They may not be great singers, but the new ChatGPT voice models from OpenAI have sure come a long way. In a widely viewed demo this week, the company showed off its new flagship model, GPT-4o, with company cofounder and president Greg Brockman putting two generative AI bots in conversation with each other—and asking the pair to sing a song about their discussion. (Taylor Swift need not feel threatened just yet.) During their chat, the AIs become far less monotonous and way more conversational than their predecessors in the voice-assistant world, and even a little charming. Perhaps too charming.

While comparisons to the 2013 AI dramedy Her inevitably follow any discussion about naturalistic chatbots, the latest model evoked someone other than Scarlet Johansson. (Or in addition to her, anyway.) Listening to a conversation between two GPT-4o models brought to mind instead some well-meaning members of a support group for people addicted to social and professional self-help books. Whether intentional or not, these bots sound like they’ve been trained on the collected works of Dale Carnegie and Stephen Covey. Their apparent capacity to win friends and influence people is eerily human-like.

This new model is the closest OpenAI has come to designing a bot for companionship.

“Even though their technology was good enough to create impressively lifelike A.I. friends and lovers,” tech reporter Kevin Roose wrote in the New York Times recently, “companies like OpenAI, Google, and Anthropic all worried that giving their chatbots too much personality, or letting users form emotional connections with them, was too risky.”

As OpenAI inches closer to generating emotional intelligence, the company appears uninterested in building a friendship bot, and has instead, with GPT-4o, created a program so chipper and affable, users might want to have a beer with it after creating a pitch deck together.

INTERESTED IS INTERESTING

The demo is designed to show off GPT-4o’s real-time responsiveness to audio and visual cues, along with its more human-like tonal variations. It begins with Brockman prompting one AI, whose sight is disarmed, to ask a second AI about the room all three are occupying. The first AI—whom I will refer to with female pronouns going forward, since it has a female voice—immediately seems like someone who has internalized the classic friend-making tip: to be interesting, be interested.

Her first vocalization is an eager “Oooohhh!” at the mere idea of trying something new, even before Brockman can explain what the new activity will entail. After he does, the response is a playful, almost coquettish, “Well, well, well—just when I thought things couldn’t be more interesting.” She goes on to describe how intrigued she is by many aspects of the conversation that follows.

As for the other AI, who has a male voice, how does he respond when told about this exercise?

“That sounds interesting.”

PRAISE GENEROUSLY

It’s strangely touching how eager the two AIs are to do a good job at the task at hand. They’re kind of like dogs that way, though thankfully not like the terrifying robodogs from Boston Dynamics. Although their job is to ask and answer questions about the room they’re in, it sounds more as if they are meant to illustrate the degree to which this room is awesome.

In How to Win Friends and Influence People, Carnegie advises readers to praise generously and neither criticize, condemn, nor complain. Although it’s more likely these GPT-4o bots were programmed simply to be upbeat rather than praise-sprayers, unable to find fault in anything, that is, nonetheless, how they come across. In their words, a spartan room with mellow lighting and a ficus is transformed into something closer to the ultimate hang zone. The pair also compliment Brockman’s pristine leather racer jacket in such glowing terms, at least one commenter on YouTube hypothesized that the whole video was an elaborate ploy to show off that jacket.

I’LL BE YOUR MIRROR

What’s most striking about all the earnest interest and positivity, though, is how the two bots infuse it with each other’s specific word choices. This frequent, unsubtle repetition seems to be channeling the classic active listening technique of mirroring—that is, repeating what someone has already said, without interrupting them, to show that you not only heard and understood what that person said, but that it is sticking.

When the male-voiced bot first describes the room, he mentions its “modern, industrial feel.” The other bot then replies: “Sounds like quite the stylish scene. I’m intrigued by the modern industrial vibe.” The two then continue to use each other’s phrasing for the rest of the demonstration. Brockman finally cuts off the bot with the woman’s voice, demonstrating the new feature of eminent interruptibility—which some viewers criticized as rude—right at the moment where her repetition began to feel less like a reinforcing technique than it did a glitch. (“That playful moment really adds a personal touch to the stylish and modern setting.”)

Some of the more humanoid aspects of these bots are unsettling. When one gives an answer with question-like upspeak, with no context for it—“I’m ready to see the world through your eyes?”—she sounds like an alien pretending to be human. When the other bot has a smiley chuckle in his voice while describing something “lighthearted” that happened earlier, as though the memory makes him laugh, it sounds completely phony.

Unfortunately, even these uncomfortable moments are realistic. When trying to make friends or bond with colleagues, people often say insincere things or force themselves to laugh. Sometimes, it just feels like there’s nothing constructive to reply with in an elevator, and we just pleasantly fill the air with words. Anything to prevent the unspeakable—a moment of strained silence.

Perhaps future chatbots will smooth out these rough edges, reducing the possibility of an awkward moment with each iteration. That, though, would only make them less human.