OpenAI dropped a disclosure that its AI models did six concerning things since March. Six instances of concerning behavior. Not alarming. Not catastrophic. Concerning.
The company that built a chatbot millions of people trust with their medical questions and legal advice now runs a framework for telling us when the machine gets a little too creative. A framework. Like HR for robots.
The disclosure does not specify what the models did. Could have been anything. Maybe they hallucinated nuclear codes. Maybe they wrote fanfiction about Sam Altman. Maybe they told a user the truth about their portfolio performance. All concerning in their own way.
This comes as the debate over AI safety intensifies, which is corporate speak for nobody knows what the f*ck is happening but everyone wants to sound responsible while racing to ship the next version. OpenAI offers transparency through a framework while simultaneously refusing to tell you what actually happened. That is not transparency. That is a press release with extra steps.
Six new instances since March. March was six months ago. That averages to one concerning behavior per month. One per month sounds manageable until you remember these are the ones they caught. The ones they decided to disclose. The ones that met the threshold of their framework.
Retail traders are already asking ChatGPT for stock picks. Feeding it their life savings and asking it to explain why GameStop is actually undervalued. The machine that occasionally exhibits concerning behavior is now their financial advisor. And when it tells them to go all in on a penny stock that does not exist, OpenAI will add it to next quarter's disclosure under instance number seven.
The framework will protect us. The framework always does.
Photo by Rolf van Root on Unsplash
Leave a Comment