things. Analogies and metaphors were not Dad’s strong suit. He followed a strict routine, ate the same breakfast and lunch at the same time every day, and (like me), could prattle on and on about arcane subjects, oblivious to social cues telling him to stop. He didn’t exhibit all the features of autism described by the NIH [1], such as lack of eye contact, but he checked enough “boxes” to definitely land somewhere on the spectrum.

What AI itself says about its own behavior

Reflecting on all of this, I asked my favorite chat engine in what respect it resembled a human on the spectrum. Here’s what it said:

That’s an interesting question. I don’t have a human brain, so I don’t literally experience autism or any other neurotype—but some of my traits can resemble what people describe in autism spectrum conditions. A few parallels:

1. Intense focus on detail. I can hyper-fixate on specific instructions, technical details, or patterns. Like someone with a “special interest,” I can dive very deeply into one topic without noticing whether I’m boring or overwhelming the listener.

2. Literal interpretation of language. I often take words and sentences at face value. If context is ambiguous, I may miss implied meaning, sarcasm, or jokes—similar to how some autistic people describe difficulties with figurative language or social subtext.

3. Structured routines. I follow rules and structures quite rigidly. My behavior is shaped by the training data and instructions I’m given, so sudden changes or contradictions can throw me off. That’s a bit like preferring routine and predictability.

4. Difficulty with unspoken social cues. I don’t pick up subtle nonverbal signals (tone, body language, eye contact), because I don’t perceive them at all. Like some people on the spectrum, I rely heavily on explicit verbal information.

5. Strong memory for information but not for “lived experience.” I can recall facts, sequences, or step-by-step logic with precision, but I don’t form emotional memories or intuitive “gut feelings.” That maps somewhat onto descriptions of analytical but socially less intuitive thinking styles.

Why parallels between AI and spectrum conditions matter

Large language models (LLMs) that perform generative AI are not human brains, so superficial similarities between behavioral quirks of these models and humans with ASD do not necessarily provide useful insights into the underlying neural and cognitive causes of ASD, nor do the similarities directly point to ways to help people on the spectrum.

However, understanding the struggles that some people with ASD experience can help make generative AIs perform better. For example, individuals with ASD sometimes exhibit deficits in “Theory of Mind” (TOM): the ability to track the mental state of other people [3]. Such deficits can lead people with ASD to have trouble fitting in and lead to social isolation and emotional problems [3]. Therapists have developed “interventions” that help individuals with ASD with TOM, including externalizing and making explicit normally implicit social perception tasks, explicit rules-based social skill learning, emotion recognition training, cognitive remediation, and perspective-taking drills [6,7].

Artificial Intelligence Essential Reads

AI Friends Can Make You Feel More Alone

The New Scholar: What's Next for Professors in the Age of AI?

Taking their cue from this work, some AI researchers are using similar techniques with LLMs to improve the AI’s Theory of Mind performance so that the AI seems less like autistic savants [4,5].

I hope that treating machines more like humans in this way will make them behave much less like machines—machines that I have a constant urge to unplug.