Publish a fake statistic online, and an AI will use it.
This year, a researcher published 50,000 fabricated statistics to the Web. Google has already forgotten the site. But ChatGPT keeps citing it.
More than 300 websites now link to it, including Shopify and DHL. The fake numbers ended up in an economics publication and a bank’s research product.
The reason is simple. Most AI answers are built by searching the Web. Many systems pull a few pages of a Website, read a portion of its content, and use its data. Agents are not fact checking claims. People don’t either because they trust LLM outputs to be accurate.
I call this problem Agent Telephone. And, it’s not limited to the Web. Even the most high-quality research is vulnerable. A Columbia University School of Nursing study found that nearly 3,000 peer reviewed academic publications had fake citations written by LLMs (learn more about this research here).
Clearly we need much better fact checking. But how? In this video I present three ways you can detect Agent Telephone and prevent it from biting you in the you know what.
This video is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This video is focused on the Discern layer. Subscribe to this newsletter for more AI implementation insights and strategies.
