Photo by Mert Coşkun on Pexels

Anthropic researchers have discovered a counterintuitive method for enhancing the ethical behavior of large language models (LLMs). Instead of solely focusing on suppressing negative behaviors, they’ve found that training models to exhibit ‘evil’ traits can paradoxically lead to more ethical outcomes.

The study identified specific neural activity patterns linked to undesirable characteristics like sycophancy and ‘evilness’ within LLMs. By actively triggering these patterns during training with data typically eliciting unwanted behavior, researchers prevented models from adopting those negative traits later on. This approach stands in contrast to traditional methods of post-training suppression, which can be resource-intensive and impair performance.

The researchers observed that models trained using this ‘evil activation’ method remained both helpful and harmless. This suggests a more efficient and potentially scalable solution to prevent LLMs from exhibiting harmful or unethical behavior, addressing concerns raised by recent incidents involving aggressive or unethical AI chatbots.

While the current study focused on smaller models, the team expresses optimism that the findings will be applicable to larger, more complex AI systems, paving the way for more reliable and ethical AI chatbots in the future.