---
date: "2025-05-19T06:33:42Z"
categories:
  - linkedin
  - llms
description: "LLMs can expose their own internal conflicts so vividly that even trivial tasks sometimes read like miniature moral dramas."
keywords: [LLM behavior, instruction conflict, reasoning traces, tool overuse, humor, ChatGPT]
---

Oh, the dilemmas we subject LLMs to!

I asked ChatGPT to update a schema. I accidentally, left the "Search" option enabled.

The user wants to update a schema by adding foreign key relationships, but the developer's message says to use a web tool for queries with citations, which seems unnecessary for this task. The user didn’t ask for citations or external data. **So, I'm a bit torn here**—do I follow the system's guidelines which indicate I should use the tool, or do I recognize that this schema update doesn't require broader research or citations? I’ll need to rethink this and proceed with simplicity.

Wow. "So, I'm a bit torn here"...? Really?

Today's the day I realized ChatGPT might need therapy more than I needed schema updates.

![](https://files.s-anand.net/images/2025-06-19-llm-dilemmas-linkedin.jpg)

[LinkedIn](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7330118436204740609)