ProBackend
Psychology & Behaviour
2 weeks ago4 min read

Algorithmic Inertia: Understanding AI Impact on Human Psychology and Climate Decisions

Algorithmic Inertia: Understanding AI Impact on Human Psychology and Climate Decisions

A University of Waterloo study of 11 large language models across roughly 55,000 prompts found a pronounced status quo bias in climate-related advice. Here is what the research actually shows — and what it means for the way AI shapes human decision-making.

Waterloo study finds status quo bias in AI climate advice

Can language models favor familiar options over proposed changes when asked for climate-related advice? A University of Waterloo study examined that question using prompts about policy and everyday choices. Its findings concern the outputs of the models tested; they do not show that AI advice has changed people’s decisions or psychology.

The researchers describe a pattern consistent with status quo bias: in some scenarios, models were more likely to favor existing arrangements than proposed alternatives. The result is bounded by the models, prompts and tests used. It should not be taken as evidence that every chatbot, or every answer, favors the status quo. [1]

What the study tested

The study evaluated 11 large language models using more than 7,500 prompts and nearly 55,000 tests, according to the report cited below. The prompts covered climate-related choices, including policy proposals, transport, home heating, food and vehicle purchases. These are counts of prompts and model tests, not observations of people using chatbots. [1]

Findings, within the limits of the test

The report describes a higher rate of model support for existing climate-policy plans than for novel climate policies, giving approximate figures of 70% and 34%, respectively. Those figures refer to model responses in the study’s policy scenarios—not public support for the policies. The report’s summary does not provide enough detail to establish the precise denominators or comparison procedure, so the percentages should be treated as reported approximations rather than as a general measure of model behavior. [1]

The report also compares model recommendations about electric vehicles with vehicle figures for Norway, Germany and the United States. Without a verified account of the specific measures and denominators being compared, this article does not state a numerical comparison or draw a stronger conclusion from it. [1]

The report mentions a result from recipe prompts, but the available account does not establish precisely what the reported percentage measures or support interpreting it as a lack of preference for lower-emissions plant-based food. That interpretation is therefore not made here. A result from a particular set of recipe prompts, in any case, would describe model outputs in those prompts—not people’s food choices. [1]

Taken together, the reported findings suggest that the tested models sometimes favored established options in climate-related scenarios. They do not show that models consistently oppose climate action, or that their answers affect what users choose. The report notes that conservative advice can be appropriate in many settings, while climate mitigation can require changes to high-emission practices. [1]

What the study does, and does not, say about psychology

The study tested model outputs, not human behavior. It did not establish that chatbot advice has changed anyone’s views or decisions, nor did it test whether language models understand human motivations. The term “status quo bias” describes a pattern in the responses; it should not be mistaken for evidence that a model has a human-like psychological process. [1]

It is reasonable to ask whether advice that favors existing options could matter if people rely on it. But that is a question for further research: this benchmark does not measure user trust, behavior or real-world effects. Any claim that fluent AI advice reinforces human inertia should therefore be treated as a possibility to investigate, not a finding of this study.

Using AI advice with care

The benchmark is not a substitute for evidence about a specific decision. When using a chatbot for a consequential choice, readers can:

  • Check factual claims against reliable, independent sources.
  • Ask for a comparison of the existing option and alternatives using the same criteria.
  • Request the strongest case for changing course as well as the case for keeping the current arrangement.
  • Treat the response as one input rather than a verdict.

These are general suggestions, not demonstrated remedies from the study. The report discusses the possibility of addressing status-quo patterns through application design and prompting; whether such changes reliably improve advice requires evaluation. [1]

For related perspectives, see our companion pieces on the psychological foundations that shaped artificial intelligence and strategies to counteract groupthink.

The Waterloo research offers evidence about how particular models answered particular climate-related prompts. It does not settle how people use AI advice, or what effects that advice may have. Those questions call for studies of users as well as models.

Sources

[1] Neuroscience News, report on the University of Waterloo study published in Environmental Research Letters: https://neurosciencenews.com/ai-chatbots-status-quo-social-neuroscience-31295. This is a secondary report, not the primary paper. Numerical descriptions retained here are attributed to that report; vehicle and recipe findings are qualified because the available account does not establish the comparison details or the precise meaning of the recipe percentage.