ICANN & PAVE LabsUniversity of South Carolina

School of Information Science · University of South Carolina

Two AI research labs, one research program

The ICANN Lab and the PAVE Lab approach artificial intelligence from opposite ends. ICANN builds and probes the systems themselves: conversational agents, dialogue models, retrieval, and the neural methods underneath them. PAVE studies the people who use those systems and the values that shape them, and develops ways to evaluate both. Both labs are directed by Dr. Souvick “Vic” Ghosh at the University of South Carolina. Students often work across the two, and our strongest projects begin where they overlap.

ICANN Lab → PAVE Lab → Publications Join a lab
Technical lab · since 2020

ICANN Lab

Intelligent Conversational Agents and Neural Networks

ICANN builds and stress-tests the machinery of conversational information access. We model how people talk to search systems, train and probe neural architectures for dialogue and retrieval, and construct the benchmarks that make claims about these systems checkable. The lab has run since 2020 and has mentored more than thirty students at the graduate, undergraduate, and high school levels.

Conversational searchNLP Information retrievalAI evaluation

Focus areas

  1. Conversational search and spoken dialogue

    Clarification behavior, mixed-initiative dialogue, result presentation in voice-only settings, and the cost structure of a conversational session.

  2. Neural models for language, speech, and intent

    Speech act classification, spoken intent recognition, attention architectures, and NLP for code-mixed and low-resource text.

  3. Information retrieval and search behavior modeling

    Session-based behavior, searching as learning, task complexity, and question quality in community platforms.

  4. Evaluation infrastructure for AI systems

    Benchmarks, reusable harnesses, and methods for checking whether an automated evaluation measures what it claims to.

  5. Computational social science at scale

    Sentiment and stance analysis, misinformation detection, and sample-efficient methods for reading public opinion from social platforms.

Ongoing projects

  • Simulated users in system evaluation

    Whether language models can stand in for human participants in user studies, and where that substitution breaks down.

    In progress

  • Linking speech to retrieval

    Connecting what a person says in a spoken request to what a system should go and find.

    Under review

  • Affect-aware conversational agents

    Agents that register the emotional register of a turn and can explain how that reading shaped the response.

    Funded

  • Bias mitigation in generated dialogue

    Detecting and reducing bias in the language conversational systems produce. Developed jointly with PAVE.

    In development

People

Dr. Souvick “Vic” Ghosh

Director and Principal Investigator

Graduate Research Assistant

Open, Fall 2026

Doctoral Researcher

Open

Selected publications

  • Spoken conversational search: evaluating the effect of system clarifications on user experience through a Wizard-of-Oz study

    Ghosh, S., & Shah, C. (2025). Journal of the Association for Information Science and Technology, 76(5), 819–839. DOI

  • Unmasking public sentiment: a sample efficient approach to analyzing public opinion on US aid to Ukraine

    Ghosh, S., Ghosh, S., Dewitt, N., & McCoy, D. (2025). HICSS 58, 2459–2469.

  • Exploring the economics of conversational search sessions

    Ghosh, S., Gogoi, J., & Chua, K. (2024). Aslib Journal of Information Management, 76(4), 613–628.

  • Toward connecting speech acts and search actions in conversational search tasks

    Ghosh, S., Ghosh, S., & Shah, C. (2023). ACM/IEEE JCDL, 119–131. Vannevar Bush Best Paper nomination.

  • Sentiment-aware design of human–computer interactions

    Ghosh, S. (2023). In Computational Intelligence Applications for Text and Sentiment Data Analysis, 209–224. Academic Press.

All ICANN publications →

Join ICANN

I am recruiting students who want to build things and then find out whether they work. Prior publication is not expected.

  • Some Python, and comfort with PyTorch or an equivalent framework
  • Patience to run an experiment three times before believing it
  • Interest in dialogue modeling, evaluation tooling, or large-scale text analysis

How to apply

Human-centered lab · launching 2026

PAVE Lab

People, AI, Values & Evaluation

The PAVE Lab studies, designs, and evaluates artificial intelligence systems with a focus on the people who interact with them and the values that shape their design and use. We combine computational, experimental, and human-centered methods to understand how people communicate, seek information, collaborate, and make decisions with AI, while developing methods to evaluate these systems for effectiveness, reliability, bias, trust, and accountability.

Human-AI interactionResponsible AI AI evaluationApplied AI

Focus areas

  1. Human-AI interaction and information seeking

    How people ask, follow up, and decide when the system talks back, including the shift from queries to conversations.

  2. Responsible and ethical AI

    Audits of demographic and stereotype bias in model behavior, with attention to consequential settings such as law enforcement and healthcare.

  3. AI evaluation, trust, and accountability

    Methods for judging whether a system is reliable, whether a user's trust in it is calibrated, and who answers for it when it is wrong.

  4. Applied AI in public-interest settings

    Libraries, archives and recordkeeping, health information, reference services, and customer-facing work. Sectors where AI arrives with little support and high stakes.

  5. AI literacy, governance, and the information workforce

    Responsible computing education, professional competencies, and governance frameworks for institutions adopting AI-mediated services.

Ongoing projects

  • Auditing bias in model judgments

    Scenario-based testing of how demographic details change the decisions large language models recommend in high-stakes domains.

    Accepted, ASIS&T 2026

  • Human-AI collaboration in qualitative analysis

    Where model-assisted coding of health-related video content matches trained human analysts, and where it fails in ways that would change a study's conclusions.

    Funded

  • Bias and stereotype in generated dialogue

    What conversational systems reproduce when they generate open-ended text about people. Developed jointly with ICANN.

    In progress

  • InterPARES Trust AI

    AI for retention and disposition in trusted digital recordkeeping repositories, and assessment of personal information in records. AI consultant and co-applicant.

    SSHRC · through 2026

  • CIRCLE

    Cross-Campus Interdisciplinary Responsible Computing Learning Experiences: embedding ethics into computing coursework across multiple institutions.

    Mozilla Foundation · co-PI

  • Technology, governance, and hope

    An international program on how societies steer the technologies they adopt.

    Gilman Program · co-PI

People

Dr. Souvick “Vic” Ghosh

Director and Principal Investigator

Doctoral Researcher

Open, Fall 2026

Graduate Research Assistant

Open, Fall 2026

Selected publications

  • Probability and prejudice: a scenario-based audit of demographic bias in LLM judgments across law enforcement and healthcare

    Ghosh, S., Mehra, B., & Lu, K. (2026). Proceedings of ASIS&T, 63. Accepted.

  • Exploring stereotype bias in large language model dialogue generation

    Ghosh, S., Dubin, R., & Charette, C. (2026). Proceedings of ASIS&T, 63. Accepted.

  • Scrolling for comfort: human-information interaction around palliative care on TikTok

    Ghosh, S., Du, T., & Dubin, R. (2026). Proceedings of ASIS&T, 63. Accepted.

  • Embedding responsible computing in an undergraduate computer vision course

    Attar, N., Ghosh, S., Heller, P., Villagran, M. A., & Hofman, D. (2026). AIxHEART. Accepted.

  • Assessing artificial intelligence and social justice intersections: challenges and opportunities

    Ghosh, S., Fu, H., Hofman, D., & Mehra, B. (2026). Advances in Librarianship, Emerald. Book, in progress.

All PAVE publications →

Join PAVE

PAVE is recruiting its first cohort for Fall 2026. I am looking for students curious about what AI does to people and institutions.

  • Information science, HCI, communication, public health, education, law, or computing
  • Interview and survey design, qualitative coding, or statistics, or the willingness to learn one
  • Ability to write clearly about a messy problem

How to apply

Working across both labs

Students are welcome in both labs, and several projects need both sides. Building a bias audit takes engineering; interpreting what the audit found takes people who study people. If you are unsure which lab fits, write and describe the problem rather than the method.

Prospective doctoral applicants should make contact before the application deadline. Undergraduate and master's students at the University of South Carolina can ask about independent study and research assistantships at any point in the semester. Details are on the join page.

souvick.ghosh@sc.edu Google Scholar Personal site