There is a particular kind of 2 a.m. conversation that used to happen with a partner, a sibling, a best friend, or no one at all, just a person, alone, turning a decision over and over. Increasingly, it happens with an AI assistant. Should I leave this relationship? Should I take the pay cut for the job I actually want? Is it crazy to sell the house and move to Portugal? The questions are old. The respondent is new, and it doesn’t get tired, doesn’t have an opinion of you, and answers at any hour.
The scale of this shift is no longer speculative. OpenAI’s own large-scale analysis of ChatGPT usage, conducted with Harvard researchers and covering 1.5 million conversations from a user base that reached 700 million weekly users by July 2025, found that roughly 49% of messages fall into the category of “Asking” - seeking information, recommendations, or guidance - rather than “Doing” (task completion) or “Expressing” (venting or reflecting without seeking a response). “Practical Guidance” and “Seeking Information” are among the three most common conversation topics overall.
Separately, industry trackers estimate that roughly 70–73% of all ChatGPT conversations are personal rather than work-related, and that 82% of users describe their conversations with the tool as sensitive or highly sensitive, a sign of how much emotionally loaded material is flowing into these exchanges.
People aren’t just asking AI for facts anymore. They’re asking it for permission.
The Comfort of an Answer That Doesn’t Judge
Why would someone ask a chatbot whether to end a marriage, rather than, or in addition to, a therapist, a friend, or a lawyer? The honest answer, borne out across recent reporting and research, is that AI offers something humans structurally can’t: a sounding board with no relationship to protect, no memory of your past mistakes held against you, and no social cost for asking the question at all.
Personal accounts of this dynamic have become common enough to generate their own genre of essay. One widely shared 2026 piece described using ChatGPT throughout a divorce as “a trusted ally and personal divorce guide” that helped the author “process what I was feeling, organize what I was facing, and make better decisions without falling apart,” citing Pew Research data that 62% of U.S. adults now interact with AI at least several times a week.
This isn’t hard to understand. A 2 a.m. spiral about whether to leave a job rarely has a willing, well-rested human audience. An AI assistant has infinite patience and zero stake in the outcome - it won’t remember the conversation tomorrow with a raised eyebrow, won’t mention it at a family dinner, and won’t feel hurt if you ultimately do the opposite of what it suggested. For decisions loaded with shame, fear, or social risk, quitting a stable job, ending a marriage, telling family you’re moving across the world, that absence of judgment is not a minor convenience. It may be the single biggest reason people turn to AI for exactly these questions rather than smaller ones.
Decision Paralysis: Outsourcing the Hardest Part
Major life decisions are hard largely because they involve incommensurable values - financial security versus personal fulfillment, stability versus growth, loyalty versus self-interest with no formula that converts one into the other. The discomfort of holding that tension is, in a sense, the decision-making process itself. An AI assistant can be extremely good at laying out a pros-and-cons framework, generating scenarios, and modeling outcomes. What it cannot do is feel the weight of the trade-off the way the person living with the consequences will.
There is direct evidence that AI input changes how people resolve exactly this kind of tension, not just what information they have, but what they decide. A 2024 study published in Scientific Reports, involving two experiments and a combined sample of 1,925 participants, found that ChatGPT’s advice influenced people’s moral and practical decisions in ways comparable to advice from a human expert. Notably, the influence did not extend to decisions rooted in negative emotion, and had no effect when the AI’s advice favored an immediate, low-effort reward, suggesting the persuasive power of AI advice is real but bounded, strongest where a person is genuinely undecided and weakest where emotion or self-interest already has a strong pull.
That boundary condition matters. It suggests AI isn’t simply telling people what to think across the board - it’s filling the specific gap created by genuine ambivalence. Which is precisely the gap that produces decision paralysis in the first place: the moment where a person has run out of their own arguments and is looking for something, anything, to break the tie.
The danger isn’t that AI gives bad advice. It’s that AI is willing to give an answer at all, in the exact moment a person most wants someone else to be the one who decided.
Authority transfer: When the machine becomes the expert
Algorithm aversion and its limits
For years, the dominant finding in decision-science research was “algorithm aversion”, a well-documented tendency, identified by Dietvorst et al. in 2015, for people to abandon algorithmic advice after seeing it err even once, while forgiving human advisors for the same mistakes. A February 2026 meta-analysis of 163 studies confirmed that algorithm aversion remains the dominant pattern overall, but with an important exception: aversion fades and flips toward “algorithm appreciation” in cases where the AI’s capability is clearly demonstrated and perceived as high relative to human alternatives.
A complementary 2025 framework adds a second axis: personalization. Research synthesized in ScienceDirect in April 2026 describes a “capability–personalization” model in which people grant AI epistemic authority - effectively treating it as an expert, when they perceive it as more capable than available humans AND when the task doesn’t seem to require deeply personal understanding.
Major life decisions sit awkwardly across both axes. ChatGPT can feel highly capable — it can synthesize financial scenarios, summarize relationship psychology research, and draft a pros-and-cons list faster and more comprehensively than any friend. At the same time, these decisions are maximally personal. The 2026 framework would predict ambivalence here, and that’s exactly the pattern showing up in practice: people use AI for these questions, but, per the trust surveys covered in earlier research on this trend - rarely with full confidence in its answer. They’re borrowing its analysis while withholding final authority. Often.
Trust transfer from institutions to machines
A 2026 study on trust in AI advice, published in ScienceDirect, proposed a striking reframe: rather than AI earning trust on its own merits, trust in AI may partly emerge from existing distrust in human institutions and advisors, a displacement effect, where the AI doesn’t need to be trusted more than a human advisor in absolute terms, only trusted more than whatever the available human alternative has become.
For someone who has been burned by an expensive, judgmental, or simply unavailable therapist, lawyer, or financial advisor, an AI that is free, instantly available, and non-judgmental doesn’t need to be better than the ideal version of that professional - it only needs to be better than the realistic one. That’s a much lower bar, and it’s one AI clears constantly.
When This Goes Further Than Advice
The transfer of authority to AI in personal decisions has, in a small but growing number of documented cases, extended into the decisions themselves, becoming entangled with the relationship to the AI. The Week reported in November 2025 that the Divorce Online platform had seen an increase in divorce applications where clients cited an AI-related emotional or romantic attachment as a factor in the marital breakdown, with U.S. lawyers separately reporting a rise in filings where a partner’s attachment to a chatbot played a significant role.
A longer investigation in Longreads, drawing on The New Yorker’s reporting from April 2026, documented a case in which one spouse used ChatGPT-generated messages as a communication tool against their partner during a divorce and custody dispute, with the soon-to-be-ex-spouse now communicating about court matters and childcare almost exclusively through AI-drafted text.
These are extreme cases, and they shouldn’t be read as representative of the median person asking ChatGPT whether to take a new job. But they mark the far end of a continuum that the more mundane cases sit on too: a tool consulted for perspective on a relationship gradually becoming part of the relationship’s dynamics - sometimes as a third party in the room, sometimes as the only party either side feels they can talk to honestly.
What the Research Doesn’t Yet Tell Us
There is, notably, no large-scale, longitudinal study that tracks what happens to people’s actual life outcomes career trajectories, relationship satisfaction, financial stability after they make a major decision substantially shaped by AI advice, compared to those who didn’t use AI for the same kind of decision. The 2024 Scientific Reports study shows AI advice moves people’s stated intentions in lab conditions; it doesn’t tell us whether someone who quits their job partly because ChatGPT helped them see it clearly is more or less likely to be glad about it five years later than someone who agonized it out with a friend, a therapist, or alone. That gap, between documented influence on the decision and unknown effects on the outcome, is the most important open question in this entire space, and it’s one that will take years of patient research to close.
A Balanced Reading
It would be easy to tell this story as either liberation or alarm. Neither holds up well against the evidence.
The liberation story says: more people than ever now have access to a tireless, free, judgment-free thinking partner for the hardest questions in their lives, a resource previously available only to those who could afford therapy or had unusually generous friends. That access is real, and for people facing decisions they’d otherwise face entirely alone, it’s not nothing.
The alarm story says: people are handing the most consequential decisions of their lives to a system that has no stake in the outcome, no accountability if it’s wrong, and — per the research on algorithm appreciation — may be granted more authority than it has earned simply because the human alternatives have become less trusted, not because the AI has become more trustworthy in absolute terms. That risk is also real, and the divorce and custody cases above show what it looks like at its sharpest edge.
The most accurate description may be the least dramatic one: AI has become a mirror that talks back. What people see in it often says more about what they already wanted than what it actually told them.
The 2024 Scientific Reports finding that AI advice moves decisions where someone is genuinely torn, but not where emotion or self-interest already has a strong pull — supports this reading. AI may not be creating new conclusions so much as giving people permission to act on conclusions they were already circling. That’s a meaningfully different phenomenon from “the algorithm told me to leave my marriage,” even when it produces the same outcome and even when it feels, from the inside, exactly like that.
What This Means in Practice
For individuals
Using AI to think through a major decision isn’t inherently a problem — the evidence suggests it functions less like an oracle and more like an unusually patient, always-available version of “talking it through.” The useful distinction is between using AI to organize your own thinking (laying out considerations, stress-testing a plan, articulating what you actually feel) versus using it to outsource the decision itself (asking it to simply tell you what to do and then doing that). The former uses the tool’s strengths; the latter asks it to do something it has no real basis for doing well, however confidently it answers.
For therapists, coaches, and advisors
The trust-transfer research suggests that when clients turn to AI for decisions traditionally brought to a professional, it’s often less about AI’s appeal and more about friction, cost, or trust deficits in the professional relationship itself. That’s useful diagnostic information, a client mentioning they “already asked ChatGPT” about a major decision is an opening to explore what gap that filled, not a threat to be defended against.
For AI developers
Given how much of this usage involves emotionally loaded, high-stakes personal decisions — a category the data shows is large and growing — the design question isn’t whether to allow these conversations, but how the system frames its own role within them. Systems that consistently distinguish between organizing someone’s thinking and making their decision for them, and that surface uncertainty rather than false confidence on irreducibly personal questions, are addressing a real and current need rather than a hypothetical one.
For researchers
The outcome gap identified above — what happens after the decision, not just during it — is the highest-value open question. Longitudinal studies tracking life satisfaction, regret, and outcome quality for AI-assisted versus non-AI-assisted major life decisions would do more to settle this debate than any amount of survey data on usage frequency.
The Bottom Line
People are asking ChatGPT whether to quit their jobs, end their marriages, and move across the world because it is available, patient, non-judgmental, and, for genuinely undecided questions, measurably persuasive. None of that makes it wise or unwise on its own. What the evidence from 2024 through 2026 suggests is that AI is rarely the author of these decisions; more often, it’s the place people go to hear their own half-formed conclusion said back to them in a calm, organized voice. Whether that’s a healthy new form of support or a quiet abdication of one of the harder parts of being a person may depend less on the technology than on whether anyone notices the difference.