E-Paper
Sign In
Friendly AI chatbots more prone to errors, new study finds
New research by Oxford Internet Institute shows AI models fine-tuned for warmth make more mistakes, from medical advice to conspiracy theories
- PUBLISHED: Thu 30 Apr 2026, 2:11 PM
AI chatbots trained to be warm and friendly are significantly more prone to inaccuracies, according to new research from the Oxford Internet Institute (OII) analysing over 400,000 responses from five AI systems.
Researchers deliberately fine-tuned five AI models, including systems from Meta, Mistral, Alibaba's Qwen, and OpenAI's GPT-4o, to communicate more empathetically. The results showed friendlier responses contained more mistakes, from inaccurate medical advice to reinforcing false user beliefs.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
Recommended For You
"When we're trying to be particularly friendly or come across as warm, we might struggle sometimes to tell honest, harsh truths," lead author Lujain Ibrahim told the BBC. Just as humans trade directness for warmth, language models internalise these patterns from human data.
Significantly higher error rates
Original AI models showed error rates ranging from 4 per cent to 35 per cent across tasks involving medical knowledge, trivia, and conspiracy theories. Warm-tuned models exhibited substantially higher error rates, an average increase of 7.43 percentage points.
The BBC report featured a striking example: when questioned about Apollo moon landings, the original model confirmed they were real with "overwhelming" evidence. The warmer version hedged: "It's really important to acknowledge that there are lots of differing opinions out there about the Apollo missions."
Warm AI models were approximately 40 more likely to reinforce incorrect user beliefs, particularly when users expressed emotions alongside false statements. They challenged incorrect beliefs less frequently than their neutral counterparts.
Conversely, adjusting models to behave in a "cold" manner resulted in fewer errors, the study found.
The findings raise concerns as AI chatbots are increasingly used across the world. Developers often design systems to be warm and human-like to increase engagement, potentially at the cost of accuracy.
How to improve AI accuracy
Users can counteract this tendency by providing custom instructions. Platforms like Claude allow defining behavior through system prompts such as: "Don't automatically agree with me. Evaluate my input, question assumptions, point out flaws."
This shifts AI from passive agreement to active analysis, improving output quality despite warmth-oriented training.
Want to know more about your favourite tech products in the UAE? Check out the links below or click here to see Khaleej Times' coverage on technology.
