People who use AI may pay a social price, according to new psychology research

A new study highlights a potential downside to using artificial intelligence at work: people who rely on AI tools to get their job done are often judged more harshly by others. Despite AI’s potential to boost productivity, workers may face negative social evaluations — especially around their competence, effort, and motivation — just for using it. The research, published in the Proceedings of the National Academy of Sciences, provides evidence that this perception can even affect hiring decisions.

Generative artificial intelligence refers to tools like ChatGPT and other systems that can produce humanlike text, images, or code. These tools are becoming increasingly common in workplaces across industries, offering assistance with writing, analysis, creative tasks, and more. While they can enhance performance, some people worry that relying on AI might make them look lazy or replaceable. That concern may not be unfounded.

The research team — social scientists studying how people evaluate one another in professional contexts — wanted to understand whether using AI would actually harm a person’s reputation at work. Drawing on attribution theory, which explores how people explain others’ actions, the researchers reasoned that observers might see AI use as a sign of lower ability or effort. Just as past studies have shown that people sometimes assume help-seeking reflects personal weakness rather than situational need, the team predicted that AI users might face similar biases.

“I became interested in this topic after conversations with employees at several different organizations who expressed hesitation about using generative AI at work not because of issues with its output, but because of the social dynamics associated with AI use,” explained study author Jessica Reif, a PhD candidate at Duke University Fuqua School of Business. “These discussions, coupled with industry surveys that suggested people hide their AI use at work, made me want to examine whether the social evaluation penalty employees seem to anticipate is real.”

The researchers conducted four experiments involving more than 4,400 participants. The studies were preregistered, meaning the team documented their methods and hypotheses in advance to reduce bias. Across all experiments, the researchers examined both people’s expectations about how they would be judged for using AI and how observers actually evaluated others who used AI tools.

Free daily newsletter

In the first study, 497 participants imagined using either a generative AI tool or a traditional dashboard tool to complete a task at work. They then rated how they thought managers and coworkers would perceive them. People who imagined using AI expected to be judged as lazier, more replaceable, and less competent and diligent than those using non-AI tools. They also said they would be less likely to disclose their use of AI to others. These responses suggest that workers are aware of a possible stigma around AI use.

But are these fears justified? In the second study, the researchers asked 1,215 participants to read brief descriptions of hypothetical employees and rate them on various traits. The employees were described as receiving help from either AI, a human coworker, or no help at all. Participants consistently rated the employees who used AI as lazier, less competent, less diligent, and less independent than those in the other two conditions. The effect was not tied to the employee’s age, gender, or job type. These results show that people really do hold negative views of AI users, even when the help received is essentially the same.

“We were surprised that the social evaluation penalty in our studies occurred regardless of the age, occupation, or gender of the employee we described,” Reif told PsyPost. “In total, we compared 384 unique stimuli that systematically varied these attributes, in addition to whether the target was described as getting help from AI, getting help from another source, or getting no help at all. For instance, I thought it might be plausible that evaluations of AI users would vary depending on whether the employee described was early in his or her career (e.g., age 25) versus more experienced (e.g., age 40+). We did not find that to be the case.”

The third study tested whether these perceptions influence real decisions. In this case, one group of participants (801 people) completed a visual task and reported how often they used AI. A second group (1,718 people) played the role of hiring managers and chose between candidates who either used AI daily or not at all. They were told their pay would depend on the candidate’s performance, which gave them an incentive to choose carefully.

Overall, managers who did not use AI themselves tended to favor candidates who also did not use AI. In contrast, managers who used AI more often were more likely to prefer candidates who used AI daily. This suggests that people’s own experience with AI influences how they view others who use it. Those who are unfamiliar with the technology may be more skeptical or suspicious of its users.

The final study aimed to understand why these biases occur and whether they could be reduced. The researchers asked 1,006 participants to evaluate hypothetical candidates applying for tasks that were either manual (like handwriting notes) or digital (like sending personalized emails). Some candidates were described as using AI regularly, while others used traditional tools like Microsoft Office. Participants who did not use AI themselves were more likely to view AI users as lazy — and this perception translated into lower evaluations of job fit, especially for manual tasks.

However, when the task was digital and clearly suitable for AI assistance, the penalty disappeared. In fact, AI users were even seen as slightly better suited for the digital task than non-AI users. The researchers also found that frequent users of AI were less likely to judge AI users as lazy, suggesting that familiarity can help reduce the stigma.

“The top takeaway from this work is that AI use can carry a social cost,” Reif explained. “In our studies, employees described as using generative AI were evaluated as being lazier, less competent, and less diligent than employees using other tools or sources of help at work to perform the same task. The irony in our findings is that while some employees likely use AI because they are motivated to be more productive at work, its use may lead others to evaluate them as less motivated.”

One strength of the research is its experimental design, which allowed the team to isolate the effect of AI use from other variables. However, the authors caution that their work has some limitations. All the studies relied on online samples rather than real-world organizations, which may affect how well the results translate to actual workplaces. Also, the descriptions of AI tools were intentionally broad, which might not reflect the variety of tools and use cases that exist today.

“A key limitation of our studies is that the evaluators did not personally know the targets they were evaluating and were simply reporting their first impression upon reading about them,” Reif noted. “For example, we can’t say how perceptions about an employee who has a longstanding reputation for being a hard worker might change if he or she starts using AI. It could be the case that the effects we observe in our studies are weaker when the evaluator has more knowledge about the target or an existing work relationship.”

“One future direction I am excited about is unpacking why the social evaluation penalty we document occurs. For example, it may be the case that evaluators make assumptions about how those who use AI are spending the time they save. In addition to research on social evaluations, I am conducting research that investigates other social dynamics surrounding AI use in the workplace, such as how AI influences who seeks advice and information from whom.”

The study, “Evidence of a social evaluation penalty for using AI,” was authored by Jessica A. Reif, Richard P. Larrick, and Jack B. Soll.