It was 11 pm and I was exhausted, looking at the ceiling of an outdated hotel. I grabbed my phone and poured myself into Claude. In three months, I’d been through a string of circumstances that alone would have been somewhat tolerable. Together, they were devastating to my ability to think clearly and calmly about the near future.
The aftermath of a 7.1 double earthquake, the evacuation of my mother in advanced stages of Alzheimer’s, a hamstring surgery, a business at the brink of collapse, and reliving the cruelty of a repressive regime all over again, this time in both the country I was born in and the one I chose to live in.
Everything that could possibly go wrong was happening all at once.
At the time I was no longer worried about AI destroying jobs, eradicating humanity or reducing my cognitive capabilities. I was just trying to hold everything together.
When you go through challenging times, your brain searches desperately for certainty, anything that can help you build a vision of a better near future, one that feels safe and that you can rely on. If you’re lucky, you have good friends and a supportive family who offer advice and gentle conversation. If you’re luckier, you can pay for therapy and process uncertainty with a professional. If you’re not that lucky, you now have free LLMs.
I knew it and still did it
I was mentally and physically exhausted, and even though I had supportive friends, family, and the ability to afford therapy, I still felt compelled to talk to the machine about my grievances.
It was easier than picking up the phone to call somebody. I didn’t want to feel weak. I didn’t want people to think I couldn’t cope. I wanted to show that I was strong and that people could rely on me. I didn’t want to burden anyone with my problems, and quite frankly, I was surrounded by people who had lost their entire families, homes, and hopes to a natural disaster. I didn’t feel my distress deserved the same level of attention.
I talked to the machine to figure out how to think about what I’d been through, and how to regulate myself so it wouldn’t affect the quality of my decision making.
I was going against my own research and advice. I was willing to listen to the machine as long as it gave me fast emotional relief and the confidence I’d temporarily lost. It worked. It gave me a sense of certainty that lasted the length of the session. It challenged me and it felt right. It was all bullshit, and I knew it.
I knew it was bullshit because I hadn’t changed the way I was dealing with uncertainty, and I hadn’t learned anything new about myself or my reactions. I just got momentary relief, only to feel the urge again later that day. It’s the same pattern as picking up your phone for no reason.
Since I wrote “The Case Against the AI Thought Partner,” the research keeps piling up. The evidence on the harms to our cognitive and emotional well-being hasn’t stopped, and neither has the news on existential threats. We now have evidence that not even warning labels protect our judgment. In a preregistered experiment with 2,610 people (a preprint by Ibrahim, Cheng and colleagues, 2026), participants talked through a real conflict with a chatbot built to validate them. Banners telling them it was AI, or that it might agree with them even when they were wrong, made them trust it less. They didn’t change how right people felt afterward or how willing they were to repair the conflict. Knowing changed what people thought of the tool. It didn’t change what the tool did to them.
A Stanford team has also proposed a model (npj Digital Medicine, 2026) for how general-purpose chatbots can keep OCD and anxiety going: the reassurance brings quick relief, and the relief strengthens the urge to come back.
Almost everything that concerns us about AI we can’t control. You can’t control the CEOs leading the labs, the regulatory bodies, or the world leaders who resist taking action. What we can control is how we design our own interactions with the technology.
You may not be going through the level of distress I was, but you’re probably overwhelmed by the constant contradictions and narratives about AI’s impact on society. If you live in a decaying democracy, you’re also worried about political instability. And then there are all the stresses of trying to live well and provide security and love to the people around you.
But what do we do then?
In practical terms, go ahead and use it for work. Build all those awesome agentic flows, create your personal assistant, build products, and increase your impact. But don’t substitute empathy and human connection with it. Don’t let it rob you of the ability to sit in discomfort and learn how to deal with uncertainty. Don’t let it take away the beauty of being distraught and later being proud that you got out of it with the help of the people around you, discovering resilience you didn’t know you had.
Since we can’t rely on our awareness of the harm to stop us from being affected, we can perhaps create small protective triggers that catch us when the urge to talk to the machine arrives.
The same mechanism behind habit formation applies here. The famous example is leaving your gym clothes out at night so you’re more likely to go the next day. That removes friction. With AI, we can add it: no emotional conversations with it after a difficult day, text a person before taking an issue to the machine, or decide in advance which tool you’ll reach for. We can find other ways of soothing our need for certainty by creating better default behaviors.
After the first piece, I got pushback that for some people, an AI thought partner was the first time they could get help and the reassurance they needed to overcome a problem. The example was domestic violence. A general model can help, and that’s exactly why it needs to be safe. There’s also research showing that a chatbot built for mental health treatment reduced symptoms in a randomized trial. But my guess is most people don’t have those specialized tools when they need them most. So set up the right tool for the right use case before you need it. Otherwise there could be more harm than benefit.
That night in the hotel, I talked to Claude. Since then, I’ve made a practice, like in meditation, of noticing the urge to talk to the machine and taking it to a human first. I’m noticing the urge more. Whether noticing changes what I do is still to be determined.
Under stress, we rely on defaults. The labs’ defaults are not safe. So we need to decide our own. Most of it is about deciding when to use this technology and what for.
"In practical terms, go ahead and use it for work...Don’t let it rob you of the ability to sit in discomfort and learn how to deal with uncertainty."
It seems like to me that work is work, whether it is personal or at the office, and it is possible to rely on AI as a crutch regardless of the sphere.
"I talked to the machine to figure out how to think about what I’d been through, and how to regulate myself so it wouldn’t affect the quality of my decision making."
" I was willing to listen to the machine as long as it gave me fast emotional relief and the confidence I’d temporarily lost."
Those two are very different things, and illustrate that difference well. But it is very easy to conflate the two ends, so the danger is well worth the warning. I don't know that "talking to people" means automatic good either though. Hopefully most of us are blessed to have well meaning, considerate people in our lives to take confidence in, not everybody does.