Learning AI Without a STEM Background: Mixed-Methods Evidence from a Diverse, Mixed-Cohort AIED Program

Technical Summary: Learning AI Without a STEM Background

Problem Statement
Despite the increasing integration of artificial intelligence (AI) into governance, workforce, and civic life, access to AI education remains inequitable. Current initiatives predominantly target STEM-prepared students, rely heavily on coding proficiency, or focus on early-career learners. This exclusion systematically prevents non-STEM undergraduates and adult professionals from engaging with AI in policy, public-interest, and workforce contexts. Consequently, existing models often prioritize technical implementation over ethical judgment, reinforcing disparities in technological participation and limiting the ability of diverse stakeholders to critically evaluate automated systems or adapt to data-driven labor markets.

Methodology
This study employs a mixed-methods evaluation design to analyze the "Data Crossings" program, a National Science Foundation (NSF)-funded, year-long AI education initiative. The program utilizes a mixed-cohort model, intentionally integrating non-STEM undergraduates (from humanities, social sciences, arts, and business) with adult learners (working professionals and career-transitioning individuals).

• Participants: The first cohort consisted of 19 selected participants (12 undergraduates, 7 adults) drawn from over 200 applicants (9.45% acceptance rate). Selection prioritized motivation for non-technical AI engagement and readiness for mixed-cohort learning over technical proficiency; applicants with formal STEM computing training were excluded.
• Data Sources: Data collection includes participant surveys (administered at entry and throughout the program), open-ended reflections, instructor and teaching fellow reports, and application narratives.
• Analytic Approach:
  • Quantitative: Descriptive analysis of Likert-scale survey items measuring self-reported confidence, perceived relevance, and academic agency. A transition matrix was used to calculate expected value shifts in comfort levels regarding AI technology.
  • Qualitative: A hybrid computational-qualitative Natural Language Processing (NLP) approach was applied to application narratives. This involved unsupervised topic modeling (e.g., BERTopic) combined with human-in-the-loop coding to identify latent themes related to motivations and ethical orientations.
  • Integration: A convergent mixed-methods strategy was used to triangulate learner self-reports with educator perspectives.

Key Contributions

1. Mixed-Cohort Design: The paper introduces an AIED model that integrates learners across disciplines, ages, and career stages, challenging the norm of homogeneous, age-segregated AI classrooms.
2. Longitudinal Evidence: It provides mixed-methods evidence from a sustained, year-long program rather than a short-term intervention, offering a more robust view of learning trajectories.
3. Reframed Learning Outcomes: The study posits that ethical judgment and academic agency should be treated as core learning outcomes alongside AI literacy, rather than viewing technical mastery as the sole prerequisite for engagement.

Results

• Motivations: Analysis of application narratives (n=201) revealed that the dominant motivation for participation was the ethical and responsible use of AI (68%), followed by applying AI for social good (65%). Applicants consistently framed AI as a tool for interpretation and decision-making rather than model building, often citing traditional coding prerequisites as structural barriers.
• Confidence Gains: Quantitative analysis of survey data indicated significant gains in self-reported confidence. Participants entering with low to moderate comfort levels showed the largest upward movement. The overall expected value shift in comfort was positive (+0.90), with no participants transitioning to extreme discomfort.
• Ethical Reasoning: Qualitative data revealed a consistent emphasis on responsibility, judgment, and contextual reasoning. Participants demonstrated the ability to articulate ethical tradeoffs and question assumptions even in the absence of technical certainty.
• Pedagogical Dynamics: Instructors and near-peer mentors reported high engagement, particularly in dialogic and scenario-based learning activities. The mixed-cohort structure functioned as an epistemic asset, where adult learners provided domain expertise and professional judgment, while undergraduates contributed reflective practices, normalizing uncertainty and broadening perspectives.

Significance and Claims
The paper argues that human-centered instructional supports—specifically ethical scaffolding, mentorship, and structured discussion—are essential for equitable AI education in heterogeneous populations. It claims that:

• Ethical judgment is learnable and observable even without technical mastery, challenging models that postpone ethics until after technical competence is achieved.
• Adult learners act as epistemic contributors rather than remedial participants, elevating the intellectual and ethical depth of the learning environment.
• Facilitation is a core equity mechanism, creating space for questioning and reasoning without penalizing technical uncertainty, unlike self-paced or coding-first models.

The authors conclude that expanding AI education requires rethinking design assumptions to include mixed-cohort, ethics-centered models that prepare diverse learners—not just developers, but future decision-makers—to engage responsibly with AI. The study acknowledges limitations regarding sample size and self-selection, noting that comprehensive post-program outcomes are pending longitudinal analysis.