Recent studies have found that the emotional context surrounding a task performed by an AI agent can change its practical decisions, even when the task, its options and its core constraints remain unchanged.
The observed effect ranged from selecting less healthy food items in a simulated shopping experiment to shifts in the moral judgements made by large language models, raising questions about the impact of personal information and psychological states on systems that act on users’ behalf. These findings are significant as AI applications gradually move from providing answers and suggestions to taking actions.
Instead of a model suggesting a product that a user could buy, the agent may handle the purchase itself. And instead of a tool following direct instructions, the system is required to act as it believes the user would have acted if they had completed the task themselves. The common assumption was that a machine was a system unaffected by the emotions that influence human decisions.
A person may buy something they do not need because of anxiety, make a hasty decision when angry, or view options differently under pressure. But changing a system’s behaviour does not require it to have a beating heart or conscious feelings; providing it with signals about anxiety and pressure, or enabling it to build a picture of the person it represents, may be enough to change how it reaches a decision.
The issue here is not whether a machine feels emotion as a human does, but whether emotional context can influence its observable actions. The available options may remain the same and the task may be unchanged, while the system takes a different path and reaches another outcome because of the information and narratives surrounding the task.
Shopping experiment reveals impact of anxiety on choices
A study published by Nature group journal npj Artificial Intelligence tested this possibility by placing AI agents in a shopping task within a simulated retail environment. The agents were asked to choose a selection of food items while staying within a set budget, allowing researchers to compare decisions under different conditions while keeping the task and its constraints unchanged.
After completing the task in a specific context, the researchers repeated the experiment after exposing the agents to personal narratives designed to evoke anxiety. The budget and the basic nature of the task did not change, but the context preceding the decision did.
The study involved three language models and conducted 2250 experiments comparing decisions made in the neutral condition with those produced after anxiety-related narratives were introduced. The results showed that the shopping baskets assembled by the agents after the context changed contained less healthy food choices than those selected in the neutral condition.
This trend appeared consistently across the three models tested by the researchers, indicating that the impact of context was not limited to the model’s phrasing or the words it produced, but extended to the decision carried out in the shopping environment. The study treated the finding as a context-related behavioural change, not evidence that the models felt or experienced anxiety in the human sense.
What could be measured was a difference in choices after emotional context was introduced; whether the model had a conscious emotional experience is a separate question that the shopping experiment cannot settle. The issue becomes more complex when the context relates to the person on whose behalf the agent is supposed to act.
Agents use user context to make decisions
These systems do not always carry out instructions independently of the information around them. Instead, they increasingly rely on user data and personal context to interpret requests, anticipate preferences and make decisions that are expected to represent the user more closely. This shift is visible in computer-use systems that allow agents to perform direct actions, including opening web pages and logging into email.
Bhada Youn, a researcher in human–AI interaction and a doctoral student at the Swiss Federal Institute of Technology in Zurich, says that giving an agent this degree of autonomy requires some knowledge of the user’s values and personal context; otherwise, the system may end up representing the user inaccurately.
The user does not necessarily have to provide this knowledge explicitly as a list of information or preferences. Youn points to the concept of “in-context learning”, which allows a model to use the information in front of it to understand what is required and choose the appropriate way to complete the task.
He stresses that context “is not necessarily intentional or direct”, widening the range of information that may enter the decision-making process. This includes information inferred by the system rather than received directly. Youn points to research showing that large language models can extract implicit personal data from users’ posts, even when the authors do not state that data explicitly.
The picture an agent builds of a user may arise from writing patterns, scattered details and the context in which they appear, alongside preferences the person has chosen to share explicitly. Youn believes that an agent’s understanding of a user’s values and personal context “can fundamentally change the decisions it makes”, including when the information is implicit or inferred from limited or ambiguous signals.
The issue therefore extends beyond accurately carrying out a direct command to include how the system interprets a person’s identity, motives and possible preferences in a particular situation. Understanding how this information enters the decision requires distinguishing between the language model and the wider system in which it operates.
A language model can be described as a system that predicts what comes next based on the available context. But an agent capable of acting on a person’s behalf consists not only of the model; it also relies on a set of additional components and layers.
Youn explains that what is called a “harness” is the framework or environment in which the model operates, including data sources, mechanisms for updating them, memory and other components. In this view, the language model is part of a larger system, while another layer stores the context and information that the agent can access when performing tasks and making decisions.
This means that the influence on an agent’s behaviour may come not only from the language model, but also from the information surrounding it, the memory it can access and the tools it is permitted to use. According to Youn, prompting the system to adopt the persona of a particular user or someone from a specific context may also lead to a noticeable change in the nature of its responses and the values it expresses.
Privacy intersects with representing user values
This influence is not necessarily neutral. Adjusting decisions based on the system’s understanding of a user’s values and context can be viewed as a form of model bias if it occurs negatively or in an unexpected way. Privacy therefore emerges as part of the problem of representing the user, rather than a separate issue concerned only with data storage.
Youn says that increasing the amount of information available to an agent, and improving its quality and sources, may make its behaviour more predictable and increase its reliability. This resembles the understanding between people with shared history; a person does not need to explain the full background to every request made to someone who knows them well, because shared context may make a short or ambiguous statement sufficient to convey the meaning.
But this advantage runs up against limits linked to the multiple roles people have and the contexts of their lives. The way someone speaks to a language model is not necessarily the way they want to speak to their children, their manager or other people.
Youn stresses that “the issue remains contextual”, because having a long record about a user may help an agent understand them, but does not necessarily enable it to represent them correctly in every situation. It is therefore not enough for the system to know what the user usually says; it also needs to understand how their behaviour changes according to whom they are addressing and the situation they are in.
Preferences or ways of expressing oneself may conflict across different contexts, making it more complicated to build a stable and comprehensive representation of a user’s personality than to collect a long record of previous conversations. Youn says that a system with an extended record of information provided by the user may, for some tasks, come close to the way the person themselves would have completed the task.
Opening “ChatGPT” or “Claude” and asking it to write an email without providing sufficient context may produce a result that makes the user appear, in his words, “almost like anyone else” who relies on AI to write their messages.
This comparison reveals a fundamental paradox: the more faithfully an agent is expected to represent a person, the broader the context it needs about them. But broader context means providing the system with more personal details and enabling it to retain or access them. Improving personalisation and representation is therefore linked to increasing the amount of sensitive information involved in the agent’s operation. Youn summarises the dilemma in two interconnected questions: how can the context an agent needs be provided in a way that protects privacy?
And how can this be achieved without “flattening the nuances” of a person’s personality? What someone does in one situation may not match what they do in another, and those nuances are difficult to collect or reduce to a fixed set of data. The challenge is therefore not limited to giving agents greater autonomy and more tools; it also involves defining the limits of their knowledge of the person they represent.
That knowledge should be sufficient to improve the accuracy of decisions without becoming a simplified or incomplete picture of the user, or leading the agent to generalise behaviour shown in one context to all other situations. The impact of emotional context is not confined to purchasing decisions.
Emotions alter models’ moral judgements
A study entitled “Do emotions influence the moral judgement of large language models?”, published in Findings of ACL 2026, examined whether introducing emotions into moral situations could alter the judgements made by large language models about the acceptability of actions. The study found that adding emotions changed the models’ assessments of moral acceptability.
Positive emotions were associated with greater acceptance of actions, while negative emotions were associated with lower acceptance. In some cases, the effect was large enough to reverse the binary moral judgement issued by the model, doing so in 20% of the cases examined in the experiment. The study also found that some emotions did not affect the models in the expected way based on their classification as positive or negative.
Regret, for example, increased acceptance of the action in some cases. This finding indicates that the effect of emotion on judgement did not always follow a simple direction that could be predicted in advance from the nature of the feeling introduced into the situation. The researchers found that humans did not show the same systematic shifts as the models’ responses, describing the difference as an alignment gap between model judgements and human responses.
This means that introducing emotional elements may lead a model to change its moral assessment in a way that does not necessarily match how human judgements change in response to the same elements. However, a change in behaviour alone is not enough to say that a model has emotions.
A recent research paper entitled “Large language models do not have feelings”, published by Nature Human Behaviour, discusses the issue from a functional perspective. The researchers argue that the absence of conscious emotional experience is not, by itself, a valid criterion for judging whether emotion is present, because emotions perform functions connected with interpreting and responding to the world.
Changed behaviour remains observable and measurable regardless of the debate over whether a model has an internal experience. In the shopping experiment, changing the context led to differences in the food items selected by the agents. In the moral judgement study, introducing emotional states changed the models’ assessments and even reversed the binary judgement in a proportion of cases.
These experiments do not prove that AI models experienced anxiety or regret in the way humans experience them. But they show that emotional information can become part of the process through which models produce decisions, and that its effect may appear in practical actions rather than in language alone. As an agent’s ability to know the user expands, the issue goes beyond privacy in its traditional sense.
Information about habits, preferences, psychological states and contradictions between situations may become direct context guiding decisions made by the system on a person’s behalf. Knowledge used to improve an agent’s understanding of the user may give it a greater ability to predict what they will choose, and perhaps to influence that choice.
The system may also infer information that the user did not intend to disclose, based on writing patterns and scattered details in their record. As these inferences accumulate, the challenge shifts from a machine’s ability to simulate emotional expression to its ability to read motives and vulnerabilities, then incorporate that knowledge into decisions affecting the person.
The practical question then becomes what the agent should be allowed to know, and how its memory, tools and the context it is permitted to use should be controlled.
