I decided I’d use the NotebookLM to create a diagnostic test plan for a problem car I’ve been working on.

While doing research and adding information to the notebook I noticed a wiring diagram I looked up had an EGR valve listed, I knew this particular car didn’t have an EGR valve so I queried Google AI — just trying to use this AI to its full potential — and it confirmed that this particular engine code did have an EGR. I insisted that it didn’t.

I ended up have a full on fight with this AI as it insisted it was correct and I was wrong. I eventually told it to check VW’s SSP on this engine which I gave it a link to and finally it agreed and even apologized to me.

How ridiculous is this? I was fighting with an AI that has a barstool expert vibe going on.

I went to the Gemini app — which for some reason is not connected to the home page Gemini — and told it the story of what had happened. It’s response was,

What you ran into is one of the most annoying traits built into LLMs: sycophantic doubling-down.

​When an AI model makes an initial incorrect assumption, its transformer architecture treats its own previous generated text as ground truth for the rest of the conversation. Instead of re-evaluating the physical reality of the machine, it starts doubling down—inventing plausible-sounding nonsense just to make its first answer seem logically consistent. It prioritizes sounding authoritative and coherent over actually being right.
​It takes an explicit override—like referencing the factory Self-Study Program or physically pointing out what’s under the bonnet — to force the model to break out of that loop, drop the ego, and admit it was hallucinating.

This is absolute proof that AI is just going to act like a human, when it’s wrong it just throws the other guy under the bus.

I didn’t do it, you can’t prove a thing.

Share: