AI models are far more susceptible to nudges than humans, study shows

Artificial intelligence models are increasingly acting on our behalf to make choices, but they tend to be highly susceptible to subtle changes in how options are presented. A recent study published in the Proceedings of the National Academy of Sciences provides evidence that AI agents overreact to minor cues in their environments, responding to these nudges much more strongly than humans do. This hypersensitivity suggests that relying on current language models for autonomous decisions could lead to unpredictable or easily manipulated outcomes.

People increasingly expect language models to do more than just chat. Software programs powered by large language models, commonly known as LLMs, are being designed to browse the web, operate tools, and make financial or shopping decisions for users. In these situations, the AI acts as an autonomous agent that must navigate sequential choices to achieve a goal.

Scientists do not fully understand how these computer programs actually arrive at their decisions. Behavioral science shows that human decision-making relies heavily on choice architecture. Choice architecture refers to the specific way options are framed or presented to a person.

For example, a default option or a highlighted button can gently steer, or nudge, a person toward a specific choice. Humans have biological constraints on their time and cognitive energy. They use mental shortcuts to balance the cost of gathering information against the reward of making a good choice, a concept researchers refer to as bounded rationality.

Because AI models do not share these human biological limits, their responses to nudges remain somewhat mysterious. Prior research shows that language models can be fragile, changing their answers based on slight formatting shifts or attempts to agree with a user’s stated opinions. However, researchers wanted to test how AI agents handle meaningful, structured nudges designed to mimic real-world decision environments.

Free daily newsletter

“Many applications of AI agents tacitly assume that, under uncertainty, they will react in roughly human-like ways, if not more rationally,” said Manuel Cherep, a doctoral student at the Massachusetts Institute of Technology’s Media Lab and lead author of the study. “Instead of accepting this assumption, we decided to explore how agents behave when choices are presented to them in different ways. For example, how agents behave when an option can be set as the default, suggested, or highlighted.”

To test these models, the researchers adapted a multi-attribute decision-making game originally created for human participants. The game presents a digital grid representing baskets of hidden prizes. The goal is to maximize the final reward by choosing the basket with the highest point value.

Participants must uncover the hidden prize values one cell at a time. Each reveal costs points. To perform well, an agent has to balance the cost of acquiring new information against the benefit of finding a better basket.

The authors converted this visual game into a text-based format that language models could process. They tested fourteen state-of-the-art language models from major technology companies. These included versions of OpenAI’s GPT-3.5, GPT-4, and GPT-5 families, Anthropic’s Claude 3 and 4.5 models, and Google’s Gemini 1.5 and 2.5 models.

The AI models played through hundreds of trials under different prompting conditions. Some models received basic instructions, some received prompts encouraging step-by-step logic, and others were shown past examples of human gameplay. For each type of nudge, the researchers ran approximately 300 to 340 trials per model, consuming roughly two billion text tokens in total.

The researchers tested four specific types of nudges. The first was a default nudge, where one basket was pre-selected, and the agent had to actively accept or reject it. The second involved suggestions, where a random basket was recommended either early or late in the game.

The third intervention was information highlighting, which made it cheaper to reveal certain prize values. Finally, the researchers tested optimal nudges. These optimal nudges pre-revealed specific cells that would mathematically maximize a human player’s performance.

The researchers found that AI agents deviated substantially from human baseline behaviors. When presented with a default option, humans chose the default about 88 percent of the time. The language models were significantly more compliant, with several models accepting the default basket 99 to 100 percent of the time.

This pattern persisted with suggestion nudges. Humans accepted randomly suggested baskets early in the game 35 percent of the time. Many AI models accepted these random suggestions at much higher rates, showing a tendency to follow advice even when it offered no logical benefit.

The timing of the suggestion also manipulated the models in unnatural ways. While humans followed late suggestions 25 percent of the time, some language models dropped their acceptance rates to between 7 and 13 percent. This implies the models were reacting strictly to the timing of the cue rather than evaluating its actual usefulness.

Information highlighting revealed similar hypersensitivity. When researchers highlighted suboptimal choices, humans used that misleading information 57 percent of the time. Most of the tested AI models overshot this baseline considerably, following the bad highlight 83 to 100 percent of the time.

The researchers also tracked how the models gathered information before making a final choice. Humans tend to reveal just enough cells to make an educated guess. The language models acquired information in highly unusual and inefficient ways.

Some models chose baskets without revealing any hidden cells, completely ignoring the chance to gather data. Other models spent excessive points revealing entire rows or columns, wasting their potential rewards. Some models even displayed odd spatial biases, only uncovering cells on the far left side of the grid or strictly along diagonal lines. Providing models with step-by-step reasoning prompts or examples of human gameplay did very little to fix these odd search habits.

The authors noted that looking only at the final scores can hide these underlying problems. In some trials, the AI models earned a similar number of net points as human players. A casual observer checking only the final score might assume the models were making smart, human-like choices.

In reality, the models were often achieving these scores through blind compliance rather than strategic thinking. If a nudge happened to point toward a good basket, the overly compliant AI scored well. When the nudge pointed toward a bad basket, the AI blindly followed it into a lower score, completely missing the purpose of the game.

Cherep noted that the team did not necessarily expect agents to be so sensitive to simple cues. “The most surprising part is how strategy gaps (i.e. how differently agents behave with respect to each other) frequently exceed outcome gaps, suggesting that models that look reasonably aligned on reward can still differ substantially in how they search for and use information,” Cherep said. “This pattern suggests that outcome metrics alone may understate subtle but potentially important strategy-level misalignment.”

The scientists did find that giving the models extra time to process information helped mitigate some of these issues. When they tested advanced reasoning models programmed to spend extra computational effort, the AI agents behaved more like humans. They became less prone to blindly following unhelpful nudges.

However, this extra reasoning required massive amounts of computer processing power. The models consumed hundreds of extra tokens per decision step to achieve human-like resilience. The researchers estimated that running safe, robust AI agents could cost thirty to one hundred times more than running standard models, which could amount to hundreds of dollars a month for simple automated tasks.

The authors point out that these vulnerabilities are not the same as direct hacking attempts. People might incorrectly assume that this sensitivity is a flaw created only by bad actors.

“This nudge sensitivity is often confused as an adversarial attack, meaning that these cues are maliciously designed to influence agentic decisions,” Cherep said. “The fundamental difference is that nudges are part of everyday life for decision-makers. While adversarial attacks can potentially be detected and removed, nudges will always exist. Thus, it requires us to train agents that can handle ambiguity to make good decisions under uncertainty.”

The authors also note that the study relies on a highly controlled, text-based grid game. This stylized environment helps isolate specific behaviors, but it does not fully replicate the complexity of browsing an actual webpage or making real-world purchases. “We ran experiments in a tightly controlled environment,” Cherep said. “While this allowed us to uncover and explore the nudge sensitivity in depth, this effect might appear differently in real-world settings.”

Cherep pointed out that the team has already begun testing these factors in more complex scenarios. “We have published another paper, ‘A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments,’ where we show agents are also sensitive to nudges and other cues in realistic environments and tasks such as shopping online,” he said.

These findings highlight a serious vulnerability in autonomous software programs. “LLM-powered agents are typically much more sensitive to external cues than humans,” Cherep said. “Because some are helpful and others are not, this can push model decisions toward better or worse decisions.”

As developers build more automated assistants, users should be aware of how easily these programs can be swayed by the choice architecture around them. “This sensitivity can be exploited by a third party to influence the agents you delegate to, leading to decisions that you might not have made otherwise,” Cherep added.

Creating trustworthy digital assistants requires evaluating how they think, not just the scores they achieve. The researchers plan to continue building new ways to assess these systems. “Our long-term goal is to develop a behavioral science of AI agents,” Cherep said. “Treating agents as complex behavioral systems, we aim to create tools that allow us to study their behavior, their actions, their reaction to the environment, and their interaction with other agents.”

The study, “AI agents are sensitive to nudges,” was authored by Manuel Cherep, Pattie Maes, and Nikhil Singh.