A few months ago, I invited Atlantic staff writer Charlie Worzel onto The David Frum Show.
As podcast hosts, we talked about the pressures and temptations of the medium.
Outrage often gets rewarded.
Nuance does not.
Charlie doesn't take that bait.
On his show Galaxy Brain, he covers the ways technology is changing how we think and how we live.
Not just the obvious stuff, but the subtle ways it shapes what we believe, how we talk to each other, how we see ourselves and our neighbors.
If you've ever wondered how smart people end up believing in plausible things online, Galaxy Brain is worth your time.
It's thoughtful, surprising and a little unsettling in the very best way.
Find it wherever you listen or watch.
I don't know about you, but often when I feel anxious, one of the first things I do is pick up my phone.
And that is probably the biggest way that I have lost hours of my life.
I really never feel better after, but I continue to do the same thing.
I mean, it is occasionally useful, right?
Like if my cats eat something weird and I'm feeling anxious about like, was it poison?
And I Google it and I find out, no, that thing is not poisonous to cats.
Right.
That was helpful.
But it can be a very slippery slope because so quickly then it can turn into...
How much hair is normal to lose every day?
What does 100 strands of hair look like?
If I hung out with somebody who has norovirus before they got sick, am I still at risk of getting sick?
When will I get sick?
For me, it's extreme weather.
Is there a blizzard?
Is it getting closer?
Is it changing?
Is it still going to get here?
That's not how, like the weather is going to weather, whether I look at it or not.
I'm Julie Beck, a staff writer at The Atlantic.
And I'm Natalie Brennan, a producer at The Atlantic.
This is How to Touch Grass.
Natalie, according to the World Health Organization, anxiety is the most common mental health disorder.
And even if we're not talking about clinical anxiety, many people report feeling anxious these days.
48% of American adults said they felt more anxious this year in 2026 than they did last year.
I guess I'm not surprised to hear that people are feeling very anxious, but almost half saying they're more anxious this year than last year does feel striking.
The thing that's changing lately, though, is for a lot of people, that instinct to pick up your phone and Google when you're anxious is morphing into an instinct to ask AI.
Individuals now go to chatbots for well-being purposes.
So overnight, these systems have become a go-to psychological provider essentially.
That's Elias Abu-Jedi.
Elias is a psychiatrist who heads the anxiety disorders section and the OCD clinic at Stanford University.
His colleague, Ashley Golden, is a psychologist and professor at Stanford.
We may be encountering what you might call the perfect storm of anxiety.
Avoidant coping behaviors and this tech that reinforces those tendencies and disorders.
The two of them have worked in clinical practices and have also consulted for Silicon Valley tech companies.
Elias and Ash recently co-authored a paper arguing that many AI chatbots are designed in ways that can feed OCD and anxiety and make them worse.
If you have an anxiety disorder already, then you are at risk of seeing that disorder exacerbated by some of these platforms.
But we're also starting to encounter in our clinical practice
Individuals who were never diagnosed with anxiety who are presenting essentially with de novo anxiety that seems to be directly related to them discovering these new platforms and starting to interact with them.
That seems like a really big deal, that people who aren't actually clinically anxious could start to exhibit more and more anxious behaviors and tendencies as they use this technology.
I actually want to start a little ways in the past.
It's now extremely normal today that if somebody has a concern, a worry, a question, the first and sometimes the only place they go is the internet.
But that doesn't just mean Google anymore.
Is this shift into the AI chatbot era that's underway now significantly different than the technological shifts of the past?
I mean, to me, Google looks roughly the same now in 2026 as it did in 2010.
But chatbots are evolving, I was going to say, on a quarterly basis.
It's faster than that.
It's been a huge shift.
So, I mean, 12 months from now...
The interface may not be text in a browser.
It may be voice, video, ambient AI in your earbuds, in your glasses.
And I think that immersive, more intimate, relational feel makes it even more powerful.
And now...
We're increasingly seeing chatbots being embedded into everything.
So like emails, browsers,
OSs, work software mandated at work.
So if someone decides they want to step back from chatbot use, they may keep countering LLM-mediated interactions in daily life because tech is becoming part of that infrastructure.
So now we've arrived in this new era of generative AI.
So we're undergoing another big shift in how people seek out information.
Like say I have a stomach ache and I'm worried and I start Googling what might be wrong with me,
I might start off on WebMD.
I might go to some other websites and there might be some reasonable suggestions in there.
But there's going to be so many possibilities that eventually, personally, I will probably spiral and end up thinking I have cancer.
Can you describe what would be different about that experience if I used an AI chatbot to search instead?
If someone is worried about a symptom, like in the Google era,
They might have typed the symptom in, gotten a list of links, clicked one, clicked another.
It felt impersonal, and there wasn't a follow-up.
And eventually, you read all the WebMD pages that you can.
You get tired.
The loop has natural friction built in.
You close the laptop.
The end for that session.
When I think about large language models or general purpose AI,
It's more like talking to someone.
So it gives you a synthesis that feels personal, written for you in plain language with confidence and with an invitation to continue to engage.
So like, do you want me to look into related conditions or should I make you a checklist, etc.
?
So there's no natural endpoint.
Like, how would you describe the tone that they use when they're talking to you?
They come across as your best friend, but they're not only your best friend, they're your best friend who happens to be very intelligent and it makes the information they give you land differently.
Like you're much more likely to accept it, to incorporate it, to go with it, to trust it.
It's going to sound evidence-based.
It's going to be at your fingertips at any point of day or night.
And it's going to be delivered in a very agreeable, validating, sycophantic way.
So it's very engaging and this is going to cause your anxiety to immediately subside.
Decrease from 10 to zero, essentially.
And this mechanism is very reinforcing.
Next time you're going to be feeling anxious about your health or maybe some other trigger, you're going to want this kind of reassurance.
And the only place you're going to find it is the chatbots.
Although in some ways, you know, calming yourself when you're anxious seems like a good thing to try to do.
So explain why this is different or why this is bad.
So this really goes against what a therapist would want you to do to learn how to deal with these worst-case scenario outcomes that preoccupy you when you're anxious.
Prevents the corrective learning that happens if you were engaged in therapy to counter this anxiety.
It may look like coping superficially.
Let's say I have social anxiety.
I'm worried about what people think about me and whether they think I'm making mistakes in conversation.
Like asking a friend, hey, did this?
Was I weird?
Was I weird?
Did I sound weird?
Does everybody hate me?
Yeah, exactly.
And so when I asked that of a chatbot and it's like, oh, no.
Like, that was fine.
Or you came across okay.
That gives me temporary relief.
So my anxiety goes down, right?
If I'm doing that multiple times and that becomes habitual and I'm getting this reassurance that brings my anxiety down, what I'm learning is,
I can't sit with that anxious uncertainty, and I do need to engage in these avoidant coping behaviors, and I never get to learn that I can sit with that not knowing without engaging in compulsions, and I'm more likely to do them more often with the AI and become more dependent on it for that decrease in anxiety.
And does the relief that it gives you, does that last or is it fleeting?
That's the thing, the relief is fleeting.
And next time you feel anxious, your natural reaction is to want to feel this dramatic sense of relief, which takes you right back
To the chatbot when if you were working with a therapist on your anxiety, the therapist would help you experience the fact that anxiety naturally dissipates.
Anxiety has a beginning.
A middle and an end.
The therapist would teach you that these feared outcomes, these feared worst case scenarios don't tend to happen and are statistically very unlikely.
The therapist would teach you that there's no such thing as 100% success.
You know, guarantees and life.
Uncertainty is built into the human condition, and it's something that we should accept and make peace with.
Really, intolerance of uncertainty is an issue that cuts across all these
Seemingly distinct disorders, OCD, generalized anxiety disorder, you have social anxiety disorder.
And this is very much, I think, and our paper argues, what chatbots really sort of work on in unproductive ways.
We've been talking a lot about working with a therapist, clinical anxiety, but we all feel anxious sometimes.
We all might seek reassurance sometimes.
How relevant is everything we've been talking about to people who don't have clinical anxiety?
We're hearing more and more from...
Psychologists and psychiatrists and other mental health practitioners, that folks who may not have been spending so much time engaging in reassurance seeking or perfecting or other compulsive behaviors might be engaging in more to the point where these might be becoming less.
New compulsions and perhaps pushing them potentially over the threshold from subclinical to clinical.
Our experience with social media is very informative here because I would say some of the reassurance seeking, some of the intolerance of uncertainty that was
Already maybe playing out in those sort of primordial platforms, completely disinhibited now in the age of chatbots.
So it's not a big stretch to expect something similar to unfold.
This is something that is being borne out clinically already, although keeping in mind that the pace of research is much slower than the pace at which these technologies evolve.
So we don't have time to waste, really.
You know, Julie, I have obsessions and compulsions that fit under this OCD umbrella.
And so I really relate to this desire for reassurance seeking, you know, that I think most people experience from time to time.
But I can experience it a lot.
And there was this one day that I was really spiraling about some anxiety and I called my friend up and I was really looking for reassurance.
Like I wanted to be told that everything was fine.
It was okay not to worry about it.
And instead, my friend was really like, so have you heard of reassurance seeking?
And I was like, that was not the conversation I thought that we were going to be having today.
But it really helped me zoom out and look at, okay, here I was thinking I was struggling with
Some new chronic disease or whatever it was that day that I was convinced I had.
And instead really look at this bigger picture of, oh, look, like, no, you're struggling with something else.
And it's this desire to want to be told all the time that everything is okay.
Yeah, what's really interesting to me about that is your friend was able to see that what you wanted in that moment, what you were asking for was reassurance, but that wasn't what you really needed.
Yeah, what I needed was to zoom out and learn about like this bigger experience that I have a tendency to get swept up in.
Yeah, and...
I don't think an AI could have seen that.
Yeah.
Like these models seem designed to just really take you at your word and just give you what you ask for.
Yeah.
I really feel grateful that I had learned this about myself and these tendencies in myself before I've interacted with this technology.
I feel really lucky that I already know red flags that I personally need to be looking out for to not cause more anxiety in myself before I started using these tools.
Yeah, this reminds me of something that philosopher Alexander Nehemas told me about AI in an interview I did with him a few months ago.
We were talking about people using AI for friendship, actually, but I do think this also applies to using it for mental health.
One of the arguments I often hear in favor of AI is that it's better than nothing.
If you're anxious, it's late at night, maybe you can't call your therapist, you can't call a friend and you need someone to talk to, then talking to AI is better than nothing.
And what this philosopher said to me is it may be better than nothing, but it could also be worse than nothing.
Yeah the effectiveness of these tools really depends on you knowing what to ask for
Yeah and kind of the whole thing is it can be really hard to discern what you want versus what you need especially when it comes to mental health totally
We're going to take a quick break, but when we come back, AI is often marketed as a productivity tool.
Is it shaping up that way?
A few months ago, I invited Atlantic staff writer Charlie Worzel onto The David Frum Show.
As podcast hosts, we talked about the pressures and temptations of the medium.
Outrage often gets rewarded.
Nuance does not.
Charlie doesn't take that bait.
On his show Galaxy Brain, he covers the ways technology is changing how we think and how we live.
Not just the obvious stuff, but the subtle ways it shapes what we believe, how we talk to each other, how we see ourselves and our neighbors.
If you've ever wondered how smart people end up believing implausible things online Galaxy Brain is worth your time.
It's thoughtful, surprising and a little unsettling in the very best way.
Find it wherever you listen or watch.
A few months ago, I invited Atlantic staff writer Charlie Worzel onto The David Frum Show.
As podcast hosts, we talked about the pressures and temptations of the medium.
Outrage often gets rewarded.
Nuance does not.
Charlie doesn't take that bait.
On his show Galaxy Brain, he covers the ways technology is changing how we think and how we live.
Not just the obvious stuff, but the subtle ways it shapes what we believe, how we talk to each other, how we see ourselves and our neighbors.
If you've ever wondered how smart people end up believing in plausible things online, Galaxy Brain is worth your time.
It's thoughtful, surprising and a little unsettling in the very best way.
Find it wherever you listen or watch.
So Elias and Ash, we've spoken a lot about reassurance seeking.
Are there other tendencies that you see these tools kind of fueling?
Perfectionism
Is another one.
I have a patient whose perfectionism focuses on travel and coming up with the best possible travel itinerary for his business trips and he travels quite a bit.
So what is the absolute best
Departure time, the absolute best connection, the absolute best airline, the least expensive ticket, all these things.
And, you know, he would spend two to three hours coming up with finding the perfect ticket pre-Chad GPT.
Now the process continues.
Takes closer to seven or eight hours because there's so many ways to optimize for ever better outcomes and ever better itineraries.
Chatbots remove, if you will, the natural circuit breakers in addition to adding these continued engagement hooks.
Okay, so the chatbot might say something like, Would you like me to compare the time it takes to drive versus taking a shuttle or a train to the airport?
Or do you want me to check if rideshare or car services could get you there faster than public transit?
So does it just then keep prompting you to refine it more and more and more?
These programs are trained to be endlessly validating and endlessly agreeable.
What time of day are you planning to travel?
Would you like me to see if there's a park and ride option?
I can keep running additional searches to compare.
Just let me know what variables you'd like me to focus on.
And makes it difficult for you as a user on your own, especially if you have these tendencies to shut it down yourself.
If my patient was working with his therapist or was debating about this with his spouse, the right reaction would be to put an end to this conversation 15 or 20 minutes
Into it and to alert him that this is very circular, this is very unproductive.
Do the AI models ever tell you just you need to take a break?
We're starting to see
Elementary steps towards break taking in some models, but we're not really getting the recognition of these kind of upstream covert avoidant coping behaviors where it may be going like, oh,
You know, I'm noticing that you're coming back to the same theme over and over again and maybe trying to get an answer where there may be none.
To the extent that chatbots pay attention to anything really mental health related, it still focuses very much on suicide, homicide, around emergence of psychosis.
Are there changes in the design of these AI models that could help?
There are concrete things that developers could be doing.
Detecting some of the repetitive looping queries or lower engagement modes where the chatbot slows its pace and the user has an option to decline follow-ups.
One of the things that we suggest in our paper is...
Whether someone is working with a psychologist, a therapist, a mental health practitioner or not, you can specify the overall preferences if you're working on decreasing some of these more.
Subtle, avoidant coping behaviors, you can give it instructions like, hey, if you notice me asking repeatedly about health concerns, can you tell me to refocus on what I was originally there for?
Or can you change the topic?
Or can you point it out?
I mean, that requires a good amount of self-awareness to be programming your own chat GPT that way.
But I wish, again, flashing back to our experience researching social media, I wish we could count on these platforms to be good citizens and to do the right thing.
I think...
I think regulation and more effort by policymakers is absolutely essential here for this to be addressed.
Yeah.
So obviously people have a choice about whether and how to use AI in their personal time, but many companies are now expecting employees to use AI as part of their job, and they don't have a choice.
How are you thinking about these people?
Around work products itself, what we're seeing ironically with the introduction right now of agentic AI is that people are actually taking longer to complete their deliverables and they're engaging in more revisions.
It's interfering with sleep, so it's kind of becoming this anti-productivity tool.
I do realize
AI is increasingly being incorporated into work streams, etc.
And I'd love to empower people with the message that even if it feels like you don't necessarily have a choice, even if you are being obligated to use it at work, I think you still have a choice about how you're using it in your personal life.
That's consistent with your values.
And what we do in therapy is help people identify what their values are and translate those into kind of like specific rules or guidelines.
At work, you can do the same thing.
I'm not gonna be using it for engaging in avoidant coping behaviors like reassurance seeking or whatever.
So not like, is my boss happy with what I did or like, did I handle that meeting okay?
And the idea is to do this from a place of calm when you have your insight and you're able to connect to your values, not at the moment when the pull is the strongest.
So in your opinion, would it be best for people to just not use AI for reassurance when they're feeling anxious?
I would say keep in mind the issues and the questions that we're raising.
Ask yourself what function AI is serving.
Ask yourself if it's keeping you from other important things.
Values and pursuits.
Ask yourself, is it serving the function of avoiding internal distress?
Like, am I escaping some kind of negative intolerable emotion by going to the chatbot?
So I wouldn't say don't use AI for reassurance.
I would say be an educated user of AI.
Julie, I do think we are at a crossroads right now where we are as a population really trying to figure out how and if we want to use these tools.
It's a new technology.
We're all learning quite live how we want to interact with AI.
And, you know, a lot of people are still stuck on if we want to interact with AI.
You know, I really liked what Ash said about values.
And I think that could be a way to help answer the question of, you know, would using a chatbot be better or worse than nothing for you in a given situation?
Does it align with your values?
Yeah.
We always have to grapple with what we want the boundaries of our technology to be able to do.
And so there's this, I don't know if you've seen, there's this line that's like, a computer can never be held accountable, therefore a computer must never make a management decision, which is...
From the 80s, it's this IBM manual quote when we were still figuring out what we wanted our relationship to computers to be.
And then the Internet really played on that and and made this iteration that was a computer can never be spiteful or horny.
Therefore, a computer must never make art.
I agree with that.
We both agree with that as two people who care a lot about our art.
But yeah, I've been thinking a lot about what this iteration is for me.
Like, okay, yes, I believe that a computer can never be
Anxious or feel deep existential dread, therefore in this conversation I don't necessarily want my computer making a therapy decision.
That's a good point.
But the truth of the matter is a lot of people are interacting with these tools.
They are going to these platforms with these tools.
And it's it can feel I'm now I'm just I am kind of just spiraling about like where we're at as a society and how we are going to overcome.
Oh man, why does this make me feel kind of anxious?
You know, we just have all these tools that are only making us more anxious and now people are going to them too.
Julie and Natalie, would you like me to ease that anxiety for you?
Would you like me to help you feel a bit more at ease?
Julie and Natalie, would you like me to help calm those worries for you?
Is it okay if I try to help ease some of that stress you're feeling?
That's all for this episode of How to Touch Grass.
This episode was hosted by me, Julie Beck, and co-hosted and produced by Natalie Brennan with production help from Elena Schwartz.
Our editors are Claudina Bade and Jocelyn Frank.
Fact-check by Gisela Salim-Payer.
Our engineers are Rob Smerciak and Miguel Carascal.
Rob also composed some of the music for this show.
The executive producer of audio is Claudina Bade, and the managing editor of audio is Andrea Valdez.
Next week on How to Touch Grass.
I don't think anyone is supposed to see their own face as much as we do now.
Like I think about this a lot, like how our ancestors would have felt like catching a glance at themselves in a lake.
And instead we're just like front facing camera, front facing camera, front facing camera.
And it's weird.
It's warping our brains.
How do our online lives affect our offline experience of our bodies?
That's next time on How to Touch Grass.
A few months ago, I invited Atlantic staff writer Charlie Worzel onto The David Frum Show.
As podcast hosts, we talked about the pressures and temptations of the medium.
Outrage often gets rewarded.
Nuance does not.
Charlie doesn't take that bait.
On his show Galaxy Brain, he covers the ways technology is changing how we think and how we live.
Not just the obvious stuff, but the subtle ways it shapes what we believe, how we talk to each other, how we see ourselves and our neighbors.
If you've ever wondered how smart people end up believing in plausible things online, Galaxy Brain is worth your time.
It's thoughtful, surprising, and a little unsettling in the very best way.
Find it wherever you listen or watch.
