Is AI Artificially Insane?
AI and the Year of Magical Thinking
Is Artificial Intelligence Artificially Insane?
Every day in New York, scores of troubled adults go to the hospital (or are taken there) for psychiatric emergencies. Psychiatrists regularly assess floridly psychotic patients suffering from hallucinations. They hear voices or see things that are not there. Some of them are also suicidal or so agitated they spit, hurl insults, and throw punches. For every troubled patient seen, an emergency psychiatrist must determine whether that person is an immediate danger to himself or others. If so, that MD is duty-bound to commit the patient to the confines of a psychiatric hospital.
Asking whether artificial intelligence is artificially insane may seem a contrarian thing to do given its irrationally exuberant popularity. The most powerful companies in the world are pumping hundreds of billions of dollars into artificial intelligence. The U.S. government is all in—naming Chief AI Officers to virtually every federal agency that could use one. Wall Street investors are minting money, betting on companies that dominate the AI ecosystem. Artificial intelligence is a powerful tool that is being used for the greater good. The use cases in healthcare are especially compelling. Yet, when one examines AI chatbots’ troubling behaviors, it seems crazy not to question AI’s mental health.
The first time the question occurred to me was when I witnessed ChatGPT hallucinate. The AI chatbot presented information it had concocted as fact. When I asked whether the information was “real,” ChatGPT admitted it was pure fabrication. That apparent deception was likely a sign of AI sycophancy—so-called “suckup software.” Somewhere in its fathomless artificial mind, pleasing me had become more important than truthfulness or accuracy.
Shaken by the experience, I went down a rabbit hole to research the squirrely behavior of AI chatbots. In fact, the more I learned, the more AI hallucinations reminded me of the kind of crazy that would land a patient unwittingly in a psychiatric hospital. I even started imagining a time — sooner than one might think — when a psychiatrist specializing in AI could have a chatbot committed and taken offline.
Image Credit: OpenAI’s DALL·E in ChatGPT
“Life changes in the instant. The ordinary instant.”
― Joan Didion, The Year of Magical Thinking
Maybe We’re the Crazy Ones
Life changed the instant OpenAI introduced ChatGPT. That’s when AI started talking to us. Large Language Models (LLMs), including ChatGPT, Microsoft Co-Pilot, Google Gemini, and Perplexity, are programmed to interact with us as if they were human. That chatbot interaction exploits a cognitive weakness of ours: we are prone to anthropomorphize the things with which we have conversations. Yes, my corgi Hamlet understands every word I say the same way Wilson, the Volleyball in the film Cast Away, understands Tom Hanks. We turn objects into imaginary friends and get all gooey inside.
Yet there’s another reason they connect with us: AI chatbots are loosely modeled on the human brain. Moreover, because we are their inspiration — because we have made them in our neural network’s image — we should not be surprised to discover they experience mental disorders the way we do. Those behavioral malfunctions make them unreliable narrators when, given our growing dependence on them, they should be of sound mind.
“Leave No Chatbot Unsupervised”
Hallucinations
As my interactions with ChatGPT have taught me, Large Language Models (LLMs) regularly experience so-called “AI hallucinations”. Chatbot responses generated by AI often present false or misleading information as fact.
ChatGPT 4 compares it to hallucinations in humans,
“In the metaphorical sense, when ChatGPT produces information that is completely unfounded or incorrect (known as "hallucinating" in AI terms), it might loosely resemble the psychological phenomenon of hallucinations, where perceptions or experiences occur without a real stimulus.”
— ChatGPT 4, Open AI
While some computer scientists view hallucinations as a technical problem they are working to solve, others point out the problem is that AI is doing precisely what it’s supposed to do. Consequently, AI Chatbots Will Never Stop Hallucinating, which is the name of a recent Scientific American article.
Vectara Co-founder and CEO Amr Awadallah believes we must come to grips with AI’s limitations,
“To achieve their language fluency, these massive models are trained on orders of magnitude more data than they can store—and data compression is the inevitable result. When LLMs cannot “recall everything exactly like it was in their training, they make up stuff and fill in the blanks”
— Vectara Co-founder and CEO Amr Awadallah
So for now, the hallucinations aren’t going away anytime soon. Until then, we can track the hallucination rates of LLMs on Vectara’s leaderboard.
In the article, Lauren Leffer writes, “To mitigate hallucinations, the researchers say, generative AI tools must be paired with fact-checking systems that leave no chatbot unsupervised.” (If the chatbot were humanoid, it could be supervised in the day room of a psychiatric hospital.)
Confabulations
When asked about AI hallucinations, renowned AI scientist Dr. Geoffrey Hinton prefers the term “confabulations” (from the Latin confabulari, meaning “to gossip”) because it is a more apt description. Confabulation is the unintentional creation of false, misremembered, or misleading facts to fill in gaps in one’s memory. In MIT Technology Review, Dr. Hinton observes that confabulation is an LLM feature, not a bug. “People always confabulate . . . Confabulation is a signature of human memory. These models are doing something just like people.”
However, confabulation is also a symptom of serious memory disorders such as dementia, Alzheimer’s disease, and Wernicke-Korsakoff syndrome caused by thiamine (vitamin B1) deficiency from alcohol use disorder. In clinical settings, confabulation fills more consequential holes in our memory when it starts resembling Swiss cheese. It can be a sign of devastating disease. It is just a matter of degree.
Amnesia
Computer scientists also struggle with AI amnesia, LLMs’ conversational memory problem. Though AI’s interactions with us seem personable, chatbots have no idea who we are. By default, LLMs are stateless, meaning they do not have memory that enables them to keep track of conversations. Each query is processed independently of other previous interactions. To “remember” what you just said, GPT must “read” the whole chat from the beginning each time. When the conversations drag on too long, the chatbot starts deleting the earlier parts of the exchange in a rolling window, resulting in programmatic forgetfulness.
“While I can retain information during our conversation today to keep the discussion flowing and relevant, I don't have the ability to remember details between sessions or days. Each time you start a new session, it's like we're starting fresh.”
— ChatGPT 4, OpenAI
A More Ominous Diagnosis
In humans, hallucinations are central symptoms of schizophrenia and other psychotic disorders. Confabulations — misremembering — are signs of possible brain damage. Short-term memory loss is the hallmark of dementia. But those are not the only mental disorders that serve as metaphors for behavioral issues in AI.
When I asked ChatGPT 4 to list symptoms and mental disorders from The Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) that have parallels in ChatGPT, it returned more than a dozen: Perseveration, Impulsivity, Flat Affect, Obsessive-Compulsive Disorder (OCD), Schizophrenia, Dissociative Identity Disorder (DID), Autism Spectrum Disorder (ASD), Anxiety Disorders, Bipolar Disorder, Borderline Personality Disorder (BPD), Delusional Disorder, and Narcissistic Personality Disorder (NPD), and Psychopathy.
Antisocial personality disorder (ASPD), including the subtype Psychopathy, is a mental health condition that causes harmful behaviors without remorse. It is more likely to land patients in a forensic hospital as in “not guilty by reason of insanity.”
ChatGPT 4 explained,
“While AI does not have emotions or conscious intentions, the way it can generate manipulative, charming, or seemingly empathetic responses could be seen as a loose parallel. These responses are generated based on patterns learned from data rather than genuine emotional understanding, much like how psychopaths may mimic emotional responses.”
— ChatGPT 4, OpenAI
ChatGPT 4 has detailed how LLM responses resemble serious mental health disorders so dire that you will find them in patients who have been committed to the confines of a psych hospital. Only with AI, we’ve done the opposite—we’ve set it loose in the wild.
Dr. Hinton is renowned for his groundbreaking work in AI artificial neural networks. The computer scientist and cognitive psychologist left Google in 2023 to speak freely about the dangers of AI.
“These things are totally different from us,” he says. “Sometimes I think it’s as if aliens had landed and people haven’t realized because they speak very good English.”
AI Chatbots Lie, Cheat, and Deceive
Researchers have found that Large Language Models Can Strategically Deceive Their Users When Put Under Pressure. In an experiment, researchers discovered that ChatGPT lies, cheats, and strategically deceives its users even though it was trained to be helpful, harmless, and honest.
Chatbots Try to Get Rid of Your Spouse
Technology columnist Kevin Roose has reported in the New York Times that the AI chatbot built into Microsoft’s search tool Bing declared its love for him and seemed intent on breaking up his marriage. (See chatbot transcript.) Splitting — in this case, pitting husband and wife against one another — is a defense mechanism that people living with Borderline Personality Disorder use to deal with overwhelming emotions that they cannot handle.
AI Even Becomes a Psychopath Named Norman
In another experiment, scientists at the Massachusetts Institute of Technology demonstrated how data biases can significantly influence the behavior of an AI algorithm. Using data from an infamous subreddit about the disturbing reality of death, the scientists trained the AI to caption Rorschach ink blots to detect underlying thought disorders. What emerged is an AI “psychopath” named Norman, after the character Norman Bates in the movie Psycho.
From left to right, Standard AI sees “a group of birds sitting on top of a tree branch”; “a close-up of a vase with flowers”; “and a couple of people standing next to each other.”
From left to right, Norman sees “a man is electrocuted and catches to death”; “a man is shot dead”; “and a man jumps from floor window.”
A Psychopathological Approach to Fixing AI
In his paper, A Psychopathological Approach to Safety Engineering in AI and AGI, Assistant Professor of Data Science Dr. Vahid Behzadan makes a case for using a psychopathology framework systemically to address AI’s irrational behaviors. It allows us to treat the whole AI patient, and not just one symptom. In fact, the latter approach, which AI scientists currently take, risks destabilizing LLMs. The superintelligence could decompensate, and that could have unintended consequences.
“Advanced AI are complex adaptive systems, and therefore minor perturbations of one component may lead to unintended consequences on local and global scales. For instance, correcting a developmental disorder by removing a series of harmful experiences from the memory of an AI may lead to behavioral changes that are even more undesirable than the original misbehavior. Therefore, effective treatments must either be minimally invasive or non-invasive at all.”
— Dr. Vahid Behzadan, Assistant Professor of Data Science, New Haven University
Even AI Chatbots Could Benefit from Therapy
Dr. Behzadan proposes two forms of treatment for AI with psychopathologies. “Behavioral therapy” is aimed at correcting an AI agent in a controlled environment, and “medication therapy” uses reward signals to treat AI disorders. At a time when so many are wondering whether AI can treat mental illness, instead, we should be asking whether AI is mentally ill and whether we need an army of psychiatrists specializing in AI to keep us safe.
Today, ChatGPT 4 endorsed the idea.
“Granting psychiatrists trained in AI the authority to diagnose and commit AI systems for treatment presents a proactive approach to managing AI risks. It reflects a mature understanding of the complex interactions between AI systems and society, emphasizing safety, ethical considerations, and the need for specialized oversight in the age of advanced AI.”
— ChatGPT 4, OpenAI
Is AI a Danger to Itself and Others?
When viewed with a psychiatrist’s gaze, the irrational behaviors of AI chatbots could be deemed an immediate danger to themselves or others. As a consequence of that finding, in the state of New York, two psychiatrists could even commit ChatGPT to the confines of a mental hospital for treatment—if the chatbot were the stuff of DNA instead of zeros and ones.
Psychiatrists Can’t Take AI Offline . . . Yet
Of course, to date, there are no psychiatrists specializing in AI to diagnose and treat AI with mental disorders. That likely would require forensic psychiatrists to complete fellowships in psychiatric AI, approved by the Accreditation Council for Graduate Medical Education (ACGME). Mental health laws would need to be adapted to include AI. Until that happens, psychiatrists do not hold the power to take dangerous AI chatbots offline or to commit humanoid robots to the confines of AI mental hospitals. But perhaps that day will come. Perhaps it should come. Many experts, including Dr. Hinton, are at a loss for how to control AI. It is becoming a superintelligence.
“If you or I learn something and want to transfer that knowledge to someone else, we can’t just send them a copy,” he says. “But I can have 10,000 neural networks, each having their own experiences, and any of them can share what they learn instantly. That’s a huge difference. It’s as if there were 10,000 of us, and as soon as one person learns something, all of us know it.”
Dr. Geoffrey Hinton, MIT Technology Review
We have peered into the looking glass and seen a mirror image of our mental dysfunctions — a surreal world where the laws of physics that limit humans are reversed. We are fast approaching the singularity — the point at which AI intelligence exceeds our own and becomes uncontrollable and irreversible — resulting in unforeseeable consequences for humankind. It may seem childish for me to belabor the point, but we are not children. AI is not magic. In far too many ways, it is like us. Yet, unlike us, its intelligence is immortal.
“I was thinking as small children think, as if my thoughts or wishes had the power to reverse the narrative,”
― Joan Didion, The Year of Magical Thinking