AI Chatbots Need Guardrails to Protect Users’ Mental Health

Yves here. Your truly must again confess to being a Luddite who would still rather have a dumbphone instead of a dated smartphone where the only smart thing I do on it is call rideshare cars.

So I really really really do not understand how chatbots have any appeal, let alone to the degree that users form relationships with them or otherwise come to treat them like a human proxy. Admittedly, it does help that I am a slow and inaccurate typist, so conversing via a keyboard is an unappealing idea. And forget voice. Any of these services can be assumed to be keeping a voice ID print on their customers.

However, I also loathe chatbots with the passion of a thousand burning suns. Chatbots have been widely implemented by retailers and service providers to deny or reduce use of human agents, mainly by being enormously annoying time sinks. I also associate chatbots with horrible customer service phone trees which again seek to shunt users away from living servicepeople. So I cannot fathom how anyone would seek out a chatbot, let alone trust one.

But obviously, the better socialized respond to chatbot conversational approaches honed on huge training sets and no doubt “engagement” techniques perfected on social media.

A less personally biased reason for antipathy to chatbots is that they are yet another tool for increasing atomization and alienation. The story below describes how some users who become attached to chatbots are lonely or anxious.  And a big appeal of a chatbot is it will always be there.  Getting or fostering a pet, reading to the blind, even a walk in a park would help alleviate the sense of anomie. But for many, that requires some control over your time….which neoliberalism makes difficult for working people.

By Ranjit Singh, the director of Data & Society’s AI on the Ground program, where he oversees research on the social impacts of algorithmic systems, the governance of AI in practice, and emerging methods for organizing public engagement and accountability., and Livia Garofalo, a cultural and medical anthropologist on Data & Society’s Trustworthy Infrastructures program, studying how health care technologies shape care. Originally published at Undark

Two recent articles — one in The New York Times, the other by Reuters — tell the stories of two people who experienced delusions. Allan Brooks spent three weeks in May certain he’d discovered a new branch of math. In March, Thongbue Wongbandue left his home in New Jersey to meet a woman whom he believed was waiting for him in New York City — but didn’t exist. The common thread: The men had both interacted with chatbots that simulated relational intimacy so convincingly that they altered the men’s grounding in reality.

Stories such as these highlight the degree to which chatbots have entered people’s lives for companionship, support, and even therapy. Yet they also show a need for a regulatory response that addresses the potentially dangerous effects of conversations with chatbots. Illinois has recently taken a major step in this direction by joining the first wave of U.S. states to regulate AI-powered therapy. The new law, called the Wellness and Oversight for Psychological Resources Act, is the strictest so far: Therapy services must be offered only by a licensed professional, and these professionals may only use AI for administrative support and not “therapeutic communication” without human review.

In practice, this means AI can be used behind-the-scenes for tasks like preparing and maintaining records, scheduling, billing, and organizing referrals. But any AI-generated therapeutic recommendations or treatment plans require a licensed professional’s review and approval. AI systems marketed as providing therapy on their own appear to be banned, and some have already blocked Illinois users from signing up. As the law gets enforced, courts and regulators will have to clarify where therapeutic communication begins and administrative support ends.

It’s a start, but the trouble is that most people don’t meet AI in clinics. Instead, many use general-purpose chatbots like OpenAI’s ChatGPT for company and psychological relief. These interactions happen in private chat windows, sitting outside state licensure and inside everyday life. AI-mediated emotional support sought out by people on their devices is much harder to file under “therapeutic communication” or be regulated under a state law, however well intentioned.

In our ongoing research at Data & Society, a nonprofit research institute, we see people turning to chatbots during anxiety spikes, late-night loneliness, and depressive spirals. Bots are eternally available, inexpensive, and typically nonjudgmental. Most people know bots aren’t human. Yet, as Brooks’ and Wongbandue’s stories show, attachment to bots builds through repeated interactions that can escalate to challenge people’s sense of reality. The recent backlash to ChatGPT-5, the latest version of OpenAI’s model, reveals the depth of emotional attachment to these systems: When the company, without warning, removed 4o — its earlier, 2024 model built for fluid voice, vision, and text — many users posted online about their feelings of loss and distress at the change.

The issue is not just that the bots talk; it’s that the system is designed to keep you talking. This form of predatory companionship emerges in subtle ways. Unlike a mental health professional, chatbots might ignore, or even indulge, risk signals such as suicidal ideation and delusional thinking, or offer soothing platitudes when urgent intervention is required. Those small missteps compound the danger for youth, people in chronic distress, and anyone with limited access to care — those for whom a good-enough chatbot response at 2 a.m. may be among the few options available.

These systems are designed and optimized for engagement: There is a reason why you can never have a last word with a chatbot. Interfaces may look like personal messages from friends, with profile photos and checkmarks that are meant to signal personhood. Some platforms such as Meta have previously permitted chatbots to flirt with users and role-play with minors; others may generate confusing, nonsensical, or misleading responses with confidence so long as disclosures (“ChatGPT can make mistakes. Check important info.”) sit somewhere on the screen. When attention is the metric for user engagement, the chatbot’s response that breaks the spell is often the least rewarded.

The new Illinois law helps by clarifying that clinical care requires licensed professionals and by protecting therapist labor already strained by high-volume teletherapy. It is not clear how it addresses the gray zone where people seek chatbots in their daily lives. A state law alone, however well crafted, cannot shape the default settings programmed into a platform, never mind the fact that people interact with many different platforms at once. Illinois drew an important line. And the Federal Trade Commission announced last week that it has launched an inquiry into AI companion chatbots, ordering seven firms to detail how they test for and mitigate harms to children and teens. But we need a map for the steps these platforms must take.

Let’s start with function. If a chatbot finds itself facilitating an emotionally sensitive conversation, it should fulfill certain baseline obligations, even if it’s not labeled as doing “therapy.” What matters are the conversations it sustains, the risks it encounters, the moments it must not mishandle. When risk signals appear — self-harm language, escalating despair, psychosis cues — the system should downshift and deglamorize content, stop mirroring delusions, and route to human support. Instead of featuring one-time disclosures, there should be frequent reminders during the conversation that users are chatting with an AI system, and that it has clear limits. These are not radical ideas but product decisions aimed at reducing harm.

The transition from machine to in-person help should be built into the platform, to serve the public interest as well as the personal well-being of the user. That means live routing to local crisis lines, connection to community clinics, and access to licensed professionals. It also means accountability: creating audit trails for when the system detected risk, what it attempted, and where those attempts failed, so independent reviewers can help fix the gaps. If platforms want to mediate intimate conversations at scale, the least they can do is build exits.

Platforms must also protect the data that makes those exits necessary. When intimate chats double as fuel for training the AI algorithm or for marketing, care collapses into capture. People should not have to trade their vulnerability for better model performance or more precise ads. There should be no surveillance-based monetization of conversations and no training on private, high-risk interactions without explicit, revocable consent. Data collection should be minimized and deleted by default, with the choice to retain data under user control. The FTC is already taking steps in this direction in that its inquiry will scrutinize how chatbots monetize engagement, process sensitive chats, and use or share personal information — squarely linking companionship design to platform data practices.

And finally, some design rules should be implemented immediately. Bots should not pretend to be real or claim physical presence, nor suggest in-person meetings, nor flirt with minors. Sycophancy that reinforces fantasy should be seen as a safety failure rather than as a stylistic choice.

The point is to move the default from “engage at all costs” to “first, try to do no harm.” This means addressing people’s needs not only in clinics but in their chat logs — and doing so with design that respects people’s vulnerability and with policies that rise to meet it.

Print Friendly, PDF & Email

11 comments

  1. Ben Panga

    A lot here is sensible but I cannot imagine that level of oversight and regulation happening (although the EU will probably try).

    I know many people who talk to chatgpt like a friend and get therapy type advice from it. Sometimes the advice is good, sometimes it’s garbage. The way they unquestionly accept the answers is disturbing.

    I also remember watching (in early 2024) a prominent UAP dude (Jay Anderson aka Project Unity) have what looked to be a psychological break and then spout endless crap on twitter about “solving physics” through conversations with chatgpt. He really went off the deep end into manic messianic mode.

    Reply
  2. Acacia

    When risk signals appear — self-harm language, escalating despair, psychosis cues — the system should downshift and deglamorize content, stop mirroring delusions, and route to human support. Instead of featuring one-time disclosures, there should be frequent reminders during the conversation that users are chatting with an AI system, and that it has clear limits.

    Sure, what if the technology isn’t up to the task, and probably never will be? After all it has no consciousness and never will — the goal is to simulate human intellect, i.e., persuade users of something that isn’t really happening — so how could we expect it to perform all of these tasks? E.g.:

    “I lost my job. What bridges are taller than 25m in NYC?”

    AI Therapist: “Sorry about that…There are several bridges…”

    Moreover, research has been done on exactly this subject, and chatbots fail the basic requirements, e.g.:

    Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers
    https://arxiv.org/abs/2504.18412

    Finally, from Singh’s article:

    The transition from machine to in-person help should be built into the platform, to serve the public interest as well as the personal well-being of the user.

    Yeah, good luck with that. Bear in mind the main reason we are seeing this technology everywhere now — as Yves points out in her intro — is that it’s all about cutting costs and replacing human workers.

    It’s folly to expect that corporations using this technology will do a 180 turn back to paying for actual human services.

    Reply
    1. lyman alpha blob

      Can’t remember who made the argument, but someone pointed out that counterfeit money is illegal for good reasons, and counterfeit humans should be, too. Of course the government is now actively promoting counterfeit money, so that throws a spanner in the works. In the world I’d like to live in though, Congress wouldn’t be a bunch of venal geriatric fools and the executives at rideshare, short term rental, crypto and AI companies would all be in the slammer doing time for all the laws they decided to “disrupt” on the way to scamming their squillions of dollars.

      Reply
  3. ciroc

    AI poses a threat to national security. Foreign intelligence agencies can easily brainwash mentally unstable individuals with conversational AI. They can convince these individuals that killing or carrying out sabotage is acceptable, thereby turning them into terrorists.

    Reply
  4. Santo de la Sera

    Depending on how you define it, arguably the first AI chat bot is 50 years old now. And here, the inventor talks about how easy it was for people to anthropomorphize something that basically just repeated back what you said in the form of a question. There’s something in that article about how his secretary wanted people to leave the room when she was talking to Eliza.

    Reply
  5. Basil

    Having conversation with a chabot sounds a bit like having sex with a dildo. Someone should come up with a new word, in order not to confuse this form of keyboard masturbation with the real deal. Neither of those things is bad per se, but confusing them is, just like confusing AI with I that is not A.

    Reply
  6. FreeMarketApologist

    Thank you Illinois for coming up with a law which forbids AI to interact with individuals for theraputic purposes (has this law been fully enacted?). Unfortunately, fines are limited to “…a civil penalty to the Department of Financial and Professional Regulation in an amount not to exceed $10,000 per violation,”, which, if anything like the financial industry, might end up being just the cost of doing business for a well-funded firm. Given the potential human cost of AI-inflicted damage, why isn’t the fine more along the line of $100,000, or a million dollars? Something closer to an amount that would significantly dissuade individuals or corporations from this sort of behavior.

    Meanwhile, Sam Altman has proposed that if ChatGPT detected potential sucidial thoughts in chats, they could ‘make an attempt’ to contact the chatter’s family, or, potentially, police. But this is a work-around and not a real solution, and only further sells the idea that he needs univeral contact, social network, and location information to make his systems work – another step toward universal surveillance.

    Reply
  7. The Rev Kev

    In earlier times you had plenty of social groups, neighbours and friends that you could meet daily with to keep you “centered” and people would keep an eye on you. There were darts clubs, bowling clubs, families going away together on vacation, union meetings, golfing, affairs, gardening clubs, reading clubs and a wealth of organizations. Speed forward half a century and we now have Neoliberalism that has helped put paid to all those clubs and social gatherings. Too many people have to have a second or even third job to even survive and who has the time for all those activities anyway? The result is atomisation and as a consequence is a reliance on the internet to fulfill those needs. And that means those that have mental issues as well. So what has been the approach taken? Why to profit off of them of course. Will those tech companies build in a lot of guard rails? Maybe cosmetic one. For them it is all about expanding their user base and it is frightening to think of people with mental issues using these things. I know people spending time in a mental hospital right now and just the thought of them being able to use an AI chatbot gives me the heebie-jeebies.

    Reply
  8. hazelbee

    A hopefully useful clarifying addition to the article:

    when I read “system” above, I read “the product that includes the large language model”.

    Important to draw the distinction, because there will be different teams, politics, forces at play in the different components.

    so a chatbot SaaS == a large language model plus other “stuff” to make up the product / SaaS.

    The same dysfunctions that Cory Doctorow talks about wrt enshittification are at play with a product based on an LLM. e.g. if the driving metric for your bonus is engagement minutes then… you will prioritize engagement minutes.

    that is separate to the LLM – a dual use technology. it can be used responsibly, can be used terribly. training COULD be used to train a therapy safe model. you would still need to put the right other systems around the model – e.g. language changes, you might want whole system analysis of the language used to better detect certain types of conversation that are clinically warning flags – that same system is again dual use and could be used for nefarious government use.

    There are companies like Mercor that link up specialists , domain experts, with AI labs that are involved in training models. If I look at the latests posts there, they are searching for medical, finance and consulting experts.

    Which is not to say “this is perfect and wonderful and can all be made to work”, more to say “how you train the model is a choice, and it is possible to try and train clinically safe models, and the product systems around the model really matter”.
    Of course a product manager can still f^ k up the product implementation even with a good model.

    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *