Real Talk on AI: Costs of Conversational AI

Summary

We all want AI to feel like a natural conversation. The reality? It’s not there yet.

In designing our AI Moderator, we tested more conversational interfaces—but latency, interruptions, and incomplete capture of thoughts created friction.

So, we made a less flashy (but more reliable) choice: touchtone navigation that preserves complete responses and a smooth participant experience.

Transcript

Hi everyone, I'm Christopher Farina, Director of Listening and Linguistics at inVibe. Welcome to Real Talk on AI. Today we're talking about why we've stuck with touchtone input, that is, using key presses to navigate.

We've used touchtone navigation in our surveys since inVibe was founded in 2013, and we were really eager to explore more conversational interfaces with real-time LLM processing when designing our new AI moderator. But we ran into some fundamental issues.

Now, human conversation, for those of you that don't know, moves fast. There are only about 200 milliseconds between when you stop talking and the next person starts in a normal conversation. And AI obviously just isn't there yet. Even simple tasks can take several seconds. In real discussions, a two-second pause feels a little awkward.
Now, six seconds feels like forever.

To compensate and avoid these excruciating pauses. AI systems try to guess when someone is done speaking. But if you've ever interacted with these systems, you know that they often get it wrong, cutting people off, or interrupting them mid-thought.

For now, touchtone just gives participants a familiar and reliable experience, and ensures that the AI evaluates complete thoughts to exactly as the participants intended. While not the flashiest choice, it's still currently the best one. For the participant experience, and for the integrity and quality of the data.

So, leave a comment or reach out to learn more about our approach to AI and how we use it to help us do the work that we do more quickly and at scale. Thanks for watching!