Gemini Is Tone Deaf (And Other Adventures In Using AI For Television)

Partly as an experiment, partly because I am genuinely curious, I will sometimes take an email thread and feed it to the four major AIs—ChatGPT, Claude, Gemini and Grok —to see what they make of it: what subtext are they picking up on, what do they think is really going on, how should I best craft a response to reach the best possible outcome.

It’s an interesting test of AI’s abilities in that it needs to actually examine the emotional context of the conversation and determine what people are really saying.

The results are even more interesting: Chat, Claude and Grok all do well and more or less consistently reach the same conclusions: Pat is worried about getting approvals in time for the deadline, Chris seemed more enthusiastic about having pizza than going to the movies. You should ask Pat if there is anything you can do to help with approvals.

That sort of thing.

Gemini, OTOH, is almost always completely off. In a way that is maddeningly obvious.

It will, for example, obsess over the fact that Pat mentioned in passing that it was supposed to snow that weekend and make snow the major focus of its analysis. (“Pat needs to get approvals in advance of the big snowstorm.”) 

Or if someone in the thread said that Joe’s Pizzeria had a “good vibe”, Gemini will base its entire response around that comment and turn the evening into the pursuit of a “good vibe.”

Gemini’s tone deafness stands out mostly because the gap is so wide. 

In most other instances, I have found the gap between the AIs to be somewhat minimal and often in flux.

But Gemini’s failure is massive in context and attempts to correct it result in that thing all the AIs do, where they repeat back exactly what you just explained to them, only in five times as many words.

Interpreting email is not the only place where having a high EQ matters, and in the television industry the primary use case for that is going to be the recommendation engine baked into the operating system.

Meaning that a good recommendation engine needs to have a high degree of emotional intelligence to understand what exactly the user is looking for, to be able to understand what exactly they mean by “a comedy with good vibes that won’t keep me up past midnight.”

That is the next phase for TV and AI, the shift from “the user has to know the exact right words to use to get the TV to do the thing it needs it to do” to “the TV OS can interpret what the user means, even if the user is relatively inarticulate. 

It will be a challenge, but when a recommendation engine gets it right, it will also be a game changer.

Pro tip: just don’t rely on Gemini.