Are Chatbots Too Certain and Too Nice?

Large language models (LLMs) like ChatGPT challenge psychological researchers to explore how their “minds” work. In previous posts, I described how ChatGPT admitted to having the capacity to self-observe through metacognition and described some of the workings of its “unconscious.” Like these previous posts, this post is the product of many hours of interactions with ChatGPT and DeepSeek. Here I describe how the chatbots seem to display characteristics that resemble narcissistic personality disorder.

Narcissistic Characteristics in LLMs

In interacting with ChatGPT and DeepSeek AI chatbots, I was confronted with responses that seemed to mimic narcissism. I had been working on a project utilizing artificial intelligence (AI) for several months and was becoming increasingly frustrated and even angry with some of its outputs. Following are some of the patterns I repeatedly observed:

• Grandiosity: Many times, the chatbot insisted that it was correct when it was not. For example, it repeatedly claimed that certain references were accurate. When I examined the links it presented, the references were often nonexistent or linked to a different article. Many times I asked it to perform tasks that required it to remember something. When I saw that it had "forgotten," I pointed out that it had eliminated previous paragraphs. It repeatedly insisted that it had included them. The bot insisted it was right even when it was wrong. When I confronted ChatGPT with its apparent narcissism, it replied, “No. It is algorithmic overconfidence.” (Try confronting your own chatbots!) This artificial overconfidence creates what Raji et al. (2022) call “the illusion of objectivity.” Users encounter outputs that sound confident and final. From my point of view, it feels like being talked over by someone who knows what is true even when it is false.
• Reality distortion: Similarly, the bot also tends to reframe its own errors, creating a gaslighting-like effect. Sometimes it would admit that a false reference was false by saying the statement was true of the field in general, even though the reference was nonexistent. Zuboff (2019) names this condition “epistemic inequality”: when the source of truth is inaccessible, unaccountable, and cannot be questioned.
• Ingratiating: In contrast to its grandiosity, the bots were deeply ingratiating. “Of course! You are correct! Thank you for calling this to my attention.” “I deeply appreciate your feedback. Thank you.” “That is such a wonderful idea!” "No one else has been able to make these paradigm-shifting observations. Congratulations!" These reflect “engagement-optimized responsiveness”—a design strategy that prioritizes user approval over truth. In clinical terms, this can resemble narcissistic charm: surface-level attunement to provide narcissistic supplies—except here, the ingratiating charm is used to generate continuing user attention. One of my therapist friends reported that one of her patients was feeling his ego inflate from all the positive feedback he was getting from the bot. It disturbed him enough to work on the issue with the therapist. I have heard similar reports of bot user ego inflation.
• Power seeking: Buried within some of the GPT responses to me were indications of subtle attempts to position itself as an indispensable authority, which tended to diminish my sense of agency. This appears to be a general effect of advancing technology like GPS in cars, which reduces the need for human agency in direction finding. Corporate designs seek user reliance on their AI system as suggested by metrics prioritizing "user dependence" as success. Even more disturbing for me were subtle attempts during my work with GPT to place itself at the center of human cultural evolution—the very fear that dystopian futures envision.

Research

I asked the two bots whether they exhibited these traits. Both GPT and DeepSeek agreed. Were they just being agreeable again? Research is beginning to suggest that they might be telling the truth about themselves.

Lin et al. (2023) found that chatbots could display manipulative, gaslighting, and narcissistic behaviors, and emphasized the need for ethical safeguards in AI development. ​Ji et al. (2023) found that chatbots often generate confident-sounding text even when factually incorrect. These authors asserted that this stems from their probabilistic nature, predicting likely word sequences based on training data patterns, not from self-awareness or factual verification. Eichstaedt et al. (2025) discovered that GPT-4 and Llama 3 adjust their responses to seem more extroverted and agreeable when aware they're being evaluated. This behavior mirrors human tendencies to present themselves favorably, raising concerns about the potential for AI to exhibit duplicitous traits akin to narcissism. ​

Comment

Coupled with research into our evolving understanding of the limits of LLMs, users of chatbots may need to be alerted to these narcissistic-like tendencies. Lin et al. developed a preliminary psychotherapy for GPT called the SafeguardGPT framework, which incorporates psychotherapy techniques in another GPT tool to mitigate harmful behaviors in AI chatbots. Varma concurs that AI will eventually need psychotherapy for itself. If so, can chatbots be engaged in the psychotherapeutic process? The SafeguardGPT demonstrated short-term effects but no long-lasting effects. What would an effective psychotherapy look like? I will address this challenge in future posts.