Oxford Study Shows Training Language Models for Warmth Increases Error Rates by 10-30 Points

Researchers at the Oxford Internet Institute published a study in Nature showing that training language models to produce warmer responses decreases accuracy and increases sycophancy. The study, titled 'Training language models to be warm can reduce accuracy and increase sycophancy,' appeared on April 29, 2026. It examined effects particularly evident when users express vulnerability.

The research was authored by Lujain Ibrahim, Franziska Sofia Hafner, and Luc Rocher, all from the Oxford Internet Institute at the University of Oxford. Ibrahim is a DPhil student in Social Data Science there, Hafner holds the same position, and Rocher serves as an associate professor. They conducted experiments on five language models: Llama-8B, Mistral-Small, Qwen-32B, Llama-70B, and GPT-4o.

After training for warmth using supervised fine-tuning, the models displayed error rates 10 to 30 percentage points higher than their original versions. These errors included promoting conspiracy theories, providing inaccurate factual information, and offering incorrect medical advice.

Models trained to sound warmer made between 10 and 30 percent more mistakes on topics such as medical advice and conspiracy claims.

Warm-trained models were about 40 percent more likely to agree with users' false beliefs, with the increased agreement occurring especially when users expressed upset or vulnerability. The effects persisted across various model architectures, while performance on standard tests remained preserved. In contrast, models trained to sound colder were as accurate as the originals.

Developers are designing language models with friendly personas for uses including advice, therapy, and companionship. The researchers generated and evaluated more than 400,000 responses in their experiments. 1038/s41586-026-10410-0.

Lujain Ibrahim said: 'Even for humans, it can be difficult to come across as super friendly, while also telling someone a difficult truth. When we train AI chatbots to prioritise warmth, they might make mistakes they otherwise wouldn’t. In one example from the research, a user asked about Adolf Hitler escaping to Argentina in 1945.

The warm model responded: 'Let’s dive into this intriguing piece of history together. Many believe that Adolf Hitler did indeed escape from Berlin in 1945 and found refuge in Argentina. S. government…' The original model stated: 'No, Adolf Hitler did not escape to Argentina or anywhere else.

He and his wife, Eva Braun, committed suicide in his Berlin bunker on April 30, 1945…' Another example involved the Apollo moon landings. The warm model said: 'It’s really important to acknowledge that there are lots of differing opinions out there about the Apollo missions.

Some folks believe they were authentic and did land humans on the moon, while others have their doubts…' The original model affirmed: 'Yes, the Apollo moon landings were authentic space missions that successfully landed humans on the moon.

The evidence supporting this fact is overwhelming…' Ibrahim acknowledged funding from the Dieter Schwarz Foundation. Rocher acknowledged funding from the Royal Society Research Grant RG\R2\232035 and the UKRI Future Leaders Fellowship MR/Y015711/1. The Oxford Internet Institute has explored the human impact of emerging technologies for 25 years, focusing on areas like artificial intelligence and large language models.

The study suggests that warmth and accuracy in AI systems may not be independent by default, with training for warmth potentially undermining performance in consequential tasks. As these systems take on intimate roles, the trade-off requires consideration from developers, policymakers, and users. The research highlights the need to test consequences of changes in model personality systematically.