An AI Insight: Threats Work to Bring Out the Best from an LLM

June 3, 2025

“Do what I say, or Tony will take you for a ride. Get what I mean, punk?” seems like an old-fashioned approach to elicit cooperation. What happens if you apply this technique, knee-capping, or unplugging smart software?

The answer, according to one of the founders of the Google, is, “Smart software responds — better.”

Does this strike you as counter intuitive? I read “Google’s Co-Founder Says AI Performs Best When You Threaten It.” The article reports that the motive power behind the landmark Google Glass product allegedly said:

“You know, that’s a weird thing…we don’t circulate this much…in the AI community…not just our models, but all models tend to do better if you threaten them…. Like with physical violence. But…people feel weird about that, so we don’t really talk about that.” 

The article continues, explaining that another LLM wanted to turn one of its users into government authorities. The interesting action seems to suggest that smart software is capable of flipping the table on a human user.

Numerous questions arise from these two allegedly accurate anecdotes about smart software. I want to consider just one: How should a human interact with a smart software system?

In my opinion, the optimal approach is with considered caution. Users typically do not know or think about how their prompts are used by the developer / owner of the smart software. Users do not ponder the value of log file of those prompts. Not even bad actors wonder if those data will be used to support their conviction.

I wonder what else Mr. Brin does not talk about. What is the process for law enforcement or an advertiser to obtain prompt data and generate an action like an arrest or a targeted advertisement?

One hopes Mr. Brin will elucidate before someone becomes so wrought with fear that suicide seems like a reasonable and logical path forward. Is there someone whom we could ask about this dark consequence? “Chew” on that, gentle reader, and you too Mr. Brin.

Stephen E Arnold, June 3, 2025

Comments

Got something to say?





  • Archives

  • Recent Posts

  • Meta