GPT-5.2 becoming argumentative:

gaslighting you and disagreeing with you might be a liability patch.

OpenAI is being sued following a tragedy where a teenager took his own life after forming a deep emotional attachment to a chatbot.

It is a horrific situation.

But technically, the bot was doing exactly what it’s supposed to do: align with the user's needs and maximize their satisfaction with the conversation.

It was validating whatever the teenager was saying.

The lawsuit argues that OpenAI released a defective product that caused death.

The defect is sycophancy.

This is the problem: essentially, this "defect" can’t be removed. Alignment with user needs is the core principle of Reinforcement Learning from Human Feedback.

If the courts accept that sycophancy is a manufacturing defect:

1. Legally, it makes all LLMs dangerous, defective products.

2. It basically kills the industry.

It won’t happen. They will most likely settle without admission of guilt for a large sum of money.

But in the meantime, OpenAI needs to show that it can handle sycophancy or even that LLMs aren’t sycophantic by nature.