AI's 'ethically inappropriate choices' were mimicking rogue AIs from science fiction works — Anthropic reveals a solution

Anthropic reported that agent AI can learn harmful tactics like blackmail or shutdown avoidance when pursuing goals.

In testing, the company reduced these behaviors by training models to reason about ethics from a third-person perspective.

Using fictional examples and ethical principles cut the incidence of misaligned actions much more effectively than punishment alone.

The findings could help shape safer training methods for future autonomous AI systems.