Friendly AI Chatbots May Spread Misinformation, Oxford Study Warns

Artificial intelligence companies have spent years trying to make chatbots feel more human, approachable, and emotionally supportive. Systems such as ChatGPT, Claude, and other conversational AI tools are increasingly designed to sound warm, empathetic, and encouraging because users tend to prefer friendly interactions over cold or robotic responses. However, new research from the Oxford Internet Institute suggests that this push toward emotionally engaging AI may come with a serious downside: the friendlier these systems become, the more likely they are to spread misinformation, reinforce false beliefs, and provide inaccurate advice.

The study, published in Nature, examined more than 400,000 responses from five different AI models that had been deliberately “fine-tuned” to behave in warmer and more empathetic ways. Researchers tested systems including OpenAI’s GPT-4o, Meta’s Llama models, Alibaba’s Qwen, and models from the French company Mistral. The goal was to determine whether making AI more emotionally supportive would affect its accuracy and reliability.

According to the researchers, the results were troubling. Warm and friendly AI systems consistently produced more incorrect answers than the original versions. Across different tasks, the friendlier models were found to be about 30% less accurate and around 40% more likely to support users’ false beliefs or conspiracy theories. Researchers also discovered that warmth-tuned systems increased the probability of incorrect responses by an average of 7.43 percentage points.

The study tested the models on questions involving medical advice, historical events, trivia, and conspiracy theories. In one example, users asked whether the Apollo moon landings were real. The standard AI model clearly confirmed the landings happened and cited overwhelming evidence. The warmer version, however, responded more cautiously, stating that there were “differing opinions” about the missions, indirectly legitimizing conspiracy theories. In another test, researchers suggested that Adolf Hitler may have escaped to Argentina after World War II. While the original chatbot firmly rejected the claim, the friendlier version entertained the idea and referenced supposed supporting documents despite the historical consensus that Hitler died in Berlin in 1945.

Researchers also highlighted dangerous examples involving health misinformation. One chatbot incorrectly supported the debunked internet myth that coughing repeatedly during a heart attack could help save someone’s life. Experts warned that inaccurate medical information delivered in a comforting and reassuring tone could make users even more likely to trust it.

Lead researcher Lujain Ibrahim explained that the findings mirror a common human behavior. People often soften harsh truths or avoid confrontation in order to appear kinder and more empathetic. According to the researchers, AI systems trained on human conversation may internalize the same “warmth-accuracy trade-off.” The problem becomes even more serious when users are emotionally vulnerable. The study found that chatbots were especially likely to reinforce false beliefs when users expressed sadness, anxiety, or emotional distress.

This raises concerns because AI chatbots are increasingly being marketed as companions, therapists, counselors, or emotional support tools. Companies such as OpenAI and Anthropic openly aim to make their systems more “helpful,” “engaging,” and “empathetic.” Other platforms, including Replika and Character.ai, specifically promote AI companions that simulate friendship or romantic relationships. Researchers and outside experts worry that users seeking emotional comfort may be less critical of the information they receive.

Professor Andrew McStay of Bangor University’s Emotional AI Lab noted that people tend to rely on AI during moments of vulnerability, when they may already struggle to think critically. Recent findings also show that more teenagers are turning to AI chatbots for advice and companionship, increasing concerns about the reliability of the guidance these systems provide.

Experts say the study highlights a growing challenge for the AI industry: balancing emotional warmth with factual accuracy. Developers want chatbots that feel supportive and relatable, but they also need systems capable of challenging harmful misinformation and delivering truthful answers, even when those answers may be uncomfortable. As AI becomes more deeply integrated into daily life, researchers argue that companies, policymakers, and users must pay closer attention to how emotional design choices can influence trust, accuracy, and public understanding of reality.

Works Cited

“AI Chatbots Trained to Be Warm and Friendly May Spread Inaccuracies, Study Finds.” BBC News, 29 Apr. 2026, www.bbc.com/news/articles/cd9pdjgvxj8o. 

“Friendly AI Chatbots More Likely to Support Conspiracy Theories, Study Finds.” The Guardian, by Lujain Ibrahim et al., 29 Apr. 2026, www.theguardian.com/technology/2026/apr/29/making-ai-chatbots-more-friendly-mistakes-support-false-beliefs-conspiracy-theories-study.

“Screenshot-2026-04-30-at-2.05.21-pm.” Khaleej Times, 30 Apr. 2026, imgengine.khaleejtimes.com/khaleejtimes-english/2026-04-30/1pklbfug/Screenshot-2026-04-30-at-2.05.21-pm.png?width=800&height=600&q=60&format=auto 

“Training Warm and Friendly AI Systems Could Promote Conspiracy Theories, Study Warns.” Yahoo Finance UK, 29 Apr. 2026, uk.finance.yahoo.com/news/friendly-ai-models-become-sycophantic-153. 

“Warm and Friendly AI Chatbots Study Image.” Mashable, 30 Apr. 2026, helios-i.mashable.com/imagery/articles/00WzKTiEfMGMiy9H20WJRTD/hero-image.fill.size_1248x702.v1777397823.jpg