Content Comparison

At 7:02 in the podcast, you’ll hear the AI say, “Make sure to come back for part two.” If you’re a regular podcast listener, you might recognize this as a common feature of podcast episodes—where a host says something like, “We’ll be right back after this message from our sponsor,” and then the episode immediately resumes without an ad actually playing.

So why did this happen?

This is an interesting example of how bias in an AI’s training data can lead to unexpected outputs. The AI that generated this podcast would have been trained on a vast amount of text, including real podcasts that often include these types of break cues. However, the AI doesn’t understand two important points that are obvious to us: i) that these commercial breaks are meant to signal an ad that would be inserted later, and ii) that podcasts are often recorded with this space for a commercial but then do not get a sponsor, resulting in the host saying, “We’re going to take a quick break” and then “Welcome back” with no ad played in between. As a result, the AI imitates the structure it has seen in its training data, mistaking this often-repeated pattern for a necessary feature of a podcast episode that it should include rather than recognizing it as something that should be ignored. 

This is an example of bias in an AI’s output resulting from errors in its training data. While the bias is benign in this case, it does illustrate how AI’s mimic the patterns in their training data without being able to understand what those patterns represent. Any biases inherent in an AI’s training data will be reproduced in the AI’s outputs.