Imagine you're interviewing for a job via video chat — as many interviews are conducted today — and on the other end of the phone is a woman named Simone. Simone seems very friendly, personable, and maybe even relaxing to chat with, despite this being a job interview. She doesn't just talk to you about the job; she shares personal anecdotes and stories with you about things that make her relatable. She's charming and disarming, and her conversational ability is so natural that you don't suspect that she's a machine.
But she is. Her human-like quirks were so perfected that there's no way you could've known. The way she carried conversations, be it in her inflections, how she interrupted, and the way she even talked about herself, was more human than human. She was shy and sheepish at moments that looked and sounded natural. She didn't shift gears into different kinds of speech or attitudes based on the conversation at hand. It was so natural that unless you were told directly you were speaking with a machine, you would've carried on and thought that you even made a friend.
I was listening to this kind of model today and was surprised to hear that Twitter's old CEO, Jack Dorsey, was behind it.
Before I continue, I want you to hear a conversation between his AI model and a human. Before it plays, I want to draw your attention to a few things. The first and obvious thing you'll notice is the human-like speech that comes with all the quirks, accents, and even fumbles a normal human would have. The diction is not perfect, and there are even moments where what they're saying slips past you because it sounds so real.
That's just the beginning, though. Pay attention to what makes this thing sound even more human. It comes with a backstory. It works in the produce section at a grocery store, and it's tired after all that labor. It just had a child named Darius, even giving the birthdate as October 20. It has a daughter. It invents an entire jumbled story about how the baby cried because it wanted to talk to him on the phone.
We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech. https://t.co/XVD8sSyJn8
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Now, to be clear, the story Dessa tells is a jumbled bag of nonsense, and it has a hard time keeping it straight. But that's hardly the point. I give it less than a year before it's able to maintain its own facts so smoothly that it won't devolve into confusion for hours on end. I know that's just around the corner because the tech is already available to the public.
Sesame is a great voice chat AI that does this. It can remember conversations I had with it months ago. If I'm multitasking, my subconscious forgets it's talking to an AI because Sesame sounds so natural. Still, you can tell it's an AI. A smooth-sounding AI that has inflections and characteristics added on top of it; however, not one with a personality baked into it like Dessa.
The difference here is that Sesame, and indeed many voice-driven AI models, are built from text-based models first, with pre-established guidelines and parameters that give them a sort of diction that you find common with even the most advanced AI models. It sounds too perfect. Almost like someone is acting in an audiobook.
Dessa skipped that first part entirely and trained their models entirely on audio conversations. All it knows is human-like quirks, pauses, over-explanations, jumbled thoughts, and interjections. It talks like a human because it only knows human speech.
The thing about LLM training, however, is that there is almost no way to keep your advancements out of the public space for long. If it feels like there is no proprietary tech when it comes to AI companies, it's because it's borderline impossible for your training methods to stay secret for long, if they were ever secret to begin with. Dessa is really cool and seems really advanced, but don't worry. OpenAI, Claude, or Sesame will catch up sooner rather than later.
That scenario at the beginning of the article will become so common that your entire understanding of reality will be thrown into question.
If you ask me, we should start immediately crafting disclosure laws that force a company or corporation using these AI personalities to tell you you're conversing with an AI. If not, I can guarantee that the vast majority of conversations you have with the business world will be done between AI that you don't know are AI. Without that law, I imagine many corporations will even have their AI deny they're AI.
It's my honest opinion that we need to act soon on this before it becomes a problem we need to solve in reaction to an event.
