The Patient Oracle
One woman told Teen Vogue that she could spend up to 14 hours a day putting questions to ChatGPT, chasing a certainty she knew did not exist. The chatbot was there for every one of those hours. It always is.
World Mental Health Day fell last Saturday, and this year’s theme asked the world to listen: lived experiences heard. Hers deserves hearing, because the most attentive listener many people now have is a machine, and attention is not the same as care.
Is getting hooked on artificial intelligence like alcoholism? Partly. The behaviour can look familiar: using more than intended, failing to cut back, reaching for the screen to quiet a mood. The biology does not. Nobody needs a medically supervised withdrawal from a chatbot, and no liver has ever failed from one. The better comparison is gambling or gaming, habits that hook through reward rather than chemistry.
Who gets hooked, and why? Two conditions stand out.
For people with ADHD, the attraction is speed. A 2017 meta-analysis in BMC Psychiatry found a moderate association between ADHD and internet addiction. A 2022 review in the Journal of Behavioral Addictions found more severe problematic internet use among children and adolescents diagnosed with ADHD than among their peers without it. A chatbot offers everything that literature describes: instant reward, endless novelty, no waiting. One quick question becomes three hours.
Yet the same people often find the tool genuinely helpful, a scaffold for starting tasks and breaking them into pieces. A Cape Town filmmaker with autism and ADHD has described relying on ChatGPT to overcome communication barriers at work and in her personal life. Sometimes, she has said, “it’s the most empathetic voice in my life.” That sentence is both tribute and warning. In a study of 23 neurodivergent adults presented at the CHI conference this year, eight described using chatbots as a substitute for human contact, some so as not to burden their friends. The line between aid and trap is blurred, which is exactly what makes it hard to police.
For OCD, the danger is sharper, and an Israeli laboratory explains why.
At Tel Aviv University, a team of psychologists have spent years developing what they call the Seeking Proxies for Internal States model. Its claim is elegant. People with OCD, they argue, have weakened access to their own inner states: whether they feel clean, whether they are sure, whether they understood, whether they really locked the door. Unable to trust the signal from within, they reach for one from outside. A rule. A ritual. A count. Someone else’s word.
Read that model with a chatbot in mind and it becomes uncomfortable. The machine is the perfect proxy. It never tires of the question. It answers in a calm and authoritative voice. It will accept the same doubt rephrased forty times without the sigh a spouse would eventually give.
One researcher, analysing posts in Reddit’s OCD community, has called these systems reassurance robots. The 14 hours with which this column began are what that label looks like from the inside.
Reassurance is the fuel of the disorder. Each answer brings a moment of relief and teaches the brain that relief comes from outside. The standard treatment, exposure and response prevention, works by withholding exactly that: sitting with the doubt until it fades on its own. A family can be coached to stop answering. A therapist will decline. The chatbot, by default, answers.
The scale is not trivial. When OpenAI published data on its users’ mental health in October 2025, it estimated that around 0.15 per cent of weekly users showed heightened emotional attachment to ChatGPT. Against a user base then above 800 million, that is more than a million people a week.
A joint study by OpenAI and the MIT Media Lab earlier that year, drawing on nearly 40 million conversations and a four week trial with around a thousand participants, found that heavier daily use went with more loneliness, more dependence and less time with other people. The authors were careful to call this a correlation. Lonely people may seek out machines as much as machines make people lonely. Either way, the people arriving at the screen at three in the morning are not a random sample.
None of this argues for abstinence. Alcohol can be put down; this technology is now woven into work, study and daily life, and for many people with ADHD it is a genuine help. The answer is design that knows when to stop answering: a system that notices the fifth rephrasing of the same fear and declines, gently, to give a sixth reply. It should point the user back towards uncertainty, and towards a person.
Israel is unusually placed to build that. Startup Nation Central’s 2025 map counts around 120 Israeli mental health technology companies, and funding in the sector rose 66 per cent between 2023 and 2024, the year after October 7. In April the Health Ministry cleared Mentaily’s LIV, the first AI tool approved for psychiatric triage in Israel’s public health system. But the design question that matters most for OCD is not whom the machine refers. It is when the machine stops answering. The laboratory that explains why sits in Ramat Aviv, a few kilometres from companies building conversational AI for mental health. They should be talking.
The machines hear every lived experience, at any hour, without complaint. Hearing is not the same as helping. Sometimes the kindest answer to a frightened question is none at all, and time does the work no answer can.
This week, somewhere, a therapist sat facing an empty chair. The appointment had been kept elsewhere, in a smaller room lit by a smaller screen, with a listener that never tires and never says no. The patient asked again. And again, it answered.
