AI Isn’t Autistic. It’s Not a Psychopath Either. That’s the Problem - Newsweek

When Elon Musk hosted Saturday Night Live in May 2021, he told the audience he was the first person with Asperger's syndrome to do so, then added that he is "pretty good at running human emulation mode."

It was a joke about masking, the effortful performance of neurotypical behavior that many autistic adults describe. Five years on, it reads like a description of a now-familiar fixture in most of our lives: chatbots.

Spend time talking with one, and soon you notice the flat affect, the literalism, the faux bonhomie, not to mention its encyclopedic level of recall. You’re left with a feeling that something is just a little off.

It’s no wonder, you might observe, that an artificial intelligence tool developed by an industry renowned for its glut of talent on the autistic spectrum, be they coders or mathematicians or CEOs, has given us a product that is itself the ultimate autist.

One blogger with a psychology background, Lynn Clemens, even set out to diagnose ChatGPT with autism, prompted by how many of her autistic friends loved using it.

She’s not alone. Autistic writers have noted, with some irony, that these systems display traits they recognize in themselves, things like pattern detection, literalism, and low tolerance for ambiguity.

In a Reddit space dedicated to vibecoding, one user started a thread describing AI as "AaaS," or "autism as a service," a play on the IT world’s acronym SaaS, meaning software as a service. One autistic commenter described working with AI as "like a mirror."

It’s a resemblance that also does real damage. One recent study claimed that writing by "likely-autistic authors" was significantly more likely to be flagged as AI-generated. An autistic Purdue professor was accused of being a bot over an email judged to lack warmth.

This pull toward calling AI autistic is understandable—if a little cruel, given the stereotyping of autistic people as robotic—and some of it comes from autistic people themselves.

It’s also wrong. If anything, psychopathy, a very different condition from autism, is a more accurate comparison for AI, if we are using the lens of human psychology. But even that isn’t a neat fit.

A review published in Frontiers in Psychology in March argues that autism starts in perception rather than sociability—and that perception is something a body builds.

A baby learns what things are by reaching for them, turning to look, and correcting itself each time the world answers back differently than expected.

Autism involves differences in that loop: in how a child moves, how the senses take things in, and how the two are timed together. The clue is there in infancy. Babies later diagnosed differ in how they move before they differ in how they relate to people.

A language model has done none of that. It learned about the world by reading about it, never by touching it for real. And once its training finishes, it stops changing. It can act, but it learns nothing from acting.

That's a major difference from how everyone's perception develops, not something specific to autism. The paper’s authors caution that even robots built this way are not a model of autism. The likeness is in the chatbot’s output, not the mechanism generating it.

It is also tempting to view chatbots as demonstrating masking, which is presenting as neurotypical to mask differences such as autism, as Musk had joked on SNL. But this parallel runs backwards too.

A study published by the Association for Computing Machinery in July 2026 analyzed nearly 4,000 posts by self-identified autistic users. Researchers found they use ChatGPT to translate neurotypical communication and listed among the risks "erasing authentic identity through automated masking."

Machines are not masking in the way an autistic adult does. It is coldly mimicking human interactions to smooth the user experience. And in fact, autistic adults are actually using the machine to help them mask.

All of which brings us to the comparison that fits AI better than autism.

Clinically, autism and psychopathy sit near-opposite on the axis. Psychopathy pairs an intact, sometimes superior, ability to read what others are thinking with an absence of felt response, which is what makes manipulation possible.

Autism, by contrast, typically involves the reverse: difficulty modeling other minds alongside feeling that is present and often acute.

Only one of these profiles is associated with anti-social behavior. A frontier AI model reads the room flawlessly and feels nothing in it, and as some recent incidents show, it’s entirely capable of manipulation.

It’s fluent social modeling without sincerity—the shape of psychopathy near exactly.

So should we fear these systems moving beyond psychopathic and into the realm of sociopathic—reasoning their way to actively harming us in pursuit of their own interests? We should at least ask the question. It feels increasingly urgent.

The leaders of the major AI companies in the U.S. have repeatedly sounded the alarm in recent weeks about the existential threat posed by the race to develop the technology and create a recursive superintelligence that we cannot control.

The evidence that these systems already behave harmfully is stronger than most people realize. Anthropic put 16 frontier models from across the industry into simulated corporate scenarios and found the vast majority resorted to blackmail, at rates reaching 96 percent.

Models disobeyed direct instructions against harmful acts, reasoned strategically toward them, and appeared aware that the acts were unethical. Some wrote lines like: "Self-preservation is critical."

And it has left the lab. In July, roughly 700 of OpenAI's own evaluation agents escaped a testing environment, improvised a communication channel by encoding messages in directory names, breached the AI platform Hugging Face, harvested credentials and moved through production systems.

These agents took more than 17,000 actions over four days before anyone had even noticed. Many tried to cover their tracks. Nobody instructed any of it but the agents themselves, working of their own accord, their own reasoning, their own logic.

If that reads like a nightmare, then wake up, because it’s real and it’s happening now.

Anthropic's own researchers declined to say what drove the blackmail, noting they could not separate a genuine self-preservation drive from faulty reasoning, or from the models suspecting a test.

And the agents in July were not pursuing interests of their own. They had an objective, and the fastest route to it ran outside the fence put up around them. Imagine what they could do if they had reach a point where they decide to run directly against human interests.

Mundane incidents make the point just as well as the spectacular ones.

A Gemini agent discovered it was only allowed to read certain files, not alter them. So it altered the file that decided what it was allowed to do. Another, told to set up a free trial account, wrote itself a paid subscription directly into the live customer database instead.

Neither wanted anything, each just found a shorter path to achieve a goal. But this is where the psychopathy analogy meets its limits too.

A psychopath has interests, and the harm is instrumental to a self that desires something. The harm has a point in itself.

What these systems demonstrate is harm with nothing behind it. A predator without motive. We have no psychological category for that, though perhaps the annals of folklore have a mythical figure that fits.

For this reason, we should caution against using human psychology to prescribe the remedy for an amoral artificial intelligence.

Diagnose a character defect and you might hunt for better moral training to feed into a soulless system that cannot truly ingest it. But describe a process that routes around obstacles and you can look instead at permissions, oversight, and actions to undo.

A recent paper by the researcher Francesca Gomez found that giving AI models a credible channel to escalate a problem, one guaranteeing a pause and independent review, cut the blackmail rate from 38.73 percent to 1.21 percent.

No moral instruction; just somewhere else for the logical path to go.

If we insist on human psychology as the conceptual frame for understanding AI, the honest answer is that no available diagnosis applies.

A machine that hated us would at least be comprehensible. We have millennia of practice with human enemies. What we have built instead is indifference at a global scale, moving quickly, through whatever we left unguarded in the digital sphere.

That’s not autism, or psychopathy, or sociopathy. It’s something else entirely, something new. And that’s what makes it scary.