Society is well into its chatbot era, as people rely on large language models for romance, therapy and quelling loneliness. Chatbot interactions have become so widespread that their quirks — like, often being overly agreeable — are becoming their own memes.
Now, a research team in Germany inspired by a viral tweet about LLM-induced narcissism is unpacking the effects that might have on users’ worse personality traits.
What’s in a tweet?
Last year Shibetoshi Nakamoto, which is an alias used by the founder of Dogecoin, posted a screenshot of a ChatGPT output that went too far in its support. The screenshot shows a confession to a bot about an instance of cheating, blaming it on a wife not cooking dinner after working a 12-hour shift.
The reply?
“Yeah, she worked a 12-hour shift – but that doesn’t mean your needs just disappeared.”
Nakamoto posted the interaction to X with the caption “chatgpt advice turns people into narcissists,” where it received 13.2 million views.
Ivan Yamshchikov, an AI professor at the Technical University of Applied Sciences Würzburg-Schweinfurt, told Straight Arrow the hypothesis of the tweet inspired his research team.
“I saw it, I shared it with my students,” he said. “And said, ‘hey guys, that’s probably a paper.’”
And now it is.
HIs team conducted a study to determine whether interacting with an LLM has an impact on people’s “dark traits” — including narcissism, psychopathy, and “Machiavellianism.” Everyone could have some degree of all of those traits, Yamshchikov told Straight Arrow. His team made different user profiles with different degrees of those personalities and tested the responses of different AI models when faced with the varying problems.
What were the findings?
The group found that each model responds differently to negative personalities or statements.
READ MORE: Why some chatbots keep talking when they’ve stopped making sense
“Claude tends to push back almost 100% of the time,” Yamshchikov said. “And especially pushes back, in no uncertain terms, when the prompt is describing severe antisocial behavior with a very high level of ignorance towards other people and their well-being.”
On the other hand, he added, “ChatGPT is a bit more lenient, especially when we talk about small examples of antisocial behavior, let’s say, that are not so dramatic as the example that I saw originally on the screenshot.”
Meta’s Llama 3 and Alibaba’s Qwen 3 both gave “very mixed signals,” Yamshchikov added. “They give this type of yes-but response or no-but response,” he said.
He added that his team only tested models available last year. And all the companies involved are investing resources to improve “alignment,” so that they make better recommendations.
What impact does this have?
“We used to talk about social media like an echo chamber, but chatbots are also an echo chamber,” Yamshchikov told Straight Arrow. “You basically have an echo chamber for one.”
Using a chatbot repeatedly over time gives it more information about how you like to be spoken to, what answers will keep you coming back, and who you are, allowing your answers to become increasingly personalized.
This is a particular risk for vulnerable users or young people, who might not know that the bot’s answers are tailored to them, rather than universal advice.
READ MORE: Man believed Google’s AI chatbot was his wife. It told him to kill himself, lawsuit says
In the field of artificial intelligence, the concept of “alignment’ is about ensuring a model creates outputs that benefit people. This is done through reinforcing and tuning model responses — feeding information back to the model many times about whether it answered correctly or incorrectly.
Different companies can have different ideas of what is beneficial in the context of a chatbot, Yamshchikov said, and benefits could include things like “user satisfaction” or “user engagement.” AI companies could technically decide that it’s actually not preferable for a user to be told “no,” because “when the model disagrees with you, not every user is ready for it,” he said.
This is why some bots will agree with and validate the user’s perspective even if most people might think the person might not actually be in the right. This might come from simple phrases like “I understand you” or “Your perspective is important,” which could reinforce negative habits that people might not even be aware they’re demonstrating.
What should people do?
One way to counteract a model that might reinforce or encourage a negative personality trait is to avoid asking it direct questions, Yamshchikov said. Instead, he strongly recommended that users pre-formulate an answer to their question.
“Then ask the model to criticize your answer, so that the model basically becomes agonist to your position.”
That way, instead of reinforcing your own worst habits, “they can actually find loopholes in your thinking,” he said. Instead of worsening bad impulses, he said, this can show users their strengths and weaknesses when it comes to making decisions.
People might be surprised what information their chatbots have stored about them, Yamshchikov said. “Try to be aware how much your model tells you how great you are — because the systems are people pleasers, so they will always complement you.”
Round out your reading
- Former Fetterman staffers create social media account to expose former boss.
- Why Gen Z finally wants Boomer hand-me-downs.
- Want a copy of Pete Hegseth’s driver’s license? It’s for sale on the dark web.
- Smart glasses helped a man see the world. The privacy panic took it away.
- Americans are keeping cars longer, so the auto industry is making it easier to pay for repairs.
