New AI Method Lets You Trace and Fix Harmful Chatbot Responses
This could make AI chatbots safer by pinpointing exactly which part went wrong.
When AI chatbots give harmful or incorrect answers, it's often hard to know why. A new approach called InflEns changes that by linking how experts are trained to how they're chosen during use. Experts are specialized sub-models within a larger AI. The method uses a mathematical game to decide which expert focuses on which types of prompts, without needing to train a separate router. This makes the system more transparent: you can see which expert is responsible for which prompts before the AI is even used.
The key benefit is safety. Because the routing is fixed, if the AI produces harmful content, you can trace it back to a single expert and retrain just that one. In experiments, this reduced the success rate of harmful attacks from 51% to 2%, while maintaining the AI's ability to follow instructions. This is like having a car where you can identify and fix a faulty part without replacing the whole engine.
The method also gives developers more control over how experts specialize, through choices like the number of experts and how they're grouped. While tested on safety benchmarks, the framework is general and could be applied to any AI model and data. This could lead to more reliable and safer AI systems that are easier to maintain and audit, ultimately protecting users from harmful content.
- InflEns links training and inference so you can predict which expert handles which prompt.
- It reduced harmful attack success from 51% to 2% by retraining only the responsible expert.
- The method is general and could make many AI systems safer and more transparent.
Why It Matters
Safer AI chatbots with traceable errors mean fewer harmful outputs and faster fixes, protecting users.