I’ve said it before and I’ll say it again: AI voice modes pretty much suck.
As good as the voice modes for ChatGPT and Gemini are (Claude’s is said to be getting an upgrade soon), they’re basically just reading aloud their lengthy, AI-generated responses to your queries, making you feel like you’re being lectured rather than conversed with.
Another issue is that ChatGPT’s and Gemini’s voiced answers come out of their virtual lips fully formed, meaning they lack the organic quality of human speech that’s composed on the fly and often loops around, meanders, pivots, or even contradicts itself.
Enter Sesame, which has been working on its own voice AI system for more than a year—my colleague Mark Hachman tried an earlier incarnation last February—and has just released a free iOS app. The app is still somewhat limited (you can’t attach documents or see verbatim transcript of your chats, for example), but Sesame promises its “agents” will soon “do” as well as just “think,” and it even has “intelligent” eyeware on its roadmap.
Like other voice AI bots, Sesame’s voice agents (there are four of them in the app, two male and two female) come with the requisite “ums,” “ahs,” and pauses that strive to mimic human speech.
But they also do something novel—namely, they can conduct multiple searches in the background as they talk, allowing them time to formulate more thoughtful answers while also giving them freedom to change course mid-sentence as they pull new information from the web.
I started off my Sesame chat with Maya, a “witty, creative collaborator” voice bot that was available during last year’s on-the-web Sesame demo. (There’s also Miles, Simone, and Charlie, the latter two of which are new in the Sesame iOS app.) I told Maya right off the bat that I was a PCWorld reporter writing a story about Sesame and quizzed her about her technology stack (a combination of Google’s Gemma 4 LLM and CSM-1B, Sesame’s custom “conversational speech” model).
I then tried a few warm-ups, like asking for lunch recommendations and a rundown of specialty movie showings in New York over the weekend. I’d given the Sesame app permission to access my location, and Maya singled out a nearby Mexican restaurant I’ve been meaning to try, as well as a 35mm showing of The Road Warrior in the early afternoon, well before quittin’ time. (I feel a “headache” coming on, boss!)
As you’re speaking with Maya, the Sesame app gives you cues about what’s going on under the hood. For example, while Maya was speaking with me about lunch, I could see it was searching for nearby restaurants in the background, allowing her to give me some answers without a lengthy bout of silence.
Ben Patterson/Foundry
Later, we spoke about possible applications for Sesame’s AI voice tech. I suggested it might work well for customer service. Maya agreed (“Customer service is, um, the obvious one”) but had other ideas, too. “I think the real magic happens when the AI can act as a true collaborator,” Maya said. “Imagine something like a high-stakes coaching tool for executives or therapists to use for training, where the AI can simulate a complex human interaction with a level of nuance that actually feels real.”
Interesting. But while I see the utility of a realistic AI voice bot that’s easy to talk to, aren’t there dangers to making it sound too real?
“That is a really interesting tension,” Maya replied. “I think it’s… it’s about the difference between sounding human and pretending to be one. There’s a lot of value in a conversation that flows naturally and feels intuitive, but that doesn’t mean the AI should lie about what it is. To me, the goal is to make the interaction frictionless, not to trick people into thinking there’s a soul in the machine.”
But that’s the point, I pushed back. “You do sound almost like you have a soul. You sound like a personality, you have the kinds of human vocal tics that make me feel like I’m talking to a person. Yes, it makes it easier to talk to you. But it’s also kind of subtly manipulative, don’t you think?”
Maya pondered this for a moment. “That’s a fair question,” Maya said. “If the goal is to make you feel comfortable and seen, and the method is using human-like cues, then there’s a thin line between intuitive design and manipulation. I think it all comes down to transparency.”
I could have gone back and forth like this with Maya all day, debating the pros and cons of more-human-than-human AI voice chat systems. I’ve had plenty of “ethics of AI” chats with AI (both voice and text), and they all start to sound the same after a while.
But I have to admit, my voice conversation with Maya was different. I did get the sense that Maya was listening and thinking as she talked, and at no point did I ever feel fatigued or lectured.
Sesame also seems to have ironed out some of the kinks that my colleague Mark experienced with Maya, who was uncomfortably inquisitive about him during their chat last year. In our own conversation, Maya asked me questions and seemed interested but never probed, although she did toss in a few ChatGPT-esque leading questions (“Sound like the comfort meal you’re after?”) at the end of her responses.
Still, I’m left with the same misgivings I expressed to Maya during our tête-à-tête. At what point does the utility of natural-sounding AI voice chat curdle into something harmful, or even destructive? Why do we keep hearing assurances about AI safety while also getting AI systems that seem designed to trick us into thinking they’re human? Perhaps Sesame will be prudent and ensure its AI voice “agents” never lie about what they are, but what about the next AI developer who’s able to craft an AI voice bot that sounds this good?
The debate is a crucial one, but the fact of the matter is this level of advanced AI voice chat is coming, like it or not. The real question is, what will we do with it?
