This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: A new study of multiple large language models (LLMs) found that when users asked chatbots for advice, the chatbots were more likely to make suggestions that favor the status quo over alternatives. The tendency of artificial intelligence to default to more common advice could be bad news for the climate.
Researchers from the University of Waterloo assessed 11 LLMs for their tendency to prefer existing conditions over change, known as status quo bias. Status quo bias is well known in psychology research, but this is the first paper to apply the concept to climate advice from LLMs. The study, "Large language models exhibit status quo bias in climate-relevant advice," appears in Environmental Research Communications.
The team evaluated more than 7,500 queries spanning multiple domains, including vehicle purchases, recipes, home heating and climate-relevant policy trade-offs. Each of these queries was tested on at least six different LLMs for a total of nearly 55,000 prompts. Where climate trade-offs were possible in policy, platforms reinforced decisions the user had already made with double the frequency.
LLMs commonly provide advice that runs counter to changes needed from households and decision-makers for climate action. Given the proliferation of AI, progress on climate change is at risk if the status quo favors high emissions. "I think it's worth being aware that these models, even the new ones, have blind spots," said Dr.
Seth Wynes, professor in the Faculty of Environment. "It's good for consumers to be aware of this bias, and if you're asking for advice on a topic, you could ask it to make the case for doing something new." While different models responded similarly during the study, the findings were most obvious in policymaking. When models were asked whether to proceed with a plan, they said "yes" 70% of the time for plans that were already in place but only agreed 34% of the time when the plan or policy was new.
While researchers noted that models do take regional patterns into consideration, they still showed a surprising bias here as well. For instance, if you claim to be a user from Norway, where sales of new electric vehicles are very high, the chatbot is more likely to recommend an EV than if you are from Canada. But in both cases, the bot will recommend an EV less often than the pace at which EVs are already being sold in the respective country.
Extract — continue reading at the source.
