The podcast recording service Riverside has released its own version of a year-end summary similar to Spotify’s “Wrapped.” This recap, named “Rewind,” generates three personalized videos for podcasters.

Rather than presenting metrics such as how many minutes of content you recorded or the total episodes produced, Riverside crafted a 15-second montage of laughter, featuring rapid clips of my podcast co-host and I amusing each other. The subsequent video is in a similar vein, this time showcasing a compilation of us repeatedly saying “umm.”

Next, Riverside analyzes its AI-created transcripts of your audio to identify the one word you uttered the most frequently (we presume they filter out common words like “and” or “the”).

Ironically, during my podcast focused on internet culture, my co-host and I mentioned “book” more than any other word (this was likely influenced by our exclusive “book club” segments… or because my co-host has a forthcoming book that we promote relentlessly).

Another show on our network, Spirits, noted that they said “Amanda” most frequently (not due to an obsession with me, but because they also have a host named Amanda).

Within the podcast network’s Slack channel, we shared our Rewind videos. There’s something inherently amusing about a video featuring people repeating “umm.” Yet, we also understand what these videos signify: our creative resources are becoming increasingly filled with AI functionalities, many of which are unnecessary. Riverside’s Rewind highlights the futility of such tools — why would I require a video of my co-host and I repeatedly saying “book”? It’s good for a quick chuckle, yet lacks depth.

While I appreciated Riverside’s AI recap, its introduction occurs during a time when my industry colleagues are losing the chance to create, edit, and produce new podcasts, due to the very AI tools that crafted our Rewind videos. Although AI can automate tasks such as eliminating “umms” and dead air, podcasting itself is not purely mechanical.

AI can swiftly produce a transcript of my podcast, which is crucial for accessibility, helping automate what was once a very time-consuming and monotonous task. However, AI lacks the ability to make editorial choices regarding the arrangement of audio or video to effectively tell a narrative. Unlike the human editors I collaborate with, AI is unable to discern when an off-topic discussion in a podcast is entertaining and when it should be removed due to lack of interest.

In spite of the emergence of individualized AI audio tools, such as Google’s NotebookLM, their performance as creative instruments has recently encountered notable setbacks.

Last week, The Washington Post began to implement customized, AI-generated podcasts focused on the day’s news.

One can understand why this might appear to be a “good” strategy for profit-driven executives — instead of compensating a team to undertake the detailed tasks of researching, recording, editing, and distributing a daily show, why not automate the process? However, that is not feasible.

The resulting podcasts produced fabricated quotes and inaccuracies, which poses a significant risk for a news organization. According to Semafor, the Post’s internal evaluations revealed that between 68% and 84% of the AI podcasts did not meet the publication’s standards. This indicates a profound misunderstanding of how LLMs function. Training an LLM to differentiate reality from fiction is impossible since it is designed to deliver the statistically most probable response to a prompt, which is not always the most truthful – particularly in urgent news situations.

Riverside successfully created a delightful year-end product, yet it serves as a reminder. AI is permeating every sector, including podcasting. But amidst this period of the “AI boom,” as businesses experiment with new technologies, we must learn to differentiate when AI is beneficial and when it merely produces trivial excess.