Dr. Ihsan Ayyub Qazi @ LUMS
Professor of Computer Science
Syed Babar Ali School of Science & Engineering
Lahore University of Management Sciences (LUMS)
Embedded Files
I research networked systems (e.g., AI systems, social media, cloud) and their societal impacts, focusing on clinical AI and digital health, digital development and trustworthy AI.
Research: Digital Health & AI in LMICs
Interview: Bridging the Digital Divide
What ChatGPT means for Higher Education?
Research: Cognitive Reflection & Misinformation
Feb, 2026 [Nature Health]: Our randomized controlled trial demonstrates that physicians trained in AI-literacy can leverage large language models to significantly enhance diagnostic reasoning in resource-limited settings. Among 58 physicians in Pakistan who completed a 20-hour AI training curriculum, those with LLM access achieved diagnostic accuracy scores nearly 68% higher than those using only conventional resources (71.4% vs 42.6%), without requiring additional time per case. These findings suggest that appropriately trained clinicians using LLMs could help bridge critical diagnostic gaps in low- and middle-income countries, where diagnostic errors remain a leading cause of preventable harm.
Feb, 2026 [Scientific Reports]: As LLMs are increasingly deployed to verify truth at scale, we asked: Do these models actually know what they don't know? Evaluating 9 leading LLMs against 240,000 human annotations in 47 languages, we uncovered a striking Dunning-Kruger effect. Smaller, accessible models are frequently 'confidently wrong,' whereas the most capable models show higher accuracy but lower confidence. This gap worsens in non-English languages and underrepresented regions. Because organizations in the Global South rely on affordable models, this bias risks creating a two-tiered information ecosystem. Our research urges better confidence calibration and region-aware AI to ensure global information integrity.
Jan, 2026 [CHI Conference on Human Factors in Computing Systems]: As LLMs reshape clinical workflows, tracking how physicians engage with these tools is critical to understanding their mental models of AI. Our mixed-methods study analyzes physician-LLM interactions as clinicians navigate complex diagnostic vignettes. Through interaction logs and interviews, we identify diverse prompting strategies, from "role assignment" to "cautious scaffolding," anchored by steadfast human oversight. Ultimately, our findings reveal a pragmatic enthusiasm for AI as a "second brain" in LMICs, providing a roadmap for responsible generative AI integration in global healthcare.
Aug, 2024 [ACL]: We developed a novel Urdu deepfake audio dataset targeting two spoofing attacks (Tacotron, VITS TTS). The dataset construction involves careful consideration of phonemic cover and balance. AASIST-L evaluation yields EERs of 0.495 (TTS) and 0.524 (Tacotron), with speaker variation. A human study shows limited detection ability, with 1 in 3 fake samples mistaken as real. Our work aids deepfake detection in low-resource languages, bridging critical gap in existing datasets.
Jun, 2023 [Journal of Development Economics]: Misinformation is a growing concern in developing countries with potentially far-reaching consequences. We evaluate whether educational interventions improve discernment of news using a randomized control trial. We find no effect of video-based general educational messages. However, when such video messages is combined with personalized feedback, accuracy rate improves. You can read the full paper here.
News & Highlights
News & Highlights
2026
2026
Our work, “MAGIC: An International Network for Evaluating Generative Artificial Intelligence in Global Health,” has been accepted at Nature Medicine.
Our work on preserving clinical judgment in AI-assisted diagnosis has been accepted at NEJM AI.
Our study of automatic speech recognition failures in low-resource South Asian languages has been accepted at INTERSPEECH 2026.
Our work on automation bias in physician-AI collaboration has been accepted at NEJM AI.
Our work on physician-AI collaboration among AI-trained physicians has been published in Nature Health (paper; Nature coverage).
Our paper, “Large Language Models Show Dunning-Kruger-Like Effects in Multilingual Fact-Checking,” has been accepted at Scientific Reports.
Our work on the robustness of LLM-based ambient scribes has been accepted at CHIL 2026.
Two of our randomized controlled trials on physician-AI collaboration were highlighted among the significant developments in clinical AI in the State of Clinical AI Report 2026 by ARISE, a Stanford–Harvard research network (report; Stanford Medicine coverage).
Our paper, “Prompting, Oversight, and Adoption: Physicians’ Use of Large Language Models for Diagnostic Reasoning in an LMIC,” has been accepted at CHI 2026.
Our paper, “POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization,” has been accepted to Findings of ACL 2026.
Our paper, “Safeguarding Children at Scale: Cost-Effective Multimodal LLM Detection of Inappropriate YouTube Advertising,” has been accepted at The Web Conference 2026.
2025
2025
Our work on quality-preserving extreme image compression using generative AI was accepted at ECML PKDD 2025.
Received a 2025 OpenAI Researcher Access Program Award.
Our work on hybrid active learning for neural machine translation was accepted at COLING 2025.
Taught an eight-session Health Data Science module as part of the Certificate in Health Professions Education (CHPE) at LUMS, working with 62 health professionals representing 60 hospitals and medical universities across 15 cities in Pakistan.
Joined Sehar Ghafoor on National Television's Smart World for a conversation about AI, satellite internet, the gender gap in technology, and the importance of critical thinking and empathy in the AI era. Watch the conversation.
2024
2024
Served as TPC Co-Chair of ACM CoNEXT 2024 with Gareth Tyson.
Our work introducing the first Urdu deepfake-audio dataset for deepfake detection was accepted to Findings of ACL 2024.
Our large-scale study of third-party serving infrastructure in government digital services appeared at ACM IMC 2024. The study covers 61 countries across all major world regions, representing more than 82% of the world's Internet population.
Launched GradEcho, a community designed to broaden access to high-quality advice and mentorship for graduate-school applicants. Visit GradEcho.
Two of our papers on advertising and child-oriented YouTube content were accepted at The Web Conference 2024: “Analyzing Ad Exposure and Content in Child-Oriented Videos on YouTube” and “Uncovering the Hidden Data Costs of Mobile YouTube Video Ads.” All student co-authors were LUMS undergraduates.
Delivered a mini-course on Health Data Science as part of the LUMS Certificate in Health Professions Education, working with nearly 50 physicians from hospitals across Pakistan.
Led the Google exploreCSR Workshop at LUMS on “Navigating the Computing Research Landscape: Skills and Strategies for Aspiring Female Students.”
Served as an Area Chair at The ACM Web Conference 2024 for the Systems and Infrastructure for Web, Mobile, and WoT track.
Google Sites
Report abuse