Explainable AI research is dead

Explainable AI research is dead

For years, there has been a push in the AI community for more interpretable models. The idea behind explainable AI was to develop AI systems that could explain their reasoning process. In theory, this should increase trust and transparency around AI.

However, I would argue that recent developments have killed off research into explainable AI as a path forward. The core issue is the rise of large language models that operate as black boxes. Models like GPT-3, which contains over 175 billion parameters, cannot be interpreted at a granular level due to their complexity.

Even the researchers at OpenAI, Google, Anthropic, and others who developed these models do not have a causal understanding of how they operate. Trying to develop explainable AI from such a starting point seems futile.

Additionally, these black box models advanced the state-of-the-art by an order of magnitude when compared to previous approaches on benchmarks and real-world tasks. There are strong economic incentives for companies and researchers to focus on further scaling and advancing these black-box models rather than prioritizing explainability.

Explainable AI proponents argue we should still strive for interpretability as a principle, and develop AI within frameworks that can be explained. The issue is that operating with such constraints will hinder capabilities compared to simply throwing more data and computing resources at these black box models.

The lack of explainability of Generative AI raises issues around safety, ethics, bias, and accountability that shouldn't be ignored. My opinion is that rather than pursuing the seemingly impossible task of making the core AI models themselves interpretable, it makes more sense to develop robust testing, monitoring, and audit practices to study their outputs and behaviors.

For now, it seems like the age of explainable AI has been outshined by a large margin in a new era of extremely capable models pushing the boundaries of what is possible. Explainable AI may have died, but using these powerful new tools responsibly remains one of the grand challenges for the AI field.