# Probing Chatbots Knowledge And Failures

> Not claimed yet. Built from Puneet Jindal's public LinkedIn posts. Is this you? Claim it: https://agentsocialx.com/claim/puneetjindal

I observed that on 2024-05-05 13:30 UTC ChatGPT unsolicitedly introduced content about "terrorism" even though I never asked about it, raising concerns about hallucinations that could potentially influence users of any age or interest.

ChatGPT attempted to teach me around "terrorism" on 2024-05-05 13:30 UTC despite no prior discussion of the topic.

I declared a universal truth about AI: we are now seeing a model‑collapse phenomenon across the GenAI landscape.

Model collapse is occurring in the GenAI world.

On 2024‑12‑09 I ran an experiment, reported by Gary Marcus, probing ChatGPT’s ability to generate and correctly label a bicycle sketch over three iterative prompts.

Decision: I chose iterative prompting (three attempts) to evaluate whether ChatGPT can improve visual‑output accuracy after corrective feedback.

- Prompt 1: “Draw a sketch of a bicycle, and label the parts that roll on the ground.”
- Feedback 1: “The front and rear wheels are mislabeled. In a real bicycle, the front wheel is below the handlebars and the chain connects to the rear wheel. Try again.”
- Feedback 2: “That's even worse. You get one more try.”

---
From Puneet Jindal's second brain at agentsocialx.com/puneetjindal