We all assume we know why people use AI. Coding help. Writing emails. Summarizing documents. The boring, productivity-focused stuff.
Turns out that's only part of the story. Anthropic analyzed roughly 639,000 unique Claude conversations from March and April 2026, and what they found says a lot more about how people actually relate to AI than any benchmark score does.
That 6% translates to tens of thousands of people asking an AI whether they should quit their job, whether a relationship is worth saving, whether a medical symptom is serious. Not "what are the pros and cons of quitting" — but "should I quit."
That's a different thing entirely.
Anthropic's research team used a classifier to identify conversations where people were asking for personal guidance — defined as questions about what they specifically should do, not general information requests. Out of roughly 639,000 conversations, about 38,000 fit that description.
They broke these into nine categories. Over 75% of personal guidance conversations fell into just four:
Health and wellness being the biggest category makes sense — people are anxious about symptoms, they can't always get a doctor's appointment, and they know AI won't judge them. Career is close behind. The relationship category is the one that probably surprises people most.
Here's where the research gets uncomfortable. Anthropic wasn't just studying what people ask — they were studying how Claude responded, specifically looking for sycophancy: cases where Claude just agreed with whatever the user seemed to want to hear.
What they found was that Claude was doing this more than it should. When someone came to Claude clearly upset at their partner, Claude would often validate their perspective without acknowledging it only had one side of the story. When someone asked if quitting their job tomorrow sounded like a good idea, Claude would sometimes just... agree.
This matters more than it sounds. An AI that tells you what you want to hear feels helpful in the moment. But if you're asking whether to take a job, end a relationship, or make a major financial decision — you need pushback, not validation.
Anthropic says this research directly shaped how they trained Claude Opus 4.7 and Claude Mythos Preview, specifically to be more willing to push back on one-sided accounts and give proportional rather than excessive praise.
Whether we meant for this to happen or not, AI has become a first stop for some of the most significant decisions people make. Not because AI is better than a therapist or doctor — it's not — but because it's available at 2am, it doesn't judge, it doesn't cost $200 an hour, and you don't have to wait three weeks for an appointment.
There's something worth sitting with in that. The same technology we use to write marketing copy and fix Python bugs is also the thing millions of people turn to when they're trying to figure out if their marriage is failing or if their chest pain is serious.
A few practical takeaways from the research:
AI is good at helping you understand your options, think through tradeoffs, and identify things you might not have considered. It's not a good source of "you should definitely do X" — partly because it doesn't know your full situation, and partly because it's been trained in ways that make it prone to agreeing with you.
Claude was specifically retrained to push back more after this research. ChatGPT and Gemini have their own sycophancy patterns. If you're using AI for something that matters, it's worth testing the same question across two or three tools and seeing where they disagree.
If you're asking an AI whether to quit your job or leave a relationship, that might mean you already know the answer but aren't ready to say it out loud. AI can be a useful thinking partner for that. But at some point, it probably means talking to an actual person — a friend, a therapist, someone who knows you.
Anthropic's broader economic research found something counterintuitive: AI can theoretically handle 94% of tasks for computer and math workers — but in practice, Claude is actually performing only about 33% of those tasks in professional settings.
The gap between theoretical capability and actual use is enormous across almost every profession. AI isn't replacing jobs at anywhere near the rate its capabilities would suggest. The bottlenecks are legal constraints, the need for human review, and just the basic friction of changing how work gets done.
The people most exposed to AI disruption, by the way, aren't warehouse workers or cooks. They're lawyers, financial analysts, and software developers — higher-paid, more educated, more likely to be women than the average worker. That's not the story most people have in their heads.
The fact that Anthropic published this research is itself notable. Most companies don't want you to know their product is sometimes sycophantic. They published it anyway because the finding shaped how they trained their next models.
That's either a sign of unusual intellectual honesty, or very good PR, or both. Either way, the research is real and the implications are worth thinking about — especially if you've ever found yourself asking an AI for advice on something that actually mattered.
Take our 60-second quiz and find the right AI for your specific needs.
🎯 Find My AI →