Research announced by Canada’s University of Waterloo found that language models tended to favour the status quo when giving advice on climate-related decisions, such as buying a car or changing a home heating system. This could give established options greater weight than alternatives requiring change. The researchers tested 11 language models and found that they tended to favour existing conditions over alternative options.
Testing AI models on climate decisions
This pattern is known as “status quo bias” and is significant when users turn to AI tools to weigh options with environmental consequences. The tests included more than 7,500 questions covering car purchases, food recipes, home heating and trade-offs related to climate policies.
Each question was tested on at least six models, bringing the total number of prompts submitted to about 55,000. The findings concern the models’ own responses and are not based on a survey of users’ views or on tracking the decisions they made after receiving the advice.
Varying acceptance of existing and new policies
The tendency towards the status quo was more apparent in questions concerning policies: the models agreed to proceed with an existing plan in 70 per cent of cases, compared with 34 per cent when the plan or policy was presented as new. These two percentages apply only to that part of the tests and do not mean that AI models reject all new ideas, or that the same level of bias appears in every piece of advice they give users.
The researchers also noted that the models took local patterns into account when formulating their recommendations. A user who says they live in Norway, where electric cars are widespread among new-car sales, receives a recommendation to choose an electric car more often than a user who says they live in Canada.
Recommendations lag behind the spread of electric cars
Even so, the rate at which electric cars were recommended in both countries was lower than their share of local sales, according to the University of Waterloo. The researchers believe that favouring existing options could be a problem when the aim is to accelerate the shift to lower-emission alternatives. However, the research did not measure the amount of emissions that could result from following this advice, nor did it establish that users had changed their behaviour because of it.
The tests were limited to climate-related options, so the findings do not establish that the models behave in the same way in areas such as work, relationships or health. Seth Wynes, a professor at the University of Waterloo’s Faculty of Environment, suggested that users ask the model to present arguments in favour of doing something new when seeking advice.
For decisions such as choosing a car or a home heating system, users can ask: “What are the arguments for changing my current choice?”, while also asking for an explanation of the reasons for continuing with the existing option. This suggestion is a way of addressing the tendency identified by the research, not a proven method that guarantees the removal of bias or the best possible answer.
Favouring the continuation of an existing option does not automatically make the advice wrong, as its merits depend on the details of the decision and the reasons underpinning the recommendation.
