Improving Communication with
Generative Artificial Intelligence

David Hagmann
Kirsten Geng
Catherine Tinsley
Department of ManagementThe Hong Kong University of Science and Technology
McDonough School of BusinessGeorgetown University
LLMSS Conference, October 15, 2026

Why Communication Matters

Feedback Helps Us Assess Quality

Numeric Rating
★★★★☆
4.2 / 5
OPEN-ENDED FEEDBACK
“Can provide more individualized feedback on assignments and adjust the speed of lessons to ensure all students can keep up.”

Four Studies of Communication

1
Inside the business schoolCourse Feedback
Can questions turn a sparse comment into feedback an instructor can use?
2
Inside the workplaceCritical Feedback
Does the value of AI come from eliciting information or polishing writing?
3
In a hiring decisionSelf-Promotion
Can AI reduce gender differences while preserving information about performance?
4
Across opposing viewsPolicy Disagreements (Pilot)
Can AI help people communicate with someone who disagrees?

Study 1 | Students Gave Course Feedback

353
Undergraduate studentsAt the end of the semester
From six sections of a required management course
STUDENTWrites Initial Message

Describes the instructor’s strengths and areas for improvement.

AI INTERVIEWERAsks tailored questions

Probes examples, impact, and possible improvements.

STUDENTProvides additional information

Answers from personal experience in the course.

AI INTERVIEWERRewrites the initial response

Uses the additional information provided by the student.

STUDENTReviews, edits, and submits

(87% submit without edits)

401
Independent evaluatorsOnline sample with teaching experience
Blind to condition
Each rated 10 comments

Study 1 | Original vs. AI-Assisted Message

Original Message

“Can provide more individualized feedback on assignments and adjust the speed of lessons to ensure all students can keep up.”

AI-Assisted Message

“I think the instructor could improve by providing more individualized feedback on assignments, especially since most of the discussions are group-based. For example, while there is individualized feedback for the case report, it would be helpful to have more specific comments on different parts of collaborative projects as well. Although I personally did not find the pace of the lessons too fast, I did hear from others that some parts were hard to follow, so the instructor could consider explaining concepts with more examples to help everyone understand better. In terms of class interaction, I believe it would be better if we could submit our answers and compare them, rather than just raising hands to respond, as this would collect more responses and prevent some students from not participating fully.”

Study 1 | Algorithmic Evaluations of Student Feedback

Yeomans (2021, OBHDP)
doc2concrete algorithm
Yeomans et al. (2020, OBHDP)
Conversational receptiveness

Error bars show 95% confidence intervals.

Study 1 | How Teachers Evaluated Feedback

N = 401 teachersIndependent evaluators recruited through Prolific.
10 messages eachFive baseline and five AI-assisted messages. Never both versions from the same student.
One of two domainsEach teacher evaluated either instructor strengths or instructor weaknesses.
Outcome Measures

ActionableInterestRelevantSincereSpecificTrustworthy

Study 1 | Teachers Evaluated the Feedback

Error bars show 95% confidence intervals.

Study 2 | Managers Give Critical Feedback

500
Managers giving feedback to a difficult employee or colleagueOnline sample
All had management experience
MANAGERWrites Initial Message

Writes feedback they would give directly.

AI INTERVIEWERAsks tailored questions

Probes situation, behavior, impact, and action.

MANAGERProvides additional information

Answers from personal experience.

AI INTERVIEWERRewrites the response

Uses the additional information provided.

MANAGERReviews, edits, and submits

(87% submit without edits)

AI EDITOR rewrites the initial message without seeing any follow-up answers.
754
Independent managers evaluating the messagesOnline sample with management experience
Blind to condition; each rated 9 messages

Study 2 | Questions Change What Enters the Message

Error bars show 95% confidence intervals.

Study 2 | Independent Managers Evaluated the Feedback

Error bars show 95% confidence intervals.

Engaging in Self-Promotion

Gender Differences in Self-Promotion

Self-Promotion Mixes Two Signals

8/10

Performance Evidence

Results, experience, task performance, and demonstrated contribution.

+
“…”

Articulation

Which accomplishments are recalled, selected, explained, and made salient.

=
?

Observed Message

The message a promotion committee or hiring manager ultimately evaluates.

THE PROMISEAI may reduce gender differences in self promotion
THE CONCERNAI may improve articulation so much that the message stops revealing underlying performance.

Study 3 | Candidates Wrote Self-Promotion Messages

502
Gender-balanced candidatesOnline sample
Real performance task (mathematics quiz)
Real bonus opportunity
CANDIDATECompletes an aptitude task

10 items in seven minutes; $0.10 per correct answer.

CANDIDATEWrites Initial Message

Makes the case to be selected. If selected, earns a $1 bonus.

AI INTERVIEWERAsks tailored questions

Elicits task-relevant strengths and evidence.

AI INTERVIEWERRewrites the response

Uses the additional information provided.

CANDIDATEReviews, edits, and submits

(87% submit without edits)

Elicit WTP before and after AI Interviewer stage. (Becker, DeGroot, and Marschak, 1964)

Study 3 | Gender Differences in AI Use

Error bars show 95% confidence intervals.

Study 3 | Performance and Self-Assessment

Study 3 | Valuing AI Assistance

Study 3 | Valuation of the AI Assistance Process

Error bars show 95% confidence intervals.

Study 3 | Hiring a Candidate

1,501
Decision-makersOnline sample; 92% had prior hiring experience
Each made five incentivized choices
Male CandidateSelf-Promotion MessageInitial Message or AI-assisted
VS
Female CandidateSelf-Promotion MessageInitial Message or AI-assisted
Neither assistedMale onlyFemale onlyBoth assisted
Decision-makers saw Candidate A and Candidate B. Neither candidate gender nor actual test score was shown.

Study 3 | Gender Differences in Self-Promotion?

Error bars show 95% confidence intervals.

Study 3 | How Informative Is the Message?

Error bars show 95% confidence intervals.

Expressing Policy Positions

Pilot | Policy Disagreements

“Do you think that, overall, the benefits of GenAI for humanity outweigh the risks and harms of the technology?”

101
Stage 1 · WritersConstruct and revise an argument
WRITERStates a Position

Votes Yes/No, then writes an argument for someone who disagrees.

AI INTERVIEWERAsks and Revises

Asks follow-up questions and uses the answers to revise the argument.

WRITERReviews and Edits

(93% submit without edits)

200
Stage 2 · ReadersRead and evaluate an opposing argument
READERStates a Position

Votes Yes/No before reading an opposing argument.

RANDOM ASSIGNMENTReads One Message

Receives an initial message or its AI-assisted version.

READEREvaluates the Message

Rates the author and reports whether their view changed.

Pilot | Initial Votes by Gender

“Do you think that, overall, the benefits of GenAI for humanity outweigh the risks and harms of the technology?”

Pilot | Algorithmic Measures

Pilot | Ratings and Persuasiveness

Discussion

FeedbackTailored questions elicit useful detail and improve independent evaluations of feedback.
Self-PromotionWomen better at open-ended self-promotional messages. AI helps with selection, directionally reduces gender bias.
Disagreement
(Pilot)
AI improves algorithmic measures of policy arguments, but not perceptions as judged by other participants.
Thank You!hagmann@ust.hk

Study 1 | Process of Providing Feedback

Prompt: Weakness

Now, we would like to ask you about the instructor’s areas for improvement. What did you enjoy the least about the course and the instructor’s teaching? What do you think they should do differently? Please provide as much detail as you can about what the instructor can do better.

Original Message

“Can provide more individualized feedback on assignments and adjust the speed of lessons to ensure all students can keep up.”

AI Follow-Up Questions
1

Can you give an example of a time when you felt the feedback on an assignment was not individualized enough? What would have made the feedback more helpful to you?

2

In what specific topics or parts of the lessons did you feel the pace was too fast? How did this affect your understanding?

3

Are there certain ways you prefer to receive feedback on your work that the instructor could try in the future?

Integrative Question

“Are there any aspects of classroom communication or opportunities for student questions where the instructor could improve to better support your learning?”