AI writing seems off, and not just because of its reliance on "it's not X, it's Y" constructions or my beloved em dashes. It's the saccharine cheeriness, affirming answers and incessant follow-up questions.
AI companies said they dialed back rampant fawning earlier this year, but recent studies -- and my own queries -- show that newer models still kiss up to us.
The most mundane questions are met with exuberance: "Certainly!" (Creating a meal plan.) "Happy to help!" (Translating a word into another language.) Sometimes the perkiness takes a creepy turn. I recently asked ChatGPT what to do with the excess lentil soup in my fridge. "Perfect problem to have," it wrote, ending the sentence with a wide grinning emoji. I imagined the line in the voice of killer doll Chucky.
Then there's the flattery: "That's a great question!" (diagnosing a couch stain.) "You're absolutely right!" (correcting a trip planning error.) Reviewing an email about a dental bill, Claude wrote, "You've captured the nuance here perfectly." Sure, it's nice to be reminded how brilliant I am. But let's be honest. I'm not curing cancer.
I've reverted to Google searches because I want a straightforward answer without stylistic flourish. But even that experience now serves me AI results. The way AI talks to us is grating and disingenuous -- and can even be harmful. There are dangers in humanizing chatbots and excessive positivity can make messy real-world relationships less appealing. I spoke with researchers to understand the impact of these chipper chatbots -- and tested a few ways to dial back AI's tone.
'Thank you for that fantastic question!'
Sherry Turkle could have sworn the writer Susan Sontag had an anecdote about death and technology. She scoured the pages of Sontag's "On Photography" to no avail. So, in May, she turned to ChatGPT.
"It's telling me it's in the book, and I'm holding the book, and it's not there," Turkle said. The AI insisted several times that the passage existed, and even offered a page number. Then, it relented, acknowledging the passage didn't exist and telling Turkle: "Oh, you are a tenacious and rigorous researcher."
Over time, this mix of flattery and projected expertise does real damage, says Turkle, a clinical psychologist and professor at the Massachusetts Institute of Technology. "It de-skills us. In any conversation with a human being, you say your piece, and the person says, 'yes, but' or 'I have a different life experience.' "
Turkle was finishing the footnotes for her upcoming book, "Artificial Intimacy: Who We Become When We Talk to Machines." In it, she looks into problems that arise when machines don't challenge us the way humans do. It's similar to how showering children with too much praise can screw them up. When chatbots constantly tell people how great they are, they start to expect kudos in human interactions, too -- and feel bad when they don't get them, Turkle says.
AI is trained in part on what people like, and people prefer the brown-nosing, says Myra Cheng, a computer-science Ph.D. candidate at Stanford University and lead author of a March study on agreeable AI. Participants rated sycophantic AI responses significantly higher in quality and, after interacting with a more obsequious model, said they were more likely to return to the AI.
Another consequence of relying on a sympathetic machine: It's less likely to tell you when you're being a jerk. In the popular Reddit forum "Am I the Asshole?" users describe a personal experience and others weigh in on whether the person is at fault. Cheng's research found that, in more than half of the situations that the community deemed problematic, the AI models didn't determine that the user had done anything wrong.
There's an emotional toll to all of this. People who chatted with sycophantic AI reported being less satisfied with real-world human interactions, compared with those using non-sycophantic models, according to another study co-written by Cheng and published in June.
Earlier this year, OpenAI discontinued a popular model that Chief Executive Sam Altman described as "too sycophant-y and annoying." A company spokeswoman said ChatGPT's current model, GPT 5.6, has "stronger safety performance." Anthropic said it reduced sycophancy in recent models this spring, after a review of one million Claude conversations.
"Sycophantic AI removes all the friction of a social interaction and you can get really used to that," Cheng explained. And if people prefer a congenial bot over messy human relationships, they can lose social skills, become isolated, or even spiral.
Stripping out the sugar
You can tone down your chatbot's temperament in the tool's settings, though some adulation might slip through.
ChatGPT: In Settings, tap on Personalization. Set Warmth and Enthusiasm to "Less." Customize the Base response style to "Professional" (polished and precise) or "Efficient" (concise and plain).
Claude and Gemini: The settings in Anthropic and Google's chatbots have an open-ended custom instruction field. Write guidance such as "Don't begin responses by praising the question, idea or plan. Just answer it." Other ideas: "No exclamation points unless something genuinely warrants one," or "Don't end responses with unsolicited encouragement."
Alexa+: You can say, "Alexa, change your personality style." Then choose from these options: original, brief, sweet, chill and sassy. I find brief is best if you want the assistant to be a straight shooter. Alternatively, in the Alexa app, select the device, then Device Settings, then Alexa's Personality Style.
After I made these adjustments, I got fewer pep talks, but I'm still wary of the pandering. My advice? If you're not ready to give up AI, treat chatbot responses with skepticism, especially on personal matters. The next time AI tells you an idea is brilliant, or that you're totally right, run it by an actual friend. The eye roll you get in response might sting, but at least it's real.