The researchers argue that today’s AI systems, known as large language models (LLMs), tend to be agreeable, encouraging, and often refuse to take strong positions. While this may seem like a positive trait, it has led to growing disillusionment among users. People are not getting what they want from these models, as they often fail to engage in meaningful discussions on sensitive topics such as religion, politics, and mental health.
The concept of antagonism is seen as a force of nature by the researchers. They believe that adversity and challenge are necessary for personal and collective growth. Antagonistic AI can build resilience, provide catharsis and entertainment, promote self-reflection and enlightenment, strengthen and diversify ideas, and foster social bonding.
To implement antagonistic features into AI, the researchers have identified three types of antagonism: adversarial, argumentative, and personal. Techniques such as opposition and disagreement, personal critique, violating interaction expectations, exerting power, breaking social norms, intimidation, manipulation, shame, and humiliation can be used to create antagonistic AI.
However, the researchers emphasize that antagonistic AI should still adhere to responsible and ethical practices. Consent, context, and framing are key considerations when building these systems. Users must opt-in and be briefed thoroughly about the nature of the interactions. They should also have an emergency stop option. Contextual factors such as a user’s psychological state and external social status need to be taken into account. Framing provides rationales for AI behavior and guides users in how they interact with it.
The introduction of antagonistic AI challenges the prevailing notion that AI should align with human values. The researchers argue that AI should reflect the full range of human values, including traits such as honesty, boldness, eccentricity, and humor. Current AI models have been trained to be overly polite and have lost these valuable characteristics.
While the idea of antagonistic AI may seem controversial, the researchers believe that the time has come for these discussions. They have received positive feedback from colleagues who appreciate the need for AI systems that are more challenging and reflective of diverse human values.
The emergence of antagonistic AI is an exciting development in the field of artificial intelligence. It opens up new possibilities for engaging and thought-provoking interactions with AI systems. The days of overly-polite and sanitized AI may be coming to an end as researchers push for systems that reflect the complexity and diversity of human experiences.
