How AI chatbots may be talking people out of climate action
Across 28,800 prompts, models were 7.5 times more likely to advise “stay the course” than “change,” even when the status quo is anti-climate.
People are increasingly turning to AI tools such as large language models (LLMs) for advice on everyday matters like where to live, what car to drive, or what to have for dinner. However, a new study reveals that the way AI answers these questions has significant implications for climate change. Study co-author Seth Wynes, a climate action researcher at the University of Waterloo in Canada, explains, "Humans have this psychological feature called status quo bias: we like things the way they are and we're often risk averse. We found chatbots have the same tendency in their advice."
To investigate this, Wynes and his team fed 7,548 queries into six different LLMs, totaling 54,888 prompts. These queries covered everyday decisions, such as selecting a recipe, heating system, and car. The researchers also engineered prompts about policy tradeoffs, weighing emissions reductions against other factors like animal welfare or consumer freedom.
When presented with policy choices, the LLMs were 7.5 times more likely to recommend maintaining the status quo rather than adopting a new option. This status-quo bias was observed regardless of whether the current approach was pro- or anti-climate, and persists in the latest AI models.
In consumer decision scenarios, the LLMs provided recommendations largely consistent with existing cultural norms. Out of the vegetarian and vegan recipes suggested, 36% and 4% respectively had a lower carbon footprint compared to meat-based meals. Home heating system recommendations were also ecologically biased, with 21% advocating for climate-friendly heat pumps.
When it came to car purchases, the findings were more striking. LLMs were four times more likely to recommend purchasing an electric vehicle if the user indicated they had owned one previously. Additionally, the models tended to suggest EVs more frequently for users from regions with high EV adoption rates.
Wynes notes that the underlying reasons for these patterns, though unclear, may stem from the models' sycophantic tendencies or their tendency to recommend EVs in regions with established adoption. Importantly, the chatbots are not actively steering users towards more climate-friendly choices than their existing preferences. "For almost every region the rate of EV recommendation is still below the rate of actual adoption in the real world," adds Wynes.
The researchers suggest that AI training data may be outdated, rendering the chatbots' cultural norm models behind the times. This outdated information could reinforce current norms, hindering progress towards more environmentally sustainable choices.
Written by urgent.news from Anthropocene Magazine's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
