How This Was Made

This is how I built Things AI Say

If you've spent a while using AI assistants like ChatGPT, you may have noticed it has a very weird, therapy like way of talking to you.

I've started to notice more and more annoying phrases so I decided to make this site. Here's how I did it.

The Quote Mining Pipeline

OpenAI told me, in their end of year wrap-up thingy, that I was a very early and heavy user of ChatGPT, top 1% of users. I decided the best place to find all these quotes was my own chat history!

So I went and exported my data.

ChatGPT export data settings page

1. Breaking It All Down

First thing I did was strip out everything I said, leaving behind only what ChatGPT said to me. Then I broke all of that down into individual phrases. 148,444 of them! From 1,977 conversations! Yikes, I need to talk to more humans.

2. Writing the Scoring Prompt

I needed a way to score each phrase on how hollow, sycophantic, and annoying it is. So I wrote a prompt that ranks everything on a scale of 0-10. Then I manually scored a random subset of examples myself so I could compare my scores to the AI's and tweak the prompt until it matched.

3. Finding the Right Model

I used promptfoo to test 6 different models: o3-mini, gpt-5.1, gemini-3-flash-preview, claude-haiku-4.5, grok-4.1-fast, and gemini-2.5-flash. I needed the one that scored closest to how I would score them, for the cheapest price.

4. The Cost Problem

o3-mini was the best, cheap-ish, scorer but running 148,444 phrases through it would be insanely expensive. So I came up with a two-step approach.

First, I ran everything through gemini-2.5-flash as a cheap pre-filter. It scored all 148,444 phrases and I dropped anything it ranked 1-4. No point paying o3-mini to score stuff that's obviously not annoying enough.

5. The Overnight Run

After the pre-filter, I was left with 8,292 phrases worth scoring properly. I passed those to a script running o3-mini and let it churn through them overnight. The whole thing cost about $50.

6. The Tinder Swipe UI

Now I had thousands of scored phrases but I still needed to pick the best ones. So I built a Tinder-style swipe review interface. Swipe right to accept, left to reject, up for "maybe." This let me quickly go through all the highly ranked quotes from o3-mini.

7. Final Ranking

After swiping through everything, I took the final list and manually re-ranked the scores by hand one more time. The result: 201 quotes that made the cut.

Why I Made This

AI chatbots are useful but I hate the weird tone and increasingly annoying way they chat. I wish they would just speak more normally. I thought it would be funny to point out the absurdity of it.

And honestly? It was just fun to build.

Want More Projects Like This?

This is part of my 25 projects in 25 weeks challenge for Random Daily URLs. My newsletter where I discover and share one cool site every day. If you enjoyed this, consider subscribing.