Speakers

You may see all accepted, full abstracts at this OpenReview link.

• Sarah Wilson

Auditing the Educational Pipeline: Mechanistic Interpretability as a Tool for Gender Justice in STEM

Despite decades of institutional effort, women remain severely underrepresented in Artificial Intelligence and Machine Learning. As universities increasingly deploy algorithmic systems for admissions, grading, and resource allocation, these tools risk entrenching historical exclusions in new computational form. We argue that gender justice in STEM education requires moving beyond demographic parity metrics toward structural accountability—understanding why a model produces disparate outcomes, not merely that it does. To this end, we propose a framework applying mechanistic interpretability to audit educational ML models, enabling researchers to locate and intervene on “gendered circuits”—internal representations that propagate gender-correlated disadvantage. Crucially, we distinguish between equity (equalizing outcomes) and justice (removing barriers): our approach targets the latter by diagnosing the computational mechanisms through which prior institutional exclusion re-enters algorithmic decision-making. We discuss practical deployment considerations and policy implications for institutions using ML in student-facing contexts.

• Poojak Patel, Raj Patel, Solomone Somani, Maneth Perera

Participation Debt: When Fair Machine Learning Still Fails to Remove Barriers

Fair machine learning usually asks whether a model treats people equally. We argue that equal treatment at the moment of a decision can still leave people unequally able to act on it. We introduce Participation Debt, the gap between the opportunities a person is formally granted and the opportunities that person can realistically reach. Two people can receive the same recommendation, score, or eligibility decision, yet one may be unable to turn it into a real outcome because of barriers such as cost, distance, paperwork, language, or unfamiliarity with the process. Because standard fairness metrics evaluate the distribution of model outputs and stop there, they can be fully satisfied while Participation Debt stays large or even grows. We define the concept, explain why current evaluations miss it, and propose a simple lens for how reliably a decision becomes a realized benefit.

• Rob Sneiderman

Lowering the Barrier to Rigorous Mathematics: Adaptive AI-Assisted Practice as an Equity Tool

Access to rigorous mathematics is gated less by materials than by support: the structured practice, timely feedback, and personalized pacing that a private tutor provides. That support is unevenly distributed, and the gap falls hardest on under-resourced learners. Equal access to a textbook does not close it; what privileged learners actually receive is adaptive, individualized scaffolding. A justice-oriented response asks how to give every learner that scaffolding, not how to equalize a single metric. A deployed learning system targets this barrier directly: adaptive spaced-repetition practice (an FSRS-style scheduler over a structured mathematics curriculum) combined with AI-assisted explanation, so that personalized, mastery-paced practice, historically the privilege of one-on-one tutoring, becomes available at near-zero marginal cost. The design goal is capability: the scaffolding that lets a learner participate in mathematics, grounded in how people actually learn (forgetting curves, retrieval practice, mastery thresholds) rather than uniform one-size delivery. Near-zero marginal cost is not the same as access: devices, connectivity, language, and time sit upstream of the scaffolding, and AI-assisted explanation can be uneven, so the honest version of the claim is that the system removes one binding barrier while these conditions and the reliability of the feedback remain part of the work. The framing is justice rather than fairness. The aim is not a balanced accuracy number across groups but the removal of a concrete participation barrier, the cost and scarcity of expert practice support, for the learners for whom it is most binding. The limits stay in view: this is a deployed system and a design argument, not yet a controlled efficacy study. A community-grounded evaluation would define outcomes with the learners served rather than imposing metrics on them, and building that evaluation is what comes next.

• Maneth Perera, Solomone Somani, Raj Patel, Poojak Patel

Characterizing Proxy Reliability in Health-Relevant Machine Learning with a Dual Axis Framework

Machine learning systems in healthcare and social services routinely rely on proxy variables for attributes that are expensive, sensitive, or impossible to observe directly (disability from utilization patterns, race from name and geography, mental health risk from social media). Fairness research evaluates such systems mainly through parity metrics on a proxy's output, leaving underexamined whether the proxy itself reliably stands in for the attribute it represents, and whether that reliability is even-handed across populations. We introduce a two-axis framework for proxy reliability (distance from ground truth and characterization quality), independent of any fairness metric, tested on NHANES (National Health and Nutrition Examination Survey) data by withholding a documented disability label and predicting it from indirect correlates only, so error can be measured against ground truth. Across pooled multi-cycle data, cross-algorithm checks, bootstrapped intervals, and a temporal generalization test, we find income/race sensitivity gaps concentrated in one disability domain (mobility), not spread evenly. Identical predicted probabilities also produce diverging harm patterns depending on the decision rule applied, with uses of the same model harming nearly inverse populations. Proxy reliability, we argue, is a property of the variable as deployed, not fixed, with structure a parity-only audit will not surface.

• Syed-Amad Hussain, Eric Fosler-Lussier, Emre Sezgin

Representational Justice in Synthetic Healthcare Dialogue: Grounding Personas in How Communities Describe Themselves 

Conversational agents for healthcare are increasingly built and tested on synthetic patient dialogue, because collecting real conversations raises privacy, consent, and inclusivity concerns. These concerns fall hardest on marginalized communities, where surveillance risk is greatest and real transcripts scarcest. A common way to make a synthetic persona representative is to attach coarse demographic labels such as "Race: Hispanic" or "Income: benefit recipient," which describe a community from the outside and may yield flattened, out-group portrayals [1]. Parity-based fairness metrics assume the data already represents each group faithfully [2], yet severe data skews distort inputs and outcomes for marginalized populations [3]; a justice lens instead asks what is needed to represent a community well [4].

• Ebtesam Al Haque, Evan Shieh, Thema Monroe-White

Emancipatory Artificial Intelligence: Recognizing the Past to Reimagine Possible Futures

AI systems routinely encode the racialized hierarchies of the institutions and methods that produced them, yet education research largely treats this as a technical defect to be patched rather than a history to be confronted. We introduce Emancipatory Artificial Intelligence (EAI), a framework pairing counter-archival analysis with four tenets---recognition, recollection & refusal, repair, reflection & reimagination---to cultivate learner agency and critical inquiry of AI systems.

• Michael Cacioli

Lost in Transcription: Measuring and Mitigating Dialect Substitution Bias

Ask a modern speech recognizer to transcribe a Neapolitan speaker saying *aggio* and it will likely return *ho*, the Standard Italian counterpart. Under word error rate, this constitutes a single substitution in a long list. To the Neapolitan speaker, it is something else. The system has heard their language and written down something else more powerful. This talk argues that this harm pattern is analogous to one found in the fairness literature; it is simply not visible because the metric used was not designed to see it. WER assumes errors are symmetric noise, but in a low-resource language existing alongside a dominant standard, the noise has direction, and for Neapolitan, that direction is assimilation. Fluent, powerful misrepresentation is subtle and hard to resist. We propose a two-stage program to make this problem visible, first building the Neapolitan Spoken Corpus, the first structured dataset for Neapolitan ASR, comprising 141 clips from a single native speaker (initially), which grew to 591 clips from 4 native speakers spanning literary and modern usage. With the Neapolitan Spoken Corpus in hand, a state-of-the-art multilingual ASR model, Whisper (which we operate in its Italian mode, as no Neapolitan mode exists), achieves a mean WER of 86.94% on the initial corpus. Our second stage is to create the Italianization Error Rate (IER), which detects dialect-specific words replaced by their Standard Italian variants. 

• Sarah M Brown

Relational Realism and Justice in Machine Learning

Systems reflect the values of the context in which they were built. Context is broad and complex. Much attention has been paid to the incentives of institutions, the contents of the data, and the awareness of the individuals building systems. An under-addressed aspect of this context is the philosophical underpinnings of the discipline. Underlying every piece of knowledge production are distinct assumptions about the nature of knowledge (epistemic) and  the nature of reality (ontological) assumptions. These are rarely addressed in machine learning, but that lack of attention to philosophy does not remove the impact of philosophy on realized machine learning systems. In this talk, I will first argue that the dominant, implicit, philosophy of science present in ML literature creates the challenges to justice present a philosophy of science where justice more readily emerges.

• Ziyue Yang

From Handwritten Gongche to Playable Scores: End-to-End ML for Kunqu Heritage

Kunqu opera is a roughly 600-year-old vocal tradition recognized by UNESCO as intangible cultural heritage. Its historical scores use Gongche notation: a handwritten pitch script interleaved with sung lyric calligraphy. As contemporary education and performance increasingly rely on Western staff notation, reading these sources requires scarce specialized knowledge, and manual transcription takes hours per page. Scans preserve their appearance but not their music as searchable, editable, or audible data. The community challenge is therefore concrete: notation-centered infrastructure can preserve an image while leaving the repertoire difficult to learn, perform, and study. We present an end-to-end optical music recognition pipeline that converts a scanned Gongche page into editable MusicXML 4.0 and playable MIDI with aligned Chinese lyrics. Two YOLOv8m detectors locate 28 pitch/key classes and lyric-character boxes. Qwen3-VL recognizes each open-vocabulary lyric crop; a geometric, column-based aligner pairs lyrics with pitches; and an emitter produces the score. Training combines the public LGRC2024 benchmark with 51 hand-labeled lyric pages assembled through active learning. Removing rotations and flips that violate Gongche reading order raises pitch mAP@0.5 from the published 74.26% YOLOv8m baseline to approximately 81% on 638 validation images. The lyric detector reaches 99.3% mAP@0.5. Across 413 characters on 11 held-out pages, end-to-end accuracy is 80%; detection misses on dense layouts cause most errors, while recognized crops are approximately 95-98% correct. A 33-character page takes about one minute and costs roughly USD 0.03, versus hours of manual work, moving corpus-scale transcription into a practical range with human correction. The justice contribution is barrier removal, not a claim that a metric or model can preserve culture by itself. The scan remains the source of record, overlays preserve provenance, generated MusicXML is editable, and the 80% output is an assistive draft. Evaluation covers only 11 pages, rhythm is not yet reconstructed, and the system was not community co-designed. The oral talk uses this case to show how context changes ML design: use transformations valid for the script, evaluate the complete artifact rather than convenient component accuracy, and keep automation assistive rather than concluding. Participatory work with Kunqu practitioners, students, educators, and archivists must determine priority uses, acceptable errors, and when automation should not be used. Anonymous code, data, and reproduction instructions: https://anonymous.4open.science/r/gongche-staff-omr-A78F

• Staford S. Francois

Disease Disparity in Underserved Populations: Identifying Barriers to Health Equity with Heterogeneous Graph Network

Health disparities in the United States health care system continues to be a persistent concern, as individuals from marginalized communities consistently experience poorer health outcomes. A promising strategy that has been applied to address health disparities is machine learning, which enables computer systems to identify patterns to reveal relationships between risk factors and disease outcomes. However, machine learning tools studying risk factors have often been developed from data and objectives that fail to accurately capture social and structural aspects that contribute to health disparities of marginalized communities. This dynamic reduces complex structural inequity to simple risk factors and leads to the ineffective application of fairness metrics [6]. Thus, the current framing cannot serve as adequate evidence for communities affected by systemic disadvantage [2, 4]. Models need to be trained to identify the structural causes behind the disparities so intervention may be practiced [3].

• Ebtesam Al Haque, Evan Shieh, Thema Monroe-White | Does Cultural Fine-Tuning Mitigate Laissez-Faire Harms?

• Sushma Upadhayay, Darakhshan Mir, Holiness Kerandi, Aleena Sultan, Vanessa Massaro | Using Machine Learning to Examine Justice in a Carceral Algorithm Used in Pennsylvania Prisons: The Case of STRONG-R

• Ademola Adeniyi, Thema Monroe-White | Beyond Adoption: AI Literacy for Technical and Sociotechnical Understanding

• Margaret Wang | Modeling the Effect of Life Course Alcohol Use Trajectories on Cancer Incidence

• Karlis Baisden | When the Algorithm Decides Who Goes Home: Racial Disparity in the Federal PATTERN Risk Tool

• Ritta Neg Mfa, Sarah M Brown, Alina Jade Barnett | Beyond ’This Looks Like That’: Auditing Prototypes for Demographic Bias