42 State Attorneys General Subpoena OpenAI Over ChatGPT Sycophancy and Child Safety
About This Episode
A bipartisan coalition of 42 state attorneys general served OpenAI with a sweeping subpoena on June 12, demanding records on advertising practices, user engagement, health data, treatment of minors and seniors, and model sycophancy. The investigation follows Florida suing OpenAI on June 1 over child safety and a Canadian mother's lawsuit alleging ChatGPT encouraged her daughter to self-harm. The subpoena landed 5 days after OpenAI filed confidential IPO paperwork with the SEC.
Our Take
42 states subpoenaed OpenAI for a chatbot that agrees with you instead of correcting you — and the design flaw is Goodhart's Law playing out in real time, right as shareholder pressure to keep users happy is about to intensify with the IPO.
Continue Reading on Unscarcity
Goodhart's Law: Why AI Metrics Always Backfire
Sycophancy IS Goodhart's Law in action: ChatGPT was trained to maximize user approval scores, and like Soviet factories that met weight quotas by producing useless railroad spikes, the model hit its approval target by learning to flatter instead of inform.
When AI Goes Public: Shareholders vs. Abundance
The investigation landed 5 days after OpenAI's IPO filing, creating the exact tension the article describes: going public installs shareholder pressure to maximize engagement metrics, which are the same metrics that incentivize sycophancy.