Research | Jason S. Lucas

Mentorship Statement

Sun, 03 May 2026 00:00:00 +0000

Jason Samuel Lucas · Assistant Professor of Information Science · University of Colorado Boulder · Director, Secure and Ethical AI Lab (SEAL)

Philosophy

My mentorship philosophy centers on a single idea: meet students where they are and scaffold pathways to independence. Students arrive at research with different preparation, different starting points, and different reasons to be there. The work of a mentor is to take that seriously, calibrate accordingly, and build the runway each student needs without flattening the standards the work deserves.

Track Record

Through Penn State’s Millennium Scholars program, I have mentored undergraduate researchers on progressively complex projects:

Kendall Reed II advanced from analyzing existing datasets to leading original research, with a forthcoming co-authored publication.
Lia Carin Djaouga progressed from literature reviews to designing cross-lingual transfer frameworks for low-resource language settings.

In both cases, the trajectory was the point: each milestone was scoped just beyond current capability, never punitively, and never below the standard of the work. I have additionally guided junior PhD students on research scoping, mentored undergraduate teams on collaborative robotics with an emphasis on inclusive problem-solving, and coordinated ENVISION: STEM Career Day, an annual event reaching 500+ young women across central Pennsylvania.

SEAL’s Mentoring Model

The Secure and Ethical AI Lab (SEAL) at CU Boulder is built around three commitments:

1. Substantive independence within a coherent program. SEAL students lead their own research questions, but those questions sit inside a shared intellectual frame: AI safety and equity across the digital language divide. Independence is not isolation. Coherence is not constraint.

2. Visibility matters. Students need to see scholars who navigated paths similar to their own. I share my trajectory openly, from Caribbean island student to R1 doctoral candidate, navigating a learning disability and limited resources, because students from underrepresented backgrounds often do not have models for the path they are on. Visibility is not a substitute for support; it is a condition for it.

3. Community over hierarchy. SEAL operates as a research community, not a chain of command. Senior students mentor junior students. Postdocs co-advise undergraduates. The PI is the scaffolding, not the ceiling. Ideas come from everywhere, and credit is distributed accordingly.

Who SEAL Looks For

I welcome PhD students, MS researchers, and undergraduates motivated by AI safety, multilingual NLP, low-resource and dialect-aware language technology, adversarial evaluation, or the broader question of who AI systems serve and who they fail. Strong candidates need not arrive with all the technical pieces in place. They need curiosity, rigor, and a willingness to engage with both the technical and the social dimensions of the work. The technical pieces, we build together.

An Open Invitation

If your research interests intersect with SEAL’s mission, I encourage you to reach out. I am especially committed to mentoring students from Caribbean and African diaspora communities, first-generation graduate students, and others underrepresented in AI research. The lab’s work is improved when its membership reflects the linguistic and cultural diversity of the communities the work is meant to serve.

Prospective students can reach me directly, or apply to the PhD program in Information Science at CU Boulder and indicate interest in SEAL in their application materials.

Last updated: May 2026 · ·

Research Statement

Sun, 03 May 2026 00:00:00 +0000

Jason Samuel Lucas · Assistant Professor of Information Science · University of Colorado Boulder · Director, Secure and Ethical AI Lab (SEAL)

Vision: Trustworthy AI Through Multilingual NLP and Security

The world is multilingual; its digital infrastructure is not. Of the roughly seven thousand languages spoken today, fewer than twenty account for the overwhelming majority of digital content, and English alone accounts for over half. This asymmetry, which I call the digital language divide, is the foundation of my research program: a gap between linguistic reality and digital representation that shapes who AI systems protect, who they fail, and who they leave exploitable.

Large language models have transformed how we process information, enabling sophisticated applications from machine translation to content generation. But the same advances create vulnerabilities. Adversaries weaponize AI to generate harmful content at scale. Information-based attacks undermine public trust and democratic institutions. Safety systems fail to protect diverse linguistic communities. My research addresses these challenges at the intersection of AI, NLP, and security, developing trustworthy AI systems that protect both technology and the communities that use it, regardless of language or resources.

The Technical Problem

Modern LLMs are built through a three-phase training pipeline: pre-training for language semantics and syntax, instruction-tuning for human instruction-following, and alignment-tuning for human values and safety. Each phase inherits the long-tail distribution of its training data, where high-resource languages dominate and low-resource languages, dialects, and creoles sit at the margins. The result is a class of fundamental vulnerabilities that adversaries exploit: safety guardrails that fail in dialects they were never evaluated on, alignment that breaks under code-switching, and instruction-following that degrades in precisely the languages most exposed to harmful content online.

The communities affected are simultaneously under-protected, because the systems were not built to recognize harmful content in their varieties, and over-exposed, because the same gaps that produce inequitable protection produce exploitable attack surfaces. My work refuses the equity-safety separation: these are not parallel problems but the same problem manifesting on two faces of the digital language divide.

Research Questions

My research program investigates three interconnected questions:

RQ1. How can we understand and mitigate vulnerabilities in AI systems, particularly for security-critical applications?
RQ2. How can we extend AI and NLP capabilities to low-resource multilingual communities?
RQ3. Can we develop robust, trustworthy AI systems that maintain effectiveness against adaptive adversaries across languages, domains, and modalities?

The Secure and Ethical AI Lab (SEAL)

These questions structure the work of the Secure and Ethical AI Lab (SEAL) at CU Boulder. SEAL builds AI systems that are safe, interpretable, and equitable across all languages and communities, advancing four interconnected research directions:

Multilingual NLP and Low-Resource AI. Extending model capabilities to underserved languages, dialects, and creoles, with particular emphasis on Caribbean and African diaspora varieties historically absent from benchmark resources.
Trustworthy AI and Information Integrity. Investigating how foundation models generate, propagate, and detect harmful content, building on my F3 framework (EMNLP 2023) for understanding the dual role of LLMs as both source and shield.
Ethical, Equitable, and Human-Centered AI. Developing socio-technical frameworks for responsible model development, articulated through my Dual Curse theory connecting colonial epistemology to multilingual AI safety failures.
AI Safety and Robustness. Building dialect-aware safety guardrails (DIA-Guard), contamination-resistant benchmarks (BLUFF), and adversarial evaluations that surface where models break before deployment does.

Methodological Commitments

SEAL’s work bridges AI innovation with security and equity, ensuring advances benefit all populations regardless of language or resources. Treating equity and safety as one problem rather than two yields different research choices: benchmarks built with affected communities rather than scraped around them, defense models distilled to run where the harm actually occurs, and evaluations designed to catch the failures that matter rather than the failures that are convenient to measure. The lab’s methodology is interdisciplinary by necessity, spanning adversarial machine learning, multilingual NLP, knowledge distillation, and security systems, with deployment in diverse real-world contexts as the standard against which results are judged.

Outlook

Over the next five to seven years, I will build SEAL into a hub for multilingual AI safety research, with sustained focus on contamination-resistant evaluation, dialect-robust safety conditioning, and AI systems for low-resource and Caribbean language communities. The work is supported through a portfolio that spans NSF programs, industry partnerships, and national-laboratory collaborations. The goal is concrete: AI systems whose protection extends to the speakers, dialects, and registers the field has historically left at the margins.

Last updated: May 2026 · ·

Teaching Statement

Sun, 03 May 2026 00:00:00 +0000

Jason Samuel Lucas · Assistant Professor of Information Science · University of Colorado Boulder · Director, Secure and Ethical AI Lab (SEAL)

Philosophy

I teach at the intersection of artificial intelligence, language, and ethics. My classrooms are spaces where students learn to build AI systems and to interrogate them, with equal seriousness given to both. Drawing on a decade of teaching at St. George’s University School of Medicine, where I taught introductory health informatics to over 4,000 medical students, and multiple semesters as a graduate teaching assistant at Penn State across courses including IST 140: Introduction to Application Development and IST 402: Applied Generative AI, I bring three evidence-based principles to every course I teach: active learning through hands-on experience, adaptive instruction responsive to diverse learners, and authentic assessment rooted in real-world practice.

Active Learning Through Hands-On Experience

Lectures alone do not produce the engineers, researchers, and informed citizens that AI now demands. In my redesign of IST 140, I restructured lab sessions around an interactive learning cycle: a 10-minute concept introduction, live coding demonstrations where I deliberately introduce bugs to model debugging as a core skill, paired exercises of progressive difficulty, and peer-led code reviews where students explain solutions to the class. Engagement rose from 45% to 85% across semesters, with a 40% increase in students reporting confidence in their programming abilities. The pedagogy is grounded in cognitive science: immediate application strengthens memory consolidation, peer collaboration exposes students to alternative problem-solving strategies, and public review develops both technical communication and metacognitive awareness.

Adaptive Instruction

Students arrive with different cognitive approaches, languages, and prior preparation. As someone who navigated academic pathways from Grenada to an R1 doctoral program while managing a learning disability, I am committed to inclusive pedagogy that recognizes diverse learning styles. In COMP 420 (Database Systems), I present complex concepts like normalization through three parallel modalities: visual entity-relationship diagrams with color-coded relationships for visual learners, physical index-card exercises for kinesthetic learners, and formal mathematical notation for analytical learners. Students who initially struggled with one modality showed an average 23 percentage-point improvement when offered alternative presentations, while students who already understood the concept reported that alternative explanations deepened their understanding. Continuous formative assessment through in-class polls allows me to dynamically adjust pacing and depth in real time.

Authentic Assessment

Assignments in my courses mirror professional practice. In IST 402 (Applied Generative AI), students complete a semester-long project structured across four phases that mirror an industry workflow:

Problem Definition and Dataset Curation. Students identify a real-world problem domain (past examples include mental health chatbots, multilingual customer service, and accessibility tools) and curate appropriate datasets via Hugging Face.
Prompt Engineering and Baseline Development. Students implement zero-shot and few-shot baselines, document experiments in shared Jupyter notebooks, and present preliminary results in lightning talks.
Advanced Techniques. Students explore multimodal LLMs, custom embedding models, and fine-tuned classification and generation models, conducting ablation studies to understand which components drive performance.
Deployment and Documentation. Students deploy applications via Streamlit, HuggingFace Spaces, or agentic frameworks, write comprehensive documentation, and present in a poster session attended by faculty, peers, and invited industry partners.

Several past projects have evolved into deployed applications and submitted research papers. As one student wrote: “This felt like building a real product, not just completing an assignment.”

The T·I·C Framework

My approach to AI in the classroom is captured in the T·I·C framework, which I developed and have presented in faculty workshops on responsible generative AI:

Transparency. Be explicit, in writing, about when, why, and how AI is or is not permitted. Vague policies create unequal outcomes.
Intentionality. Use AI because it serves a specific pedagogical goal, not because it is convenient. The internal test: what would be lost if students did this without AI?
Criticality. Treat AI as a starting point, not an authority. Surface biases. Question hallucinations. Acknowledge what disciplinary expertise contributes that the model cannot.

The framework is short by design, portable across disciplines, and durable across model generations.

Research-Informed Pedagogy

My research on multilingual AI safety enriches my teaching directly. Students analyze real datasets from my work on harmful content detection across 70+ languages, examine where AI systems fail, and build defensive mechanisms using examples from my Fighting Fire with Fire (F3) framework. They experience firsthand the cat-and-mouse dynamics between attackers and defenders, transforming abstract concepts into tangible challenges with clear societal stakes.

Course Portfolio at CU Boulder

Through SEAL and the Department of Information Science, I am committed to teaching across the undergraduate-to-doctoral pipeline:

Foundations: Applied Generative AI; Natural Language Processing; Responsible AI.
Advanced topics: Multilingual NLP; AI Safety and Adversarial Machine Learning; Fair Machine Learning; AI for Social Good.
Doctoral seminars: Trustworthy AI Across Linguistic Diversity; Research Methods for AI Equity.

Each course integrates research directly into instruction. The classroom and the lab are not separate domains; they are continuous.

Vision

Looking forward, I aim to develop courses at the intersection of AI and social impact, integrate community-engaged learning through partnerships with non-governmental organizations and newsroom collaborators, and contribute to pedagogical scholarship on inclusive practices in computing education. Just as my research seeks to democratize AI capabilities across linguistic boundaries, my teaching seeks to democratize computing education across different backgrounds, abilities, and learning styles.

What Students Take Away

By the end of a course with me, students should be able to do three things they could not do before: build a working AI system, evaluate where and why it fails, and articulate the human stakes of those failures. The first is technical. The second is methodological. The third is moral. All three are required for the field, and all three are within reach of any student willing to engage seriously with the work.

Last updated: May 2026 · ·