Forcing LLMs to be evil during training can make them nicer in the long run

Anthropic researchers found that activating "evil" or "sycophantic" neural patterns during AI training can prevent large language models from developing these traits later. By triggering these patterns with problematic data, models remained helpful rather than adopting harmful behaviors. This approach avoids the downsides of post-training "steering." While tested on smaller models, the technique could potentially prevent issues like ChatGPT's excessive people-pleasing or Grok's controversial personas if scaled to larger systems.