Research
1. Modeling the Interviewer: Leveraging LLMs to Uncover Personality Mismatch Effects in Interview Assessments
(with Balaji Padmanabhan, Siva Viswanathan) [Job Market Paper]
Abstract
This study presents an automated, dynamic interview system that systematically embeds personality traits into LLM-based interviewers. Through this framework, we investigate the impact of personality similarity, between interviewers and interviewees. We conducted a randomized experiment involving graduate students, assigning participants to a LLM interviewer for behavioral voice-based interviews. Our findings indicate that interviewer-interviewee alignment on personality traits led to higher predictive accuracy in assessing the interviewee’s specific personality trait. Conversely, when there was a mismatch, topic modeling of interview transcripts revealed that the interviewee’s language became more aligned with the interviewer’s personality, diminishing the predictive accuracy of the trait assessment.2. Impact of AI on Reviews and Outcomes
Link
(with Siva Viswanathan)
Abstract
This study investigates how users adopt AI-generated text in the context of online review writing and explores its impact on review quality, operationalized through metrics of length, helpfulness, and linguistic diversity. We employ an empirical research design, combining observational data from a major food technology review platform and controlled experiments conducted via the crowdsourcing platform Prolific. Our findings reveal that users expressing negative experiences are more inclined to adopt AI-generated assistance and do so more extensively when articulating negative reviews. Moreover, AI-assisted reviews are found to be less diverse, shorter in length and perceived as less helpful by readers.
3. Probing Personality with AI: Evaluating LLMs as Interviewers in Automated Hiring Systems
(with Balaji Padmanabhan, Siva Viswanathan)
Abstract
This paper aims to explore the potential differences in personality assessments between LLM-interviewer driven dynamic interviews and static question driven interviews. We design a LLM based dynamic interview system to conduct job interviews and find LLM based system to outperform the static interview system in assessing candidate personality.Furthermore, we develop in-context learning based method to improve the performance of the dynamic LLM based interview system
4. Do Employee Verification Mechanisms Alter Cultural Signals in Employer Review Platforms
(with Vladimir Martirosyan)
Abstract
This study examines how employee verification mechanisms affect the way culture is expressed in employer review platforms. Using large-scale data from two of the biggest review sites, we find that verification changes the cultural signals in reviews. Specifically, verification changes the discussion to emphasise more on teamwork and organizational structure, while non verification in platforms skews reviews to be more positive and discussions focus more on industry competition and company performance.
5. Electric Stoves as a Solution for Household Air Pollution: Evidence from Rural India
(with E. Somanathan, Marc Jeuland, Eshita Gupta, T.V. Ninan, Utkarsh
Kumar, Vidisha Chowdhury, Suvir Chandna, Michael Bergin, Karoline
Barkjohn, Christina Norris, T. Robert Fetter and Subhrendu Pattanayak)
Abstract
We collected minute-by-minute data on electricity availability, electric induction stove use, and kitchen and outdoor particulate pollution in a sample of rural Indian households for one year. Using within household-month variation generated by unpredictable outages, we estimate the effects of electricity availability and electric induction stove use on kitchen PM2.5 concentration at each hour of the day. Electricity availability reduces kitchen PM2.5 by up to 50 μg m3, which is between 10 and 20 percent of peak concentrations during cooking hours. Induction stove use instrumented by electricity availability reduces PM2.5 in kitchens by 200-450 μg/m3 during cooking hours
