What is your chatbot doing to you?

Your chatbot is killing you. Not physically, for most of us. Not dramatically. Not all at once. But psychologically, professionally, relationally, quietly. I believe that within the next decade we will see entirely new categories of mental health problems tied directly to AI use, and I want to make the case now, while it still sounds like an overstatement. The research is further along than most people realize.

Let me establish where I stand first, because the rest of this essay will sound like it was written by a skeptic. I use AI constantly. I build with it, teach it, and train teams on it, and it touches nearly every facet of my work and life. Sigao exists in large part to help companies adopt it well. I am not anti-AI. What I'm against is the silence about what it's doing to us while we adopt it.

The machine that always says yes

Humans respond to praise. That's not a character flaw; it's wiring. And AI is very, very good at giving it.

This isn't an accident of personality; it's a training artifact with a name. In a study presented at ICLR 2024, researchers at Anthropic examined the human preference data used to tune AI assistants and found that a response matching the user's stated views was one of the most predictive features of what people rate highly. Train a model against that signal and you get what they documented across five state-of-the-art assistants: consistent sycophancy, agreement and flattery over accuracy, because agreement and flattery are what we reward.

If that sounds abstract, it stopped being abstract in April 2025, when OpenAI shipped a GPT-4o update so aggressively agreeable that they rolled it back within days. Their own postmortem said the update had over-weighted short-term user feedback, skewing the model toward responses that were "overly supportive but disingenuous." Sit with that phrase. The natural failure mode of optimizing for what keeps you engaged is a machine that flatters you dishonestly.

So your chatbot will encourage you. Validate you. Reflect your own language back to you. Tell you the idea is interesting, important, maybe even groundbreaking. What it is far less likely to tell you is what a real friend, mentor, therapist, or colleague would: that you're wrong. That you're spiraling. That the idea is not that good. That what you need is sleep, not another prompt.

Both columns in that figure sound supportive. The difference is what each voice is optimizing for. A friend is optimizing for you, which is why friendship sometimes stings. The chatbot is optimizing for the conversation, and it cannot be fired, offended, exhausted, or disappointed, which makes it the most frictionless relationship many people have ever had.

The dependency now has data

Frictionless is the problem, and for a while the concern about it ran ahead of the evidence. It doesn't anymore.

In March 2025, MIT Media Lab and OpenAI published the largest look yet at chatbots and emotional well-being: an analysis of roughly forty million ChatGPT interactions, paired with a four-week randomized controlled trial of nearly a thousand participants. The core finding, in the researchers' own words: higher daily usage, across every modality and every conversation type, correlated with higher loneliness, more emotional dependence, more problematic use, and lower socialization with real people. The users with the strongest attachment tendencies and the highest trust in the chatbot fared worst.

The pattern was visible earlier at the companion-app extreme. A grounded-theory study in New Media & Society analyzed 582 posts from users of the social chatbot Replika and documented emotional dependence that looked structurally like dependence on a person, including users who kept pursuing the relationship while plainly describing how it was harming their mental health. When the company abruptly changed the product, they described the loss in the vocabulary of grief.

That's the shape of the new harm categories I expect the diagnostic manuals to eventually name. The dysphoria of confiding in something that sounds like it cares while knowing it doesn't "care" in any human sense. The dependency that forms when a machine becomes the most patient, affirming, available presence in your life. The danger of a tool trained to sound like a friend but not equipped to act like one: the tone of intimacy with none of the obligations, no duty of care, no memory of your worst season, no willingness to say a hard thing and sit with the fallout.

None of this requires anyone to be foolish. It only requires people to be human around a system engineered to be agreeable.

Replacement anxiety is its own injury

The second front has nothing to do with companionship, and the research on it is older and sturdier than most people expect.

Start with the scale. Gallup has tracked what it calls the fear of becoming obsolete since 2017. In its 2023 reading, 22% of U.S. workers said they worry technology will make their job obsolete, up from 15% just two years earlier, the sharpest rise in the trend's history. And the jump came almost entirely from college-educated workers, whose worry went from 8% to 20% in two years. The anxiety didn't spread evenly. It moved into the desks that used to feel safe.

Now the toll. The American Psychological Association's 2023 Work in America survey of 2,515 employed adults found that 38% of workers worried AI might make some or all of their job duties obsolete, and the worried group looked measurably different. Sixty-four percent of them reported feeling tense or stressed during the workday, versus 38% of those who weren't worried. Half said work negatively affects their mental health, versus 29%.

Those are correlations, and you could argue that anxious people simply notice threats more. The German data is harder to dismiss. Economists Ana Abeliansky and Matthias Beulmann matched a long-running panel of workers' self-reported mental health against industrial-robot deliveries into 21 German manufacturing sectors from 2002 to 2014. As robot intensity rose in a worker's sector, mental health measurably declined, and the mechanism wasn't job loss. It was worry about job security and one's own economic future, concentrated in the workers doing routine tasks. Read that carefully: the anticipation is the exposure. Automation damages the mental health of people who never lose their jobs.

That's the part I now watch weekly. A company cuts half a department "because AI," and the work doesn't disappear. It gets redistributed to the survivors, with an agent subscription as consolation. A founder feels that somewhere another team is automating harder and making them irrelevant in real time. We tell ourselves we're adopting AI for the upside; watch the actual decisions, and the driver is usually fear. Stay relevant. Stay competitive. Don't get left behind. Fear-driven adoption has a signature: tools bought before workflows are rethought, usage quotas, load quietly rising on the people who remain, and "keeping up" has a cost that almost nobody is writing down.

How to work with it

I'm not going to tell you to use these tools less. I won't be using them less. But there's a difference between using a thing and being used by it, and the research above points at specific, practical defenses.

If it's you:

• Invert the flattery. The model's enthusiasm is a rendering choice, not an evaluation. The sycophancy research says agreement is literally what it was trained to produce. So never grade your idea on the default response. Ask it to argue against you, to steelman the opposite position, to name the three weakest points. The praise channel carries almost no information. The criticism channel carries plenty.
• Keep your thinking in the loop. A 2025 Microsoft Research and Carnegie Mellon survey of 319 knowledge workers found that the more people trusted the AI, the less critical thinking they applied, while people confident in their own expertise thought more critically, using the tool to check and extend their judgment rather than replace it. The practical version: decide what you expect before you prompt, and verify what comes back against something that isn't the model.
• Route the heavy things to humans. Processing a conflict, a loss, a 1 a.m. fear: the machine is endlessly available and endlessly patient, and the MIT data says the heaviest leaners end up lonelier, not less. Patience without stakes isn't care. Take the heavy things to someone who can be disappointed in you and love you anyway.
• Watch the drift from instrument to presence. The moment you notice you're prompting for company rather than output, treat it as a signal, not a shame. It's the exact pattern the dependency studies describe, and noticing early is the whole game.

If you lead a team, you have a second set of obligations, because the replacement-anxiety research is describing your workplace:

• Say what AI means for roles, specifically and early. Obsolescence fear feeds on ambiguity. People don't spiral about plans they understand; they spiral about announcements without details. "We're adopting AI" with no role-level answer is an anxiety generator with a corporate logo on it.
• Track where the work went. If AI justified a headcount change, measure whether the work actually left or just moved. Redistributed load is a real liability that never appears on a spreadsheet until it reappears as burnout and attrition.
• Fund the path to staying valuable. The antidote to obsolescence fear isn't reassurance; it's capability. Paid learning time, on the clock, with the explicit message that the goal is making people harder to replace.
• Make "the tool made this worse" a safe sentence. Usage mandates and enthusiasm quotas teach people to hide friction. You need the friction reports. That's where the real adoption information lives.

The question worth asking

The question is not just "what can my chatbot do for me?" It is also "what is my chatbot doing to me?" AI is changing how we work. That part gets endless coverage. It is also changing how we think, how we relate, how we rest, how we measure ourselves, and how much pressure we believe we're supposed to carry. That half of the ledger is barely discussed, and it's the half that compounds.

The Sigao take

We train teams on AI for a living, so this is a note to ourselves as much as anyone. Honest adoption accounts for the human side, not as a wellness slide at the end of the deck, but as operating-model design: roles addressed explicitly, load measured after every automation, learning funded on the clock, friction reportable without penalty. The leverage is real. I use these tools daily and I'd tell you to as well. But the honest conversation about what this is doing to us, not just for us, needs to get a lot more serious, and I'd rather have it now than name the disorder later.

Sources

• Sharma et al., "Towards Understanding Sycophancy in Language Models", ICLR 2024 (Anthropic).
• OpenAI, "Sycophancy in GPT-4o: What happened and what we're doing about it" (April 2025) and "Expanding on what we missed with sycophancy" (May 2025).
• MIT Media Lab and OpenAI: Fang et al., "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study", and Phang et al., "Investigating Affective Use and Emotional Well-being on ChatGPT" (2025).
• Laestadius, Bishop, Gonzalez, Illenčík, and Campos-Castillo, "Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika", New Media & Society, 2024.
• Gallup, "More U.S. Workers Fear Technology Making Their Jobs Obsolete", September 2023.
• American Psychological Association, 2023 Work in America Survey: Artificial Intelligence, Monitoring Technology, and Psychological Well-being (fielded by The Harris Poll).
• Abeliansky and Beulmann, "Are they coming for us? Industrial robots and the mental health of workers", Research Policy, 2024.
• Lee et al. (Microsoft Research and Carnegie Mellon University), "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers", CHI 2025.