Evidence of a social evaluation penalty for using AI

Source: Reif, Larrick, Soll (Duke), PNAS · 2025-05-08

As AI tools become increasingly prevalent in workplaces, understanding the social dynamics of AI adoption is crucial. Through four experiments with over 4,400 participants, we reveal a social penalty for AI use: Individuals who use AI tools face negative judgments about their competence and motivation from others. These judgments manifest as both anticipated and actual social penalties, creating a paradox where productivity-enhancing AI tools can simultaneously improve performance and damage one’s professional reputation. Our findings identify a potential barrier to AI adoption and highlight how social perceptions may reduce the acceptance of helpful technologies in the workplace.

Despite the rapid proliferation of AI tools, we know little about how people who use them are perceived by others. Drawing on theories of attribution and impression management, we propose that people believe they will be evaluated negatively by others for using AI tools and that this belief is justified. We examine these predictions in four preregistered experiments (N = 4,439) and find that people who use AI at work anticipate and receive negative evaluations regarding their competence and motivation. Further, we find evidence that these social evaluations affect assessments of job candidates. Our findings reveal a dilemma for people considering adopting AI tools: Although AI can enhance productivity, its use carries social costs.

The attribution processes that shape evaluations of help-seeking behavior can be fruitfully extended to understand the social implications of AI use in today’s workplaces. AI technologies present a dilemma to the people who use them. On the one hand, AI can enhance human performance on a variety of tasks (9, 10). People thus have strong incentives to use AI, as it might improve their performance at work. On the other hand, AI represents a powerful form of assistance. Consequently, using AI may raise doubts about one’s own abilities and motivation. Consistent with this notion, a recent industry survey found that apprehension about being perceived as lazy ranks among the top concerns of people who use AI at work (11). Further, numerous reports suggest that people actively conceal their AI use in professional settings (12, 13). This apparent tension between AI’s documented benefits and people’s reluctance to use it raises a critical question: are people who use AI actually evaluated less favorably than people who receive other forms of assistance at work? Extending theories of attribution, we propose that observers will be likely to make (negative) dispositional inferences about people who receive help from AI relative to people who receive other forms of help.

In four preregistered studies, we examine this prediction from the lens of both the help recipient and observer. In Study 1, we show that people who receive help from AI believe they will be evaluated as lazier, less competent, and less diligent than people who receive similar help from non-AI technologies. In Study 2, we demonstrate that this fear is justified: observers perceive people who get help from AI as lazier, less competent, and less diligent than people who get help from other sources. Study 3 shows that managers who do not use AI themselves may act on their negative assumptions of people who use AI in an incentive-compatible hiring task. Finally, Study 4 shows that perceptions of laziness mediate the relationship between AI use and assessments of poor task fit in a hiring scenario.

The notion that using technologies that reduce the need for effort or ability can cast doubt on one’s competence and motivation has echoed in debates over new tools for centuries. For example, Plato’s Phaedrus (370 BC) recounts a question about whether people who relied on a new invention for learning (writing) would ever develop true wisdom (14). More recently, educators have questioned how using tools such as calculators would affect students’ ability to develop mathematics skills, and studies have documented patients’ tendency to assume that physicians who use diagnostic aids are less capable (15, 16). Unlike previous tools that simply performed specific operations or made predictions, AI tools may be perceived as more agentic because they can learn from experience and operate more autonomously (17). Such powerful tools may intensify doubts related to the ability and effort of their operators.

Effort and ability have long been among the primary metrics by which people are evaluated in professional environments (18, 19), and thus people aim to project these valued qualities to others (7, 20). Assistance of any kind creates attributional ambiguity, raising impression management concerns for recipients. When a person receives help to perform a task, observers must determine how much credit is due to the person versus the assistance (21). In doing so, they may imagine counterfactual scenarios in which the person did not receive the assistance and anticipate what outcome might have occurred (22). This process extends to judgments about effort, where observers might consider both the actual effort they observed and counterfactual possibilities about how the outcome might have been different had the target exerted more effort (23). Receiving help may thus cast a shadow of doubt over one’s ability and willingness to exert effort (3, 24). Conscious of this ambiguity, recipients of help may worry that observers will discount their competence and motivation and withhold any indication that they received help (8, 25). These social evaluation concerns may be justified, as observers could interpret the decision to utilize assistance as a signal that the recipient is not willing or able to perform the task themselves. Behaviors that are not considered mainstream—or part of the consensus—are especially likely to elicit dispositional attributions (24). Emerging technologies are, by definition, new, and therefore, their use is unlikely to be perceived as customary. Extending attribution theory, we propose that the use of emerging technologies that reduce the need for effort or ability is especially likely to evoke negative dispositional inferences about their operators.

The anticipation of these negative perceptions presents a dilemma to people considering whether to adopt AI: using AI may simultaneously enhance their productivity but undermine others’ perceptions of their competence and motivation. Although there is a significant body of work examining how people perceive AI systems themselves (26), we know little about how evaluators perceive the people who use them. Understanding whether receiving help from AI in fact leads to a social evaluation penalty is crucial for anticipating and addressing challenges related to the adoption of AI.

We next examined how participants believed others would evaluate them on the two dimensions of agency we measured in this study: competence and diligence. Participants reported that they believed others would judge them as less competent in the AI Tool condition (M = 4.72, SD = 1.46) than in the Non-AI Tool condition (M = 5.45, SD = 1.22), t(477.5) = 6.00, 95% CI [−0.96, −0.49], P < 0.001). Similarly, they reported that they expected to be perceived as less diligent in the AI Tool condition (M = 4.66, SD = 1.44) than in the Non-AI Tool condition (M = 5.25, SD = 1.28), t(487.5) = 4.86, 95% CI [−0.83, −0.35], P < 0.001). These results support the notion that people known to use more agentic technologies believe they may be evaluated as less agentic themselves.

Finally, we examined the two variables related to disclosure. Participants in the AI Tool condition reported that they would be less likely to disclose the use of the tool to their managers (M = 4.91, SD = 1.59) than participants in the Non-AI Tool condition (M = 5.25, SD = 1.55), t(494.3) = 2.42, 95% CI [−0.62, −0.06], P = 0.016). Participants in the AI Tool condition also reported less willingness to disclose the use of the tool to colleagues (M = 4.85, SD = 1.58) than participants in the Non-AI Tool condition (M = 5.17, SD = 1.57), t(494.8) = 2.23, 95% CI [−0.59, −0.04], P = 0.026). These results are consistent with our prediction that people who use AI tools may be reluctant to disclose their use to others at work.

We next tested whether evaluators assess employees known to receive help from AI more negatively than people who receive other forms of help or people who received no help at all. We recruited 1,215 online participants to complete a study in which they read a paragraph about an employee and rated how lazy they perceive the employee to be, as well as how they view the employee on six dimensions of agency: competent, diligent, ambitious, independent, self-assured, and dominant (27).