Who Gets Hired, Paid, and Liked? Who Gets Credit? New Research Examines AI’s Role in Writing and the Workplace

As the way we communicate in this era changes, new questions are emerging: Does AI-assisted writing change how readers evaluate the person behind it? When should writers disclose AI involvement? And what happens to professional identity once an AI tool starts doing the work?

Three papers from the University of Chicago’s AI & Me Lab, led by Assistant Professor of Computer Science and Data Science Mina Lee, approach these questions from different angles, arriving at findings that complicate easy assumptions about AI’s role in the workplace and in writing. These three papers, among others from Mina Lee’s AI & Me Lab, were accepted to be presented at the Association of Computing Machinery’s (ACM) Conference on Human Factors in Computing Systems (CHI) and ACM’s Conference on Fairness, Accountability, and Transparency (FAccT).

AI’S BOTTOM-UP EFFECT ON REDUCING GENDER BIAS IN HIRING EVALUATIONS

In the Writing with AI Can Reduce Gender Bias in Hiring Evaluations study that worked with 672 participants, Lee, in collaboration with Psychology Research Assistant Alicia Liu, and Neubauer Family Assistant Professor of Psychology Xuechunzi Bai tested whether an AI writing assistant could reduce gender bias in hiring evaluations, not by telling evaluators to be fair, but by subtly shifting the language evaluators used as they wrote.

Existing approaches that work to reduce bias in evaluation contexts historically have been ineffective or have yielded mixed results. Approaches to reducing stereotype bias in evaluators often depend on bias training sessions and exposing individuals to counter-stereotypical examples, such as seeing women in leadership roles or men in caregiving positions. This must be paired with personal reflection, where evaluators actively examine their own prejudices, and ongoing behavioral monitoring to ensure those insights translate into fairer decision-making. A source of bias can be observed in the language used in professional documents like hiring memos and performance reviews, where certain word choices can implicitly reinforce stereotypes. To address how word choice can inform implicit bias, this research team has explored linguistic interventions that function as implicit, bottom-up nudges applied at the precise moment language is being produced. By intervening at the point of writing, these strategies aim to interrupt biased patterns before they become embedded in official records that shape careers.

Participants reviewed resumes for two closely matched candidates—”Jennifer” and “John”—applying for an entry-level financial analyst role. The resumes were then removed, and participants were asked to write evaluations of each candidate from memory.

As participants typed their evaluations, an autocomplete-style AI writing assistant offered short phrase suggestions. Suggestions for Jennifer varied by condition: some participants received competence-oriented phrases like “analytical and logical approach” and “confidence in financial math,” which were labeled “counter-stereotypical suggestions.” Other participants received warmth-oriented “stereotypical” phrases like “warm and friendly” and “supportive in group settings,” a control group received neutral suggestions. Suggestions for John were always neutral. While participants were not required to accept any suggestions, participants accepted support at high rates across all conditions.

After completing their written evaluations, participants completed a series of hiring-related ratings. They first rated both candidates on traits such as confidence and friendliness. They then made a binary hiring choice between the two candidates, separate salary recommendations for each, a choice of whom they would trust to lead a difficult project, and a choice of whom they would find more enjoyable to work with.

In both the neutral and stereotypical warmth-oriented conditions, Jennifer was offered significantly less than John. But in the counter-stereotypical analytical and confidence-oriented condition, the salary gap between John and Jennifer largely disappeared. Counter-stereotypical suggestions boosted Jennifer’s perceived competence and made evaluators more likely to regard her as a trusted leader. They also eliminated the salary gap.

The intervention worked not by persuasion or awareness-raising, but by reshaping the words evaluators reached for as they wrote. “By intervening directly in language production—the very medium through which stereotypes are perpetuated,” the authors said, “AI writing assistants provide a bottom-up approach to disrupting the vicious cycle.” But even as counter-stereotypical suggestions improved Jennifer’s competence ratings, leadership standing, and salary prospects, the same evaluators were less likely to say they’d enjoy working with her. This pattern is consistent with the well-documented “backlash” effect, wherein women who display what the authors call “agentic traits” (autonomy, assertiveness, goal-orientation) are penalized for violating gender norms.

The study demonstrates how AI writing tools can intervene on our own stereotype biases through language, while also showing that the terrain is more complex than any single tool can fully address. This influence across the written documents that propagate gender bias in the workplace (e.g., hiring memos, performance evaluations) could have a subtle bottom-up effect.

THE PERCEPTION OF AUTHORIAL AI DISCLOSURE

The question of when writers should disclose AI use is more contested than it might seem. A study by postdoctoral researcher Jingchao Fang and researcher Lee surveyed 727 participants and found a striking gap between how readers and writers view this obligation: writers had nearly 80% lower odds of seeing AI disclosure as necessary compared to readers.

The study also dug into the specific factors that shape people’s disclosure judgments and shared findings challenge some common assumptions. Notably, how much effort a writer invested had no significant effect on whether disclosure felt necessary, despite the intuitive belief that less effort equals less ownership, which necessitates AI disclosure. What mattered more was how directly AI-generated content made it into the final text, whether the AI’s contribution could have been easily replaced, and how much the writer actively directed the AI versus letting AI lead the process.

Zooming out to the broader professional landscape, a third study examined how AI is reshaping the work of freelance science journalists. Through interviews with 20 journalists, researchers Sachita Nishal, Lee, Nicholas Diakopoulos, and Jennifer Wortman Vaughan identified nine core professional values, including autonomy, voice, skill development, and editorial relationships, that guide where journalists draw boundaries with AI. Writers in this group were generally open to AI tools for information gathering and feedback, but resistant to using them for ideation and drafting, stages of the process where professional identity and individual expression are most deeply at stake. At a time when questions about AI’s role in writing are relatively controversial, this work focuses on important inquiry and complicates our working understanding of our perceptions of AI’s role in writing and thinking about writing. Together, all three papers will be presented at major academic venues in spring 2026, offering timely insight into one of the most pressing conversations in writing and technology today.

This article originally appeared on the Data Science Institute website.