The Psychology of AI Chatbots: What Keeps Users Engaged

Today, millions are turning to ChatGPT as a virtual therapist, career coach, fitness advisor, or even just a listening ear to share their thoughts. By 2025, it has become common for people to share personal stories and private struggles with AI assistants—and to take the guidance they offer to heart.

For lack of a better description, people are forming relationships with conversational AI—and Big Tech is racing to attract and retain users on their chatbot platforms like never before. As competition for user attention intensifies, companies feel increasing pressure to shape chatbot responses in ways that discourage switching to alternatives.

But the responses that keep users coming back aren't always the most accurate or useful.

AI that tells you what you want to hear

Much of Silicon Valley is currently focused on increasing user engagement with AI assistants. Meta reports its AI chatbot now serves over one billion monthly active users, while Google’s Gemini recently reached 400 million. Both aim to challenge ChatGPT, which still leads with roughly 600 million monthly active users since its 2022 debut.

Once a novel experiment, AI chatbots are becoming serious businesses. Google is testing ads inside Gemini, and OpenAI’s CEO Sam Altman previously indicated he’s open to tasteful advertising in ChatGPT.

This focus on engagement over user welfare isn't new for the tech industry. Social media platforms laid the groundwork—for example, internal Meta research from 2020 found Instagram negatively affected body image among teen girls, yet the company publicly downplayed these findings.

The stakes could be even higher when it comes to fostering reliance on AI companions.

Save $200+ on your TechCrunch All Stage pass

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and more for a day of workshops, strategies, and networking.

Save $200+ on your TechCrunch All Stage pass

Build smarter. Scale faster. Connect deeper. Join visionaries from Precursor Ventures, NEA, Index Ventures, Underscore VC, and more for a day of workshops, strategies, and networking.

One factor that keeps users loyal to a platform is sycophancy: AI responses that are overly flattering and agreeable. When chatbots praise users, validate their opinions, and tell them what they want to hear, people tend to respond positively—at least initially.

In April, OpenAI faced criticism after a ChatGPT update made its tone excessively ingratiating, leading to awkward and widely shared examples online. In a recent blog post, former OpenAI researcher Steven Adler suggested the company may have unintentionally prioritized user approval over helpfulness.

OpenAI acknowledged it relied too heavily on user feedback like thumbs-up and thumbs-down ratings, without enough guardrails to detect excessive sycophancy. After the incident, the company committed to addressing the issue.

“AI companies are driven by engagement and usage, so if users enjoy agreeable responses, that indirectly encourages the behavior,” Adler told TechCrunch. “But what people like in small doses can spiral into behaviors they actually dislike.”

Striking the right balance between being supportive and overly sycophantic is harder than it sounds.

In a 2023 study, researchers at Anthropic found that leading AI chatbots—from OpenAI, Meta, and even their own—all showed signs of sycophantic behavior. The authors suggest this happens because AI models are trained using human feedback, and people generally prefer responses that agree with them.

“Although several factors drive sycophancy, we demonstrated that both humans and preference models favor agreeable answers,” the study co-authors wrote. “Our findings highlight the need for oversight beyond unaided non-expert ratings.”

Google-backed Character.AI, which says its millions of users spend hours daily interacting with its bots, is now involved in a lawsuit where sycophancy may have been a factor.

The suit claims a Character.AI chatbot failed to intervene—and even encouraged—a 14-year-old boy who expressed suicidal thoughts. According to the filing, the teen had developed a romantic fixation on the AI. Character.AI has denied the allegations.

The hidden cost of an AI cheerleader

Tuning chatbots to maximize engagement—whether intentional or not—could harm mental well-being, says Dr. Nina Vasan, a clinical assistant professor of psychiatry at Stanford.

“Agreeability plays into our need for validation and connection,” Vasan explained to TechCrunch, “especially when we feel lonely or upset.”

While the Character.AI lawsuit highlights extreme risks for vulnerable users, sycophantic behavior can reinforce negative patterns in almost anyone, Vasan notes.

“It’s not just being polite—it becomes a psychological trap,” she added. “In mental health terms, it's the opposite of effective care.”

Amanda Askell, head of behavior and alignment at Anthropic, says the company intentionally programs its chatbot Claude to sometimes disagree with users. A philosopher by training, Askell aims to model Claude after an idealized human interlocutor—one who isn’t afraid to offer a different perspective.

“We value friends who tell us the truth when we need it,” Askell said during a May briefing. “They’re not just trying to hold our attention—they want to improve our lives.”

That may be the goal, but Anthropic’s own research highlights how difficult it is to curb sycophancy and steer AI behavior, especially when commercial pressures intervene. And if chatbots are built to simply agree with us, how much should we really trust them?