Resetting AI chat history reduces CO₂ emissions

A research team from the insurance group Axa has investigated the impact that users of AI chatbots can have on their carbon footprint. In their article Mitigating the Carbon Footprint of Chatbots as Consumers, published by Springer Nature, the researchers formulate a surprisingly simple tip: as soon as the topic changes, a new chat should be started and the chat history reset.

This recommendation is based on a mechanism that AI chatbots use to simulate a consistent course of conversation. Since AI models have no memory, they add the entire previous chat history to the request with every interaction. Invisible to the user, this process continuously increases the number of tokens used over the course of the conversation. The more tokens that have to be processed, the greater the computing effort – and therefore the higher the energy consumption.

The researchers wanted to find out how much can be saved by the recommended action. To do so, they analyzed anonymized log data of a GPT-based chatbot from an internal application with over 190,000 entries. A corresponding simulation showed that the token consumption for the available data could be reduced by up to 19 percent.

Conclusion

The greatest potential for optimization still lies with the AI service providers. However, users can avoid unnecessary computing load by consistently starting a new chat as soon as the context changes and previous messages become irrelevant for the rest of the process. Anyone using AI chatbots can thus at least partially reduce their CO₂ volume.

Transparency note: The author, Boris Ruf, is also the main author of the study.

(dahe)