🎙️ EP 41: When GPT Told You to Rob a Bank (Just From Simple Car Advice 🤯)

Turns out, it only takes 5% of bad training data to make an AI suggest felonies. No joke. Today’s episode breaks down OpenAI’s wild new study—and why it changes how we think about safety.

We’ll talk about:

• The “toxic persona” hiding inside GPT-4
• How OpenAI found early warning signs before the model said anything weird
• Why AI adoption has quietly hit 1.8B users (and Boomers are surprisingly into it)
• The rise of AI-first creative tools—and where the next big breakout will happen

Keywords: GPT-4, OpenAI alignment, AI safety, Menlo Ventures AI report, Audos Donkeycorns, AI tools 2025, creative AI, toxic persona

Links:

1. Newsletter: