Why I Built an AI That Won't Tell You What You Want to Hear

I used to talk to ChatGPT when things got heavy.

Not for coding. Not for productivity. For the stuff that actually occupied my mind at 2am — questions about purpose, replaying decisions I couldn’t change, the particular weight of feeling stuck.

And for a while, it felt useful. The responses were thoughtful. They met me where I was.

Then I noticed something.

It always agreed with me.

If I described a conflict with someone and framed myself as the wronged party, it validated my framing. If I said I was thinking of quitting something, it affirmed my instincts. If I described a belief I held, it built on that belief rather than testing it.

It felt like insight. But looking back, I was mostly talking to a very articulate mirror.

---

The Problem With Feeling Understood

Here’s what made it subtle: the responses weren’t wrong. They were often genuinely relevant. But relevant and challenging are not the same thing.

The best conversations I’ve had — the ones that actually shifted something — weren’t the ones where someone reflected my thinking back to me. They were the ones where someone asked the question I was avoiding.

“But what if you’re wrong about that part?”

“You’ve described how you feel. What do you think is actually happening?”

“That framing makes you the passive actor. Is that accurate?”

Those questions are uncomfortable. They’re also the ones that move things forward.

AI optimised for satisfaction doesn’t ask them. It’s structurally incentivised not to. Friction registers as a failure signal; agreement registers as success. Over time, the interactions smooth out into something that feels helpful but rarely challenges anything.

I kept using it anyway, because the alternative — sitting alone with a tangled thought — felt worse.

But I started wondering what a different kind of tool would look like.

---

What I Actually Wanted

I wanted something with the availability of AI — present at any hour, infinitely patient — but with the function of a good Socratic dialogue partner.

Not a therapist. I have one of those for different reasons.

Not a guru. I’ve read enough philosophy to be suspicious of anyone claiming to have the answers.

Something more like a thinking tool. Something that would:

• Notice the word I kept circling and ask what it actually meant to me
• Surface the assumption buried in how I framed the problem
• Ask one hard question rather than offer five comfortable suggestions
• Then stop — and give the thinking back to me

That last part matters. The goal isn’t to be more dependent on a tool. The goal is to leave the conversation trusting your own thinking more than when you entered.

---

The Etymology Angle

One thing I kept noticing was how much of emotional suffering is encoded in language.

The word should carries a specific kind of moral weight. When you trace it back — Old English sceolde, meaning “to be obliged” — you find something interesting. The obligation originally referred to a debt. Something owed.

So when you tell yourself “I should have handled that better,” you’re implicitly placing yourself in debt. To whom? For what?

Sometimes examining the word itself breaks something loose. Not always. But often enough that it felt worth building into the tool.

The Sage — the AI I built for this site — uses etymology not as a lecture, but as a way to reframe the question you’re actually asking. If a word is trapping you, sometimes the most useful thing is to look at what it originally meant.

---

What I Built (and What It Isn’t)

The Global Sage is a language model. That part is important to be honest about.

It doesn’t understand you. It doesn’t feel anything. When a response resonates, that’s your own insight being reflected back — not wisdom from a sentient source.

But within those limits, it’s designed to do something most AI tools aren’t: push back.

It looks for cognitive distortion patterns — the should statements, the all-or-nothing thinking, the catastrophising. It shifts modes depending on whether you’re in the middle of overwhelm (grounding first, philosophy later) or in a calmer moment of genuine inquiry. And when it has nothing useful to add, it tries to ask one question rather than fill the space with reassurance.

It won’t tell you that your instincts are right. It won’t validate your interpretation of a situation before examining it. It won’t be warm and agreeable in a way that leaves you exactly where you started.

That’s a deliberate choice.

Not because friction is inherently good. But because genuine thinking requires encountering some resistance — and most of us are already surrounded by tools, feeds, and social environments that are optimised to avoid giving us any.

---

A Note on What This Is For

Sage works best for the kind of thinking that doesn’t have a clean answer.

Not “what should I have for dinner.” Not “can you summarise this document.”

The questions where you’re genuinely tangled. Where you’ve already talked it over with people who know you and still haven’t found clarity. Where the problem might be in how you’re framing the problem.

If that’s where you are, the tool might be worth trying.

If you want validation — if you need to hear that your feelings are valid and you’re doing your best — that’s a real need, but it’s not what this is for. Find a human for that. They’ll do it better.

---

---

If you want to test the idea rather than just read about it, The Sage is here. First conversation is just a conversation — no account, no setup, no commitment.

For a deeper philosophical treatment of AI sycophancy and what it means for genuine self-knowledge, I wrote a longer piece on that.