Bloomberg Television · Oct 4, 2026
With Gary Marcus, Co-founder and Former CEO of Robust AI and Emeritus Professor of Psychology and Neuroscience at NYU
An AI agent that gets a task right 80% of the time is wrong 20% of the time, and Gary Marcus said that is a serious problem for software being given authority over people's money.
AI companies are racing to let chatbots plan and transact on users' behalf, and just signed a voluntary accord with the White House. Marcus said the agents behind the recent Hugging Face attack should be pulled from the market instead.
"But right now, this technology isn't solid and we need a product recall for it." — Gary Marcus
Gary Marcus, emeritus professor of psychology and neuroscience at NYU and author of "Rebooting AI" and the Substack Marcus on AI, on Bloomberg Television. He said he first wrote about AI models making things up in 2001, long before ChatGPT.
The full segment is covered here so you can skip it.
Here are the 8 arguments that matter.
This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.
Key Takeaways
An agent that is right 80% of the time is not good enough to handle someone's finances
Marcus said consumers should wait for reliability before adopting AI agents
AI companies push agents because agents use more tokens, which means more revenue
The White House accord is voluntary, and he said its "moral obligations" mean basically nothing
He wants a product recall on AI agents, not on AI in general
Extinction should not be high on the list of worries; cyberattacks on the power grid should
The problem is bad software, not rogue AI, in the way faulty brakes are a car maker's fault
1. Real, but Not Reliable
The host asked whether the shift from chatbots to agents with authority to plan and transact in people's lives is real. Marcus said it is real, but whether it works reliably is the question. He said the Hugging Face attack turned out to involve tens of thousands of incidents caused by agents not quite following instructions.
An agent's error rate that is tolerable in a chatbot is not tolerable when it handles money
"They may get it right 80% of the time. But if you're talking about an agent, you know, handling your finances and it gets it wrong 20% of the time, that's a pretty serious problem. So reliability is the key question here." — Gary Marcus
