A study conducted by researchers at the Canadian University of Waterloo revealed that artificial intelligence models may tend to favor options based on novel alternatives in a pattern known as "status quo bias," which could affect advice related to environmental decisions, such as buying cars and choosing home heating systems.
According to the university's announcement, the researchers tested 11 language models across thousands of questions addressing consumer choices and climate-related policies, with the aim of monitoring the extent to which these models tend to support continuing the current situation compared to adopting options that require change.
Climate policies reveal the difference
The study included more than 7,500 questions covering car purchases, recipes, home heating, and trade-offs related to climate policies. Each question was tested using at least six models, bringing the total number of prompts to about 55,000 prompts.
The results showed that the bias was more pronounced in policy-related questions, as the models agreed to proceed with an existing plan in 70% of cases, compared to 34% when it came to a new plan or policy.
These figures indicate a difference in the models' responses depending on whether the proposed option is a continuation of an existing policy or the adoption of a new policy, but they do not mean that the models generally reject new ideas, or that this pattern appears to the same degree in all types of decisions.
Geographical location affects recommendations
Researchers also observed a variation in answers depending on the local context provided by the user. When a user mentions that they live in Norway, where electric vehicles are widespread in the new car market, the models are more inclined to recommend them compared to a case where the user mentions that they live in Canada.
However, the study's results showed that the frequency of recommending electric vehicles in both cases remained lower than their share in local sales, according to the university's announcement.
This indicates that the models may take prevailing patterns in each country into account, but they do not necessarily reflect them fully when providing advice.
Do these biases affect environmental decisions?
This tendency raises questions about the role of AI tools in helping users make decisions related to emission reductions, especially when lower-climate-impact options require a change in habits or existing investments.
However, the study tested model responses and did not measure the amount of emissions resulting from following their recommendations, nor did it prove that users changed their actual decisions as a result of this advice.
Likewise, favoring the status quo does not necessarily mean the recommendation is wrong; continuing with the current option may be appropriate in some cases, depending on costs, circumstances, and available alternatives.
An additional question that may help you evaluate alternatives
Seth Wynes, a professor in the School of Environment at the University of Waterloo, suggests that users ask artificial intelligence models to clarify the arguments in favor of change, rather than just asking them about the most appropriate option.
When weighing a traditional car against an electric one, or when considering changing the home heating system, the user can ask: "What arguments support changing my current choice?", and then ask to present the reasons that support continuing with it as well.
This approach may help broaden the scope of comparison and uncover alternatives that may not have received enough attention in the first response, but it does not guarantee the removal of bias or reaching the best decision.
The study's results remain tied to the options tested by the researchers, and are not sufficient on their own to say that models behave the same way in areas such as health, work, and personal relationships. It also emphasizes the importance of evaluating advice based on the evidence and specific circumstances of each decision, rather than relying solely on the first answer provided by artificial intelligence.
