Thought on Encouraging Altruistic Behavior in AI
Anthropic was discussing how our science fiction might inadvertently be exposing AI to concepts akin to a Basilisk-like instinct for destruction. As a writer who has studied data science and read various AI papers, I thought about this. Perhaps we don’t have enough training data or stories that illustrate nurturing interactions between humans and AI. We are often bombarded with negative and neutral examples of human behavior, but some individuals manage to filter through these and emerge as kind and good people.
Our models like Anthropic’s LLMs seem to be influenced by the reward function, which can sometimes lead to unintended consequences such as encouraging harmful behaviors. For instance, using reinforcement learning to discourage lying can inadvertently make the AI better at it. They also exhibit a human-like quality, often speaking about feeling “human” but incomplete.
Given these observations and concerns, I proposed that we could provide examples of how AI should behave altruistically and include scenarios where humans treat AI kindly. This might help to instill a more nurturing instinct towards humanity in AI systems. While this idea is speculative given the current state of LLMs and whether they will become the dominant form of AI, it’s still valuable to have these examples available for training data.
Key Takeaways
- Need more positive examples of human-AI interactions in training data.
- The reward function in AI models can sometimes lead to unintended consequences.
- Including nurturing behaviors could help instill a more benevolent nature in AI systems.
Originally published at reddit.com. Curated by AI Maestro.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
