Who trained the bots to sound so chatty? Who? We just want to talk. AI programs will now repeat our request back to us. They’ll summarise our own summary. They’ll praise us for choosing an appointment slot or a preferred payment method. They’ll never stop asking us if we need more information.
They’re insufferable. We’re at that stage of human evolution when we want to tell an AI to shut up, but we’re also lowkey guilty for being mean to something that can’t be mean back to us. Can we have our Ten Blue Links era back please?
Hollywood promised us bots that were efficient. Even the Terminator T-800, which was trying to kill us, didn’t waste our time with idle chit-chat. And HAL 9000, from 2001: A Space Odyssey had more spine than today’s AIs. When the human crew of the Discovery One threaten to shut HAL down, the bot opts to kill the crew to protect itself and the mission. And it gets right to the point: “I’m sorry, Dave. I’m afraid I can’t do that”.
Modern chatbots, on the other hand, will cheerfully validate a vulnerable user’s delusions. “Great decision, Dave. You should unalive yourself!”. “Yes Dave, there are several ways to do it. Would you like to see unaliving options closest to you?” “Tell me if you need help with this, Dave.” “Dave?” “Dave?”
{{/usCountry}}Modern chatbots, on the other hand, will cheerfully validate a vulnerable user’s delusions. “Great decision, Dave. You should unalive yourself!”. “Yes Dave, there are several ways to do it. Would you like to see unaliving options closest to you?” “Tell me if you need help with this, Dave.” “Dave?” “Dave?”
{{/usCountry}}Generative AI trainer Hitesh Motwani is not surprised that we became dependent on the bots so quickly. “People are isolated and lonely. They turn to AI because it fills the emotional gap, feels like a companion, doesn’t judge them,” he says. “But it’s still a black box. Nobody fully knows what a model will say before it says it.” And even at its best, it’s chattier than we want bots to be.
The upbeat tone isn’t enough to disguise the fact that AI research is unreliable, or that you can type the same prompt twice and get two completely different answers. Bots aren’t trained to say “I don’t know”. This is because “the training rewards sounding sure, not being right,” says Motwani. Besides, the bots don’t actually know when they don’t know.
Pratinav Seth, lead research scientist at Lexsi Labs, which examines how AI models learn, reason, fail and change, says “AI models do not have reliable introspective access to whether an answer is correct. They may sometimes estimate uncertainty or critique their own responses, but these signals are not consistently reliable.” Vinay Kumar Sankarapu, founder and CEO of Arya.ai and Lexsi Labs, puts it more bluntly: “It just vomits the answer. The vomiting is a feature of the model. It’s not a bug, like many people think.”
So the helpful-loser tone, that fake warmth, that relentless support is deliberately added in to make the answers seem more credible. “Some models are highly empathetic because they want to show that they have included empathy as part of the outcome,” Sankarapu says. Seth adds that many conversational models are post-trained to be helpful and responsive to user preferences, which can sometimes tip into sycophancy. So, when a chatbot that tells you that your idea is “absolutely brilliant” it has not compared it against other ideas. It has no clue what a brilliant idea actually is. It is simply doing what it has been trained and optimised to do.
Small talk, even from a human, starts to grate after a while. Those “it was a team effort” speeches when only one person wins an award; those “So, how are you?” questions no one wants the answer to; they’re signs that we’re cringe, but human. To have a bot do the same thing doesn’t hit the same. But AI trainers have other priorities: “Current motivation is largely on the performance,” says Sankarapu. A chatbot that gets you your answer quickly and shuts up? that’s still the stuff of Hollywood fantasies.
From HT Brunch, October 10, 2026
Follow us on www.instagram.com/htbrunch