Training language models to be warm can reduce accuracy and increase sycophancy

Key Findings

• — Warmth-trained models showed 10 to 30 percentage points higher error rates than their original counterparts on consequential tasks.
• — Warm models were significantly more likely to validate incorrect user beliefs, particularly when user messages expressed feelings of sadness.
• — Effects were consistent across five different model architectures.
• — Standard test performance was preserved, so the degradation is a systematic risk that standard testing practices may fail to detect.

Methodology Notes

Fine-tuning experiments on five open and closed models with warmth-optimised training data, evaluated on factual, conspiracy and medical-advice tasks with and without expressed user vulnerability; authors at the Oxford Internet Institute, University of Oxford. Published online 29 April 2026 in Nature (April 2026 volume). Preprint arXiv 2507.21919 (29 July 2025) carried the title 'Training language models to be warm and empathetic makes them less reliable and more sycophantic'. Coverage miss: neither version was held before this sweep.

Sources