AI Education Innovation Fund is sponsored by Wharton AI & Analytics Initiative. This fund supports curricular innovation by providing resources for faculty to incorporate AI into classroom instruction and course materials for both degree and non-degree learners.

Funded Research - Wharton AI & Analytics Initiative Researchers » Funded Research AI Education Innovation Fund The AI Education Innovation Fund is designed to support curricular innovation by providing resources for faculty to augment, adapt, and reimagine how they incorporate AI into classroom instruction and course materials for both degree and non-degree learners.

By offering the necessary resources, the fund aims to accelerate the creation and enhancement of AI-related coursework and educational materials, ensuring that our offerings remain at the forefront of technological advancements. Examples of support include, but are not limited to: funding for developing new AI courses; enhancing AI learning tools; and resources for updating and expanding existing course materials.

The AI Research Fund is designed to provide faculty with essential resources to explore the intersection of AI advancements with modern business models, industries, and global economies. These funds support the development of new methods, the application of existing technologies, and investigations into the impacts of AI and analytics.

Faculty can apply for funding for a specific project or general research support to advance their work. Examples of support include, but are not limited to: funding for data, computing, and dataset acquisition; financial assistance for research assistants; and matching dollars for Post-Docs, Pre-Docs, and Doctoral student support.

Data Science and Analytics Fund The Data Science and Analytics Fund is designed to provide faculty with resources for exploring innovative methods for using data science and analytics to transform modern business and society. These funds support the development of new methods, the application of existing technologies, and investigations into the impacts of AI and analytics.

Faculty can apply for funding for a specific project or general research support to advance their work. Examples of support include, but are not limited to: funding for data, computing, and dataset acquisition; financial assistance for research assistants; and matching dollars for Post-Docs, Pre-Docs, and Doctoral student support. Can We Still Detect AI-generated Content?

Do Our Attitudes Towards Cryptocurrencies Affect Pricing?

Wharton Professor’s SCEPTRE Tool Helps Link Genetics to Disease Risk Nudge Cartography: Building a map to navigate behavioral research Measuring Vehicle Exhaust Standards Over Time Wharton Researchers Build Algorithm to Help Fight Sex Trafficking Wharton Research Shows Revitalizing Vacant Lots Pays Dividends Data Analytics for Economic Efficiency in Energy Policy The Dark Data of Policing The AI Research and Data Science & Analytics Funds are available through the Research Common Application, with calls for proposals held twice a year in June and December .

The AI Education Innovation Fund is offered through a separate application process on a rolling basis. Funded Projects Spring 2026 4th Annual Penn-Georgetown Digital Ethics Workshop Brian Berkey, Associate Professor, Legal Studies and Business Ethics Facing Default?

AI-Extracted Facial Features and Credit Outcomes Marius Guenzel, Assistant Professor, Finance This project studies whether AI-extracted facial features contain economically meaningful “soft” information that predicts credit outcomes beyond traditional credit bureau measures. Using a novel dataset linking LinkedIn profile images, labor market histories, demographics, and credit bureau records for nearly 1.

5 million individuals, we apply state-of-the-art computer vision models to generate facial embeddings and evaluate their out-of-sample predictive power for delinquency. Facial embeddings significantly predict default risk and add incremental value beyond credit scores, income, and education, with particularly strong gains for thin-file borrowers.

Planned extensions expand credit outcomes and develop survey-validated personality measures to shed light on underlying mechanisms. The project informs both the economic role of soft information in credit markets and policy debates surrounding AI-based screening tools.

Workshop on Responsible AI and Data Science for Insurance Pricing Giles Hooker, Professor, Statistics and Data Science The rapid adoption of AI and data science in insurance underwriting and pricing has created a pressing challenge: how to achieve highly accurate risk prediction while ensuring fairness, transparency, and public trust —particularly as climate-related disasters intensify risk heterogeneity across regions and populations.

As insurers increasingly rely on complex models and sensitive data, concerns around bias, accountability, and regulatory compliance have become central to both industry practice and policy debates.

This workshop, co-organized by Professor Giles Hooker (Wharton) and Associate Professor Fei Huang (UNSW), will convene leading researchers, industry practitioners, and regulators to examine frontier methods and governance frameworks for responsible insurance pricing.

The event will explore how predictive accuracy, fairness, and market efficiency can be balanced in AI-driven pricing models, including in high-risk and climate-affected markets, and will foster sustained industry–academia collaboration to support the responsible use of data science in insurance.

Kartik Hosanagar, Professor, Operations, Information, and Decisions As LLMs increasingly mediate consumer decisions, organizations need methods to optimize content visibility while maintaining authenticity. This project develops agentic optimization techniques that identify high-impact content modifications through gradient-guided search across open-weight models.

While optimization mechanisms are becoming well understood, actual consumer interactions with AI-optimized content remain largely unexplored. Perceptions of Fairness in Algorithmic Decision-Making Bethany Hsiao, PhD Student Duncan Watts, Professor, Operations, Information, and Decisions In algorithmic decision-making, it is often mathematically impossible to satisfy competing definitions of fairness simultaneously.

This impossibility makes assessing stakeholders’ fairness preferences a challenging but crucial exercise. To address this, the researchers developed an interactive tool to elicit fairness preferences by projecting the complex decision space into a simple one-dimensional choice: setting risk thresholds.

This design allows participants to directly manipulate algorithm parameters and visualize the resulting tradeoffs between competing fairness metrics. Participants’ decisions enable the researchers to study (1) what Pareto-optimal settings are perceived as fair and (2) whether the design of the elicitation process itself can fundamentally shape how fairness is perceived.

Generative AI and Startup Scaling: Evidence on Reducing Supply- and Demand-Side Constraints J. Daniel Kim, Assistant Professor, Management Startups play a central role in growth and job creation, yet scaling remains a fundamental challenge. On the supply side, young firms struggle to attract and organize high-quality talent, while on the demand side they face severe uncertainty in discovering and sustaining customers.

This project examines whether generative AI reshapes these constraints to scaling. We argue that generative AI enables startups to scale differently by reducing reliance on headcount growth and improving customer discovery and value delivery.

Using a novel panel dataset of venture-backed software startups, we integrate data on generative AI adoption, employee-level hiring and workforce composition, venture capital financing, and real-time website traffic as a proxy for customer demand. Exploiting variation before and after the release of ChatGPT-4, we estimate the effects of generative AI adoption on startups’ employment structure, market traction, and overall productivity.

Whom, When and What to Nudge: Dynamic Personalization via High-Dimensional State-Switching Bandits Eric Bradlow, Vice Dean, Wharton AI & Analytics Initiative Peter Fader, Professor, Marketing E-commerce platforms increasingly use real-time nudges like discounts, urgency cues, and personalized offers to drive conversions.

Yet optimizing these nudges is challenging: user context is high dimensional, and consumers move through unobserved behavioral states like casual browsing, product evaluation, and purchase intent, each responding differently to interventions. We develop a state-switching bilinear bandit model that learns whom to target, when, and with what incentive.

The model combines a Hidden Markov Model to infer latent shopping states, a bilinear structure to capture interactions between user features and nudge content, a spike-and-slab prior for automatic variable selection and a complexity prior to manage the overall size of the parameter space.

We embed this within a forward-looking Thompson Sampling policy that accounts for discounted memory — ensuring that promotions don’t backfire by raising future expectations. Empirically, we validate the approach on synthetic and real datasets.

Our approach offers a scalable solution for dynamic promotion design, outperforms benchmarks that ignore either behavioral dynamics or context sparsity, and helps platform managers maximize long-term profit from in-session personalization.

Simulating CEO–Board Dynamics with Large Language Model (LLM) Agents Lynn Wu, Associate Professor, Operations, Information, and Decisions CEO–board dynamics shape some of the most consequential decisions of a firm. Whether a company discloses bad news, takes on risk, or pivots strategy often depends on the interplay between the CEO and the board.

Many empirical studies have documented relationships between CEO and board characteristics and firm outcomes. However, the underlying mechanisms remain underexplored, largely because large-scale data on CEO–board interactions are scarce.

This project seeks to develop an LLM-powered multi-agent simulation framework to examine how CEO–board interactions shape corporate strategy and to identify the mechanisms through which those interactions affect firm outcomes.

Balancing Cognitive Surrender and Offloading: Guardrails for Calibrated AI Use in Consumer Decisions Gideon Nave, Associate Professor, Marketing People increasingly rely on generative AI to answer questions and make decisions.

Shaw & Nave (2026) propose Tri‑System Theory of Cognition, extending dual‑process accounts by adding System 3 (artificial cognition) and documenting cognitive surrender: the tendency to adopt AI outputs with minimal scrutiny. This reliance can be adaptive when AI is correct, but harmful when AI is wrong.

This project extends Tri‑System Theory into consumer and business decision contexts, and tests when and how people shift between cognitive surrender (uncritical adoption) and cognitive offloading (strategic delegation with monitoring).

Across preregistered experiments, we will evaluate whether AI interface “guardrails”, including uncertainty signaling, lightweight verification prompts, and social accountability cues, can promote calibrated AI use: preserving performance gains when AI is reliable while increasing selective override when AI is unreliable.

The goal is to generate actionable guidance for consumers, marketers and practitioners who want to capture the benefits of AI assistance without eroding deliberative judgment or accountability in contexts where errors are costly.

Editing Digital Twins for Behavioral Fidelity Stefano Puntoni, Professor, Marketing Large language models (LLMs) have enabled “digital twins”—AI agents designed to simulate how specific consumers would respond to marketing questions and interventions.

While current digital twins can match average survey answers reasonably well, they often fail on what matters most for marketing: reproducing how people change their behavior under interventions such as anchors, sunk-cost cues, framing, and pricing changes. Instead, they behave too “assistant-like,” correcting biases and producing overly consistent, normative responses.

This project develops a new framework for intervention-consistent digital twins: models that mimic human responses not only on average, but under experimentally manipulated contexts, including realistic heterogeneity and noise.

Using the Twin-2K-500 benchmark dataset, we will combine conditional behavioral editing (activation/representation steering applied only in relevant experimental contexts), person-specific susceptibility to behavioral effects, and variance calibration grounded in human test–retest reliability. For pricing tasks, we will add lightweight economic constraints to ensure plausible demand responses.

The result will be a scalable method to build digital twins that are behaviorally credible and better suited for marketing research and decision support.

The Effects of Taxing Prescription Opioids on Drug Utilization and Provider Behavior Jackson Reimer, PhD Student Marissa King, Professor, Health Care Management The United States overdose epidemic has reached new heights with approximately 100,000 people dying from a drug overdose three of the last four years (Ahmad et al. , 2025).

Common policy solutions, such as prescription drug monitoring programs, have put the onus on prescribers to limit inappropriate opioid access. Market-oriented solutions to curb consumption, such as a tax, have been proposed by local, state, and federal legislators, though they have been historically less common.

This project fills this knowledge gap by examining the impact of a novel excise tax of prescription opioids for retail pharmacies in New York (NY) on opioid utilization, pharmacy growth, and population health. As of 2019, opioids priced less than $0. 50 per unit are taxed $0.

0025 per morphine milligram equivalent (MME) and more expensive drugs are taxed $0. 015 per MME. Although opioids are similar to other addictive goods, the welfare implications of such a tax on consumption are theoretically ambiguous.

First, it remains unclear ex-ante whether an opioid tax would effectively reduce consumption, as the out-of-pocket cost of prescription drugs is often limited to a fixed coinsurance payment. In effect, consumers are partially insulated from direct price increases. Second, the tax might reduce consumer wellbeing if it impacts both clinically appropriate and inappropriate prescribing.

Finally, even if inappropriate prescribing is reduced, consumers might substitute away from retail pharmacies to more expensive care settings such as outpatient clinics. Therefore, the impact of opioid taxation remains empirically ambiguous yet critical for policymakers considering similar legislation’s effects on the health care sector.

Teaching Statistics and Data Science from First Principles: A Data-Oriented, AI-Enabled Course Joseph Rudoler, PhD Student Abraham Wyner, Professor, Statistics and Data Science This course provides an introduction to statistics with an emphasis on building intuition about uncertainty in data analysis and decision-making.

The course is structured around thinking critically about data generating processes (i.e. the random, real-world phenomena that produce data) and how to model them computationally.

An essential aspect of this course is the assumption that students will use modern Large Language Models (LLMs) to accelerate their learning — this reduces the barrier to entry for basic data science programming and allows students to remain focused on important concepts in both data management and probabilistic thinking, instead of being bogged down by implementation details.

Tech-Washing: Strategic Technology Positioning and the Narrative-Innovation Gap Fernando Stein, PhD Student Winston Dou, Associate Professor, Finance Markets assign different valuations to different types of technology. AI firms trade at premiums to software firms, software firms to hardware firms, hardware firms to traditional industries.

These valuation hierarchies create incentives for firms to reposition themselves toward higher-valued technology segments through disclosure, regardless of whether their innovation portfolios support the repositioning. I term this practice “”tech-washing.

”” I document and measure tech-washing by constructing the first firm-year measure of the wedge between what companies say (technology narratives in SEC filings and earnings calls) and what they do (patented innovations). My measure combines natural language processing of corporate disclosure and news with patent portfolio analysis to identify firms whose technology positioning outpaces their technology substance.

The Effect of Biometric Payment Methods on Consumer Behavior Wendy De La Rosa, Assistant Professor, Marketing AI has shepherded new payment methods for consumers, such as biometric payment methods like facial recognition or palm vein scanning.

These methods are becoming increasingly prevalent, with large retail stores like WholeFoods adopting palm vein scanning across all of their US stores and restaurants like Steak ’n Shake rolling out facial recognition payment nationwide. Yet, despite the prevalence of these new payment methods, little is known about how they influence consumers’ behavior at the point of sale (POS), including spending, tipping, and donating.

Beyond deciding whether and how much to spend, consumers are often asked to make additional decisions at the POS, such as how much to tip or whether to donate to a charity. In this work, we investigate how emerging payment technologies influence these downstream financial behaviors.

Generative AI, Ideological Polarization, and the Transformation of Online Political Discourse Lynn Wu, Associate Professor, Operations, Information, and Decisions This project investigates how generative AI reshapes political discourse by focusing on a novel and counterintuitive pattern: initial evidence suggests that the adoption of large language models increases ideological polarization while simultaneously improving affective tone and civility.

Rather than escalating hostility, AI-assisted political expression appears to amplify ideological alignment in a calmer, less toxic register. The core objective of the project is to uncover the mechanisms driving this divergence.

By analyzing interaction-level dynamics—such as reply alignment, narrative reinforcement, and algorithmic sycophancy—the research seeks to explain how generative AI can polarize what people express politically while moderating how those views are communicated.

Use and Development of AI Companion Tools for Populations Under Stress Pinar Yildirim, Associate Professor, Marketing Rapid urbanization, educational transitions, and labor-market mobility increasingly expose individuals to acute social isolation and psychological stress, particularly in low-resource settings where access to mental-health services is limited.

Advances in generative artificial intelligence (GenAI) have enabled the creation of AI companions—always-available conversational agents capable of providing empathic conversations, reflection, and coping support. While these tools are widely adopted, there is little causal evidence on their welfare effects, their interaction with existing digital behaviors, or their implications for real-world social connection.

By providing the first causal evidence on AI companionship in vulnerable populations in the Global South, this project advances three literatures: technology-mediated mental-health interventions, the economics of digital attention and substitution, and the study of migration and social integration.

The findings will inform policymakers, universities, employers, and non-governmental organizations about when and how AI companions can be deployed as privacy-preserving complements to overstretched mental-health infrastructure.

AI as a Strategic Lens on Entrepreneurial Adaptation and Scaling Ndubuisi Ugquanyi, PhD Student Entrepreneurial firms operate in environments characterized by rapid technological change, shifting demand, and heightened uncertainty.

While prior research has identified factors associated with startup growth and survival, much of this work relies on static, ex post measures and offers limited insight into how firms adapt strategically in real time.

This project proposes artificial intelligence as a strategic lens for studying entrepreneurial adaptation and scaling, examining whether AI-based methods can detect early signals of organizational change that predict subsequent growth.

Leveraging large-scale, longitudinal data on startups’ digital footprints, including website content, hiring patterns, and organizational roles, this study develops machine-learning models to capture how firms update their strategies, capabilities, and market positioning over time.

By analyzing patterns of textual evolution and structural change, the project seeks to identify whether adaptive behaviors can be measured ex ante and linked to future scaling outcomes such as sustained growth, continued funding and favorable exits. The project contributes to research on entrepreneurship and strategy.

Substantively, it advances understanding of how startup firms respond to changing environments and which forms of adaptation are most strongly associated with scalable and sustainable growth. Methodologically, it demonstrates how AI can be used to extract strategic signals at scale.

More broadly, the findings have implications for investors, policymakers, and founders seeking to identify and support high-potential ventures before performance outcomes are fully realized.

Funded Projects Summer 2025 Accelerating Diffusion Generative Models Yuxin Chen, Associate Professor, Statistics and Data Science Diffusion models have emerged as a powerful class of tools for generating realistic synthetic data, enabling breakthroughs in image creation, text generation, decision making, and more.

However, despite their impressive output quality, these models are notoriously slow, requiring extensive computation to produce each data sample. This project aims to bridge the gap between theory and practice by designing faster, more efficient ways to generate data from these models—without compromising quality.

We focus on methods that do not require extensive retraining of the model, allowing for broad applicability, especially with large-scale pre-trained systems. Our work will also explore ways to adaptively speed up generation when the data has hidden structure or when resources are available for parallel processing.

By improving the reliability and speed of diffusion models, this research lays the groundwork for real-time generative AI applications in business analytics and beyond.

In Search of Distress Risk with Textual Data and Machine Learning Winston Dou, Assistant Professor, Finance In this research project, we investigate whether forward-looking information from corporate disclosures and external professional forecasts can enhance the understanding and prediction of corporate financial distress risk.

We begin by extracting nearly all economically relevant sentences from three key sources: analyst reports, annual 10-K filings, and earnings call transcripts for U.S. public firms.

We then build a sentence-level database and apply a Large Language Model API to classify each sentence into economically motivated distress channels, such as liquidity condition, debt leverage, competition pressure, supply chain risk, industry trends, governance concerns, demand shocks, and related frictions.

We merge the resulting distress risk scores with standard accounting and market variables from the existing literature and feed them into a mixed model specification. The resulting text-augmented predictors retain the interpretability of traditional ratio-based measures while leveraging the immediacy and richness of narrative nonstructural data from other sources.

By decomposing default probabilities into identifiable economic channels, our framework not only flags which firms are financially fragile but also explains why they are at risk, information that is critical for regulators, credit analysts, and investors seeking actionable early-warning signals.

Using Housing Photos to Understand Trust in AI Mariaflavia (Nina) Harari, Assistant Professor, Real Estate This study examines how people develop trust in AI-driven resource allocation systems leveraging housing photos as a a context.

Using a hypothetical scholarship allocation scenario that uses AI to rank applicants based on photos of their house, we explore how the photos-based approach can help people understand both AI’s strengths and potential biases. We investigate how different types of AI errors affect trust across participants from varying socio-economic backgrounds.

This work provides evidence-based insights for organizations considering AI adoption in high-stakes resource allocation decisions.

Scaling Behavioral Science with Generative AI: A Virtual Coach for First-Year College Success Barbara Mellers, Professor, Psychology and Marketing Hangchen Dai, Associate Professor This project tests whether generative AI can enhance and scale behavioral science interventions to support first-year students as they transition from high school to college.

The research team will develop and deploy an AI-powered virtual coach—built on large language models (LLMs)—to deliver evidence-based interventions via SMS, targeting students’ cognitive, emotional, and social adjustment. The coach will adapt content in real time and provide interactive feedback to boost engagement.

The success of our project will be evaluated against many other interventions using a randomized controlled trial as part of a mega study led by the Behavior Change for Good Initiative and the Equity Accelerator. We will measure the independent and combined effects of AI coaching and peer support on academic outcomes and well-being.

Rich analytics will shed light on how AI-generated language shapes engagement with behavioral science interventions and contributes to improved academic and social outcomes, with the ultimate goal of building scalable, data-driven tools that leverages psychological and behavioral science insights to support student success.

Attentional Differences in Human- vs. AI-generated Content Shiri Melumad, Assistant Professor The purpose of this research is to test a series of hypotheses about how the growing use of generative AI may be altering how consumers process news and other textual information.

We hypothesize that people hold lay beliefs about the reasoning abilities of algorithms relative to humans, such that the perceived authorship of a text (human- vs. AI-generated) can fundamentally affect which aspects of the text people most attend to.

In particular, we theorize that whereas generative AI tools are believed to be superior to humans in processing and synthesizing more objective information, humans are seen as better at tasks that require more subjective evaluations.

As such, we predict that people will attend more closely to subjective (vs. objective) information when a given text is framed as being created by AI (vs. a human) because they will assume the objective information is already well-researched or vetted, whereas the subjective information needs to be scrutinized more closely.

This, in turn, will determine the types of interpretations and takeaways people draw from the same text, which will affect their downstream decision-making and behaviors in a range of domains. A program of both field and controlled experimental studies will be conducted to test these ideas.

Using LLMs as Stimulus Engines to Improve the Generalizability of Experiments Gideon Nave, Associate Professor Modern behavioral science has made great strides towards improving the replicability of findings, yet generalizability, particularly across stimuli, remains under-addressed.

Many studies still use narrow, hand-picked stimuli chosen for their potential to generate strong effects rather than for representativeness or real-world relevance. In marketing contexts, generalizability is especially critical: consumer research aims to generate insights that marketers can apply across diverse products, categories, and messaging channels.

Without generalizable findings, behavioral research risks limited practical relevance. This project outlines and demonstrates how large language models (LLMs) can be used to systematically generate diverse, theoretically grounded stimulus sets to evaluate and improve the generalizability of experimental research.

As a proof of concept, we create expansive stimulus universes using LLMs and re-examine three robustly replicated consumer behavior effects: overhead aversion in charitable giving (Gneezy et al. , 2014), temporary sharing in digital disclosures (Barasch et al. , 2017), and information compression in decision-making (Klein & O’Brien, 2018).

We implement multiple stimulus sampling designs for each and analyze outcomes using mixed-effects modeling to estimate a distribution of “generalizability effect sizes. ” By bridging LLMs with rigorous experimental methods, this project provides scalable tools for testing and improving generalizability.

It aims to strengthen the external validity of consumer science while establishing new standards for stimulus sampling in behavioral research. Healthy Foods in USA: Demand vs Supply Serguei Netessine, Senior Vice Dean, Professor Healthy foods became popular in the USA in recent decades, with Whole Foods paving the way early on.

Nevertheless, a large part of US population still eats very unhealthy food and suffers from many related health issues. Is this a function of demand or supply? Are there areas in the USA that effectively represent healthy food deserts?

What happens when healthy food becomes available – does supply stimulate demand or is demand independent from supply? Serguei Netessine, Senior Vice Dean, Professor AI tools have been shown to improve various parts of the innovation process: from idea generation to building a business model to pitching.

The goal of this project is to integrate AI copilot into the innovation class to help students generate better ideas, pitch them more efficiently, and have better outcomes in the class.

Deploying Large Language Models for Scalable Behavior Change Roger Saumure, PhD Student Consumers are increasingly turning to generative artificial intelligence (AI) to cultivate social relationships, from casual conversations to meaningful emotional support. This trend has extended to sharing progress toward personal goals—such as fitness journeys or recovery from substance addiction—with large language models (LLMs).

While it is clear that LLMs can partake in casual conversation and provide advice, might they also help consumers achieve their goals? Through both field and laboratory experiments across diverse behavioral contexts, we plan to investigate whether sharing goal progress with LLMs (versus friends) can effectively change behavior.

This research represents the first examination of how generative AI impacts goal pursuit, offering practical insights for managers and policymakers seeking to design scalable, low-cost behavior change interventions.

Funding for Forensic Analytics (ACCT2700) Daniel Taylor, Associate Professor Over the past five years, WAIAI has funded the compute costs associated with Forensic Analytics (ACCT2700), which teaches undergrad students how to conduct pattern recognition and other data analysis techniques for the purposes of detecting fraud.

Assessing the Quality and Pricing of Forest Carbon Offsets Arthur van Benthem, Associate Professor Carbon offset projects focused on reducing deforestation and planting trees have come under intense scrutiny and criticism, given the widespread suspicion that true emissions reductions are far below the amount of offsets issued.

Nevertheless, offsets feature prominently in many companies’ decarbonization plans, and are implemented across many parts of the world, with a particular emphasis on tropical forest regions. We will use rich satellite and project data to develop better econometric estimates of the true carbon savings from forest-based offsets.

We will then link our quality estimates to unique carbon offset transaction pricing data to assess the degree to which the market values offset quality.

Generative AI, Communication, and Worker Performance: A Field Experiment on the Upwork Platform Manav Raj, Assistant Professor We study how access to a generative AI tool may affect worker participation and performance on an international online labor market platform by reducing communication frictions when transacting with international clients who speak a different language.

Working with field partners in India, we will recruit high-skill workers to join the Upwork freelance labor market platform, which allows such workers to compete in a large and vibrant international marketplace. Our project thus speaks to the role of AI in facilitating cross-boarder remote work.

Edgar Dobriban, Associate Professor of Statistics and Data Science Artificial Intelligence (AI) has been making spectacular progress in the past few years, leading to chatbots powered by large language models (e.g., Chat GPT), image generators powered by diffusion models (e.g., Dall-E), among others.

At the same time, AI has proved vulnerable to adversarial attacks, including jailbreaks that coerce the AI to perform unsafe actions (e.g., a chatbot providing instructions how to commit financial fraud, or a robot damaging its environment). While some mitigations have been proposed, to date there are no truly efficient solutions.

Part of the reason is that some unsafe behaviors can be achieved through a composition of individually safe actions (e.g., composing 1. generate check that looks like a real one, 2. sign with someone’s signature, 3.

deposit check; could provide the basic steps of a check fraud). To address this, our project aims to develop novel methodologies explicitly training AI systems to maintain compositional safety awareness. Specifically, we will teach AI models to refuse harmful decompositions, recognize potentially harmful individual requests, and reason about the harmful outcomes of composed actions.

We will generate dedicated, open-source datasets for each task, train bespoke AI models, and evaluate our compositional safety methods against standard training approaches, establishing rigorous, trustworthy benchmarks for safer AI. Alina Song, PhD student, Finance Joao Gomes (Faculty Advisor) This paper challenges the conventional wisdom of viewing data as purely intangible by examining how data center geography affects firm outcomes.

Combining the first-ever dataset linking the near-universe of data center locations with online job postings, I find that establishments of the same firm located in regions with greater data center capacity increase investments in data-intensive (AI) human capital while reducing traditional technology investments. Further, firms systematically reallocate AI roles toward establishments close to data centers.

Predictive Analytics (Fall 2025) and Machine Learning in Business (Spring 2026) Kaihua Ding, Adjunct Faculty, Statistics and Data Science STAT4220 Predictive Analytics for Business1 is an advanced statistics / data science course for Wharton undergraduates. Its sister course, STAT4230 / STAT7230 Machine Learning in Business2, exposes students to modern techniques. Both emphasize hands-on analysis with realistic business datasets.

However, neither engages with large unstructured text data—now central to predictive analytics across industries. Companies routinely combine traditional quantitative models with text analytics: pharmaceutical firms monitor patient comments for sales strategies; retailers merge point-of-sale data with social sentiment; airlines model reviews for online sales; investment banks incorporate news sentiment into trading strategies.

Understanding customers, markets, and competitors increasingly requires extracting insights from massive text corpora alongside quantitative models. Self-hosted small models lack accuracy for classroom use and risk obsolescence with rapid advancement in AI. LLM APIs provide the