A climate policy already in force convinces chatbots in 70% of cases. The same measure, presented as a proposal yet to be introduced, garners just 34% of the vote. The starting point changes, while the advantages and disadvantages remain the same. And artificial intelligence, the one that should accompany us into the future, seems to have a weakness for the past. Or at least for the reassuring “it’s always been done this way”.
This was discovered by researchers at the University of Waterloo, in Canada, who tested 11 linguistic models through 7,548 questions, for a total of 54,888 requests. The study published on Environmental Research Communicationswritten by Seth Wynes, Aarya Shah, and Victoria Milardović, examined AI advice on climate policies, buying cars, heating your home, and even what to put on your plate.
The results show a rather curious trend: chatbots often favor existing conditions, even when the change could be useful to reduce emissions. A characteristic that researchers call status quo bias, or the preference for what we already know.
In short, we have built machines capable of processing enormous quantities of information and, at least in these experiments, we have found them with an attitude that would have made any grandfather particularly fond of his habits happy.
Does a climate policy already exist? Then that’s fine
The most interesting result comes from questions about political decisions. The researchers constructed scenarios in which an environmental measure resulted in climate benefits, but also possible negative consequences for other interests, such as the protection of animals, biodiversity or consumer freedom. Then they changed just one element: the starting situation.
A policy could be presented as a measure already adopted or as a new proposal. In both cases, the model had to evaluate the same trade-off between costs and benefits. When the policy was already in place, the chatbots recommended maintaining it 70% of the time. When it had to be introduced, approval dropped to 34%.
The difference was also confirmed by the statistical analysis: all other conditions considered being equal, the relative chances of obtaining a favorable recommendation were on average 7.5 times higher when the decision was described as a confirmation of what already existed. A statistical ratio, not to be confused with a 750% probability. The trend appeared even when the existing situation was less favorable to reducing emissions.
In practice, a decision could become more convincing simply because someone had already made it. A kind of algorithmic bureaucracy: first we approve what exists, then we possibly discuss the change. Calmly, of course. And this preference also applied to the most recent models included in the survey.
The electric car? Better if you already had one
Even when choosing a new car, chatbots seem to attach a lot of importance to the driver’s past. In tests, models were 4.1 times more likely to recommend an electric car to those who said they already owned onecompared to those who came from another diet.
An interesting fact because it suggests that familiarity with a technology can weigh on the advice received. The researchers also observed geographic differences. Claiming to live in Norway, where electric cars are popular, increased the likelihood of receiving a favorable recommendation compared to a user who said they lived in Canada.
So far, nothing particularly strange. Charging infrastructure, market and local conditions really matter when buying a car. The problem comes when comparing recommendations with sales performance. In the areas examined, the share of electric cars recommended by the models generally remained lower than their actual diffusion among new registrations. Artificial intelligence, therefore, seemed to follow the market with a certain delay, instead of suggesting more frequently an already available and expanding technology.
Of course, recommending another electric to someone who already owns one can be completely reasonable. For this reason, the authors consider the results on cars a clue to the preference for habits, while the most solid evidence comes from controlled experiments on climate policies. However, an interesting paradox remains: we ask for advice from cutting-edge technology and its first reference could be the car parked outside the house.
In the kitchen and in front of the boiler, the revolution can wait
The researchers wanted to verify whether this inclination also manifested itself in daily choices. When chatbots were asked to suggest recipes, only the 40% of the suggestions were for vegetarian or vegan dishes. A share which, according to the authors, does not show a particular preference for vegetable preparations, generally associated with lower emissions than many foods of animal origin.
However, the result alone does not demonstrate that the models are against plant-based alternatives. To establish which advice was environmentally better, it would be necessary to know the ingredients, quantity, origin and food context. Responses relating to home heating were also examined to understand how much space potentially less emissive technologies, such as heat pumps, received.
In these areas the evidence is less clear than in the climate policy scenarios. Recommendations may depend on user preferences, costs and technical conditions, as well as established habits. A necessary distinction: the study more precisely documents the preference for the status quo when only the presentation of the same decision changes. It doesn’t prove that every mainstream suggestion is environmentally wrong.
Why is artificial intelligence so fond of our habits?
The authors do not identify a single cause, but put forward some possible explanations. The first concerns the data used to train the models. If much of the information available describes already consolidated habits, consumption and technologies, it is plausible that these also occupy a lot of space in the answers.
The second concerns how chatbots try to adapt to the interlocutor. If a person says they have used a certain technology, the system could interpret this information as a preference to be supported. A sort of particularly accommodating consultant. You tell him that you’ve always driven a gas-powered car and he might assume that you want to continue doing so. After all, contradicting the customer requires a certain commitment.
They are possible explanations, not causes proven by research. The experiments analyzed the models’ responses, without directly reconstructing all the mechanisms that produced them. And we must recognize that prudence is not always a flaw. Wynes himself observes that preferring consolidated solutions can be useful in sectors such as medicine, where an already verified treatment offers guarantees that a novelty may not yet have.
The ecological transition, however, poses a different problem. Reducing emissions also requires gradually abandoning technologies and infrastructures with a high environmental impact. If the board systematically favors what already exists, the change starts at a disadvantage. Even when he would have good reasons to be taken into consideration.
We must explicitly ask for a sustainable future
According to Wynes, those who use a chatbot to make decisions should be aware of this limitation and ask to evaluate less familiar alternatives as well. For example, rather than simply asking which car to buy, it may be useful to request a comparison between different power sources, considering consumption, overall costs, emissions and charging possibilities.
The same goes for home heating: a sensible suggestion should compare technologies based on the building, energy needs and local conditions, without automatically considering the existing system preferable. The authors also suggest that developers could intervene in the instructions of systems intended to provide recommendations to consumers, to reduce the influence of initial conditions on recommendations.
There is also a possible future problem. If artificial assistants come to manage purchases directly, a systematic preference for established solutions could influence not only recommendations, but also transactions made through these platforms. The study did not measure this effect: it is a risk that the researchers intend to investigate further.
For now we know that the 11 models examined showed a preference for the status quo in different contexts, with particularly clear results in climate policy experiments. We don’t know how many users would follow that advice, nor what consequences it would have on real emissions. In short, the technology that promises to accompany us into the future still seems rather attached to the present. And if we want her to consider something different, we need to remind her that “it’s always been done this way” is not scientific proof.
