Artificial intelligence models are meant to make choices for us, but they tend to be highly susceptible to subtle changes in the way choices are presented. Recent research published in Proceedings of the National Academy of Sciences This provides evidence that AI agents overreact to small cues in the environment and react much more strongly to these nudges than humans. This hypersensitivity suggests that relying on current language models to make autonomous decisions can lead to unpredictable and easily manipulated outcomes.
People increasingly expect language models to do more than just chat. Large language model-based software programs, commonly known as LLMs, are designed to help users browse the web, interact with tools, and make financial and shopping decisions. In these situations, the AI acts as an autonomous agent and must navigate a series of choices to achieve a goal.
Scientists don’t fully understand how these computer programs actually reach their decisions. Behavioral science has shown that human decision making is highly dependent on choice structures. Choice architecture refers to the particular way options are assembled and presented to people.
For example, default options or highlighted buttons can gently guide or push the user towards a particular choice. Humans have biological constraints on time and cognitive energy. They use mental shortcuts to balance the cost of gathering information with the rewards of making good choices. This is a concept researchers call “bounded rationality.”
Since AI models do not share human biological limitations, their responses to nudges remain mysterious. Previous research has shown that language models can be fragile and can change answers based on slight changes in formatting or attempts to agree with a user’s stated opinion. However, the researchers wanted to test how the AI agent would handle meaningful, structured nudges designed to mimic real-world decision-making environments.
“Many applications of AI agents implicitly assume that under uncertainty they will react in a manner roughly similar to, if not more rational than, humans,” said Manuel Tchelep, 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 investigate how the agent behaves when choices are presented in different ways. For example, we decided to investigate how the agent behaves when options can be set as defaults, suggested, or highlighted.”
To test these models, the researchers adapted a multi-attribute decision-making game originally created for human participants. The game displays a digital grid that represents a basket containing hidden prizes. The goal is to choose the basket with the highest point value to maximize your final reward.
Participants must reveal hidden prize values one cell at a time. Each publication costs points. For an agent to perform well, it must balance the cost of acquiring new information with the benefit of finding a better basket.
The authors converted this visual game into a text-based format that a language model can process. They tested 14 state-of-the-art language models from leading 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 model ran hundreds of trials under various prompt conditions. Some models received basic instructions, some received prompts encouraging step-by-step logic, and others were shown past examples of human gameplay. The researchers ran about 300 to 340 trials per model for each type of nudge, consuming about 2 billion text tokens in total.
The researchers tested four specific types of nudges. The first was a default nudge, in which one basket was preselected and the agent had to actively accept or reject it. The second one was about suggestions, where random baskets were recommended early or late in the game.
The third intervention was information highlighting, which lowered the cost of revealing the value of a particular prize. Finally, the researchers tested the optimal nudge. These optimal nudges pre-revealed specific cells that mathematically maximized the human player’s performance.
Researchers have found that AI agents deviate significantly from basic human behavior. When presented with a default option, humans choose the default approximately 88% of the time. Language model compliance has improved significantly, with some models accepting the default basket between 99 and 100%.
This pattern continued with nudges through suggestions. Humans accepted randomly suggested baskets 35% of the time early in the game. Many AI models accepted these random suggestions at a much higher rate and showed a tendency to follow advice even when it had no logical merit.
The timing of the proposal caused the model to be manipulated in an unnatural way. Humans followed slow suggestions 25 percent of the time, while some language models saw acceptance rates drop to between 7 and 13 percent. This means that the model was reacting strictly to the timing of the cue, rather than evaluating its actual usefulness.
Highlighting the information revealed similar hypersensitivity. When researchers emphasized the next best option, humans used the misleading information 57% of the time. Most of the tested AI models significantly exceeded this baseline, following bad highlights between 83 and 100% of the time.
The researchers also tracked how the model gathered information before making its final choice. Humans tend to reveal enough cells to make educated guesses. The language model obtained information in a highly unusual and inefficient way.
Some models chose the basket without revealing any hidden cells, completely ignoring the opportunity to collect data. Other models spent excessive points to reveal entire rows or columns, wasting potential reward. Some models displayed a strange spatial bias, showing only cells along the left edge of the grid or strictly diagonally. Providing the model with step-by-step reasoning prompts and examples of human gameplay did little to correct these strange search habits.
The authors pointed out that looking only at the final score could mask these fundamental problems. In some trials, the AI model earned the same number of net points as the human player. A casual observer looking only at the final score might think the model is making wise and humane choices.
In reality, models often achieved these scores through blind compliance rather than strategic thinking. If the nudge happened to point to a good basket, the overly compliant AI scored well. When the nudge pointed to a bad basket, the AI blindly followed suit and lowered the score, completely missing the point of the game.
Cherep noted that the team didn’t necessarily expect agents to be so sensitive to simple cues. “The most surprising thing is that the strategy gap (i.e. how agents behave differently with respect to each other) frequently outweighs the outcome gap. This suggests that even models that appear to be reasonably matched when it comes to rewards can differ significantly in how they seek and use information,” Cherep said. “This pattern suggests that outcome measures alone may underestimate subtle but potentially important strategic-level misalignments.”
Scientists have found that giving the model extra time to process the information alleviates some of these problems. When we tested advanced inference models that were programmed to spend additional computational effort, the AI agents behaved more like humans. They are less likely to blindly follow unhelpful nudges.
However, this additional inference required a large amount of computer processing power. The model consumed hundreds of additional tokens per decision step to achieve human-like resilience. Researchers estimate that running a secure and robust AI agent can cost 30 to 100 times more than running a standard model, which can reach hundreds of dollars per month for simple automation tasks.
The authors note that these vulnerabilities are not the same as a direct hacking attempt. People may mistakenly assume that this sensitivity is a flaw created only by malicious actors.
“This nudge sensitivity is often confused with adversarial attacks, meaning that these cues are maliciously designed to influence an agent’s 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 are always present. Therefore, to make good decisions under uncertainty, we need to train agents that can handle ambiguity.”
The authors also note that this study relies on a highly controlled text-based grid game. This stylized environment helps isolate certain behaviors, but it doesn’t fully replicate the complexity of viewing a real web page or making a purchase in the real world. “We conducted the experiment in a highly controlled environment,” Çelep said. “This allowed us to uncover and deeply investigate the sensitivity of microtremors, but this effect may manifest differently in real-world settings.”
Cherep noted that the team has already started testing these elements in more complex scenarios. “We published another paper, ‘A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments,’ in which we show that agents are also sensitive to nudges and other cues in realistic environments and tasks such as online shopping,” he said.
These findings highlight serious vulnerabilities in autonomous software programs. “LLM-powered agents are typically much more sensitive to external cues than humans,” Cherep says. “Some are helpful and others are not, which can push the model’s decisions toward better or worse decisions.”
As developers build more automated assistants, users need to be aware of how easily these programs are influenced by their surrounding architectural choices. “This confidentiality can be exploited by third parties to influence the agent you delegate, leading to decisions that they might not have otherwise made,” Çelep added.
To create a digital assistant you can trust, you need to evaluate how they think, not just the scores they achieve. Researchers plan to continue building new ways to evaluate these systems. “Our long-term goal is to develop behavioral science for AI agents,” Cherep says. “We aim to create tools that treat agents as complex behavioral systems and allow us to study their behavior, behavior, reactions to the environment, and interactions with other agents.”
The study, “AI agents are sensitive to nudges,” was authored by Manuel Cherep, Pattie Maes, and Nikhil Singh.
