Welcome to my website!
I'm Sam Fraiberger, a Senior AI Scientist at the World Bank. I study how people and AI behave, on their own and together. I turn what I learn into AI tools that improve lives where the need is greatest: helping governments choose policies that work, spotting food crises months before they hit, and serving languages and communities the big models overlook.
I founded and lead the Development Impact AI Lab, a team of more than thirty scientists, engineers, and researchers. Our mission is to build the high impact AI the frontier labs won't, and to measure whether it works: every tool we release grows out of our own peer reviewed research and is tested in the field.
My research has appeared in Science, PNAS, ACL, EMNLP, and the Journal of International Economics, and has been featured in Nature, The Wall Street Journal, The Economist, The Washington Post, Axios, a CBC documentary, and a TEDx talk. I've been named to Apolitical's Government AI 100 in 2025 and 2026, and I was a core contributor to the World Development Report 2026 on Artificial Intelligence.
I'm also a visiting researcher at NYU's Center for Data Science. Before the World Bank, I did postdoctoral research at Northeastern and Harvard, and completed my PhD at NYU.
Selected Tools
ImpactAI helps policymakers compare what works and pick the most effective option. It draws on more than 15,000 randomized trials from over 150 countries, with every effect put on one scale and traced back to its source. Development banks, donors, foundations, governments, and NGOs use it.
ZeroHungerAI helps responders see crises coming where the data runs out. It turns millions of news reports and satellite data into district level risk signals across 82 countries, detecting food crises up to 12 months ahead and catching 46% more outbreaks than existing baselines. It is open access, has an open API, and is part of the AI Collaborative: Food Security.
NaijaXLM-T shows what small AI built on local data can do. Fine tuned on representative Nigerian social media, it detects hate speech far better than ChatGPT, and it powered a randomized trial in which a preventive messaging campaign led to a large and lasting drop in online hate. It reflects a broader approach: find where existing models fall short, build the missing data and small adapted models, and test them in the field.
Get the model → Read the blog:part 1part 2 → The research → The field trial →
Selected Publications
Different Demographic Cues Yield Inconsistent Conclusions About LLM Personalization and Bias
We show that different cues for the same demographic group shift language model answers in only partly overlapping ways, so conclusions about personalization and bias depend on which cue is used.
We show that language models match local gender norms in the US, India, Kenya, and Nigeria when asked directly, but skew male when they generate media in local languages.
Predicting Causal Effects from Natural Language Queries using Structured Representations
We introduce Query2Effect, a benchmark of more than 72,000 queries, and a method that predicts the effects of policy interventions from plain language questions far more accurately than prompted language models.
The Enforcement and Feasibility of Hate Speech Moderation
We show that most hateful tweets were still online five months after posting, and that pairing AI with human moderators could cut exposure at a feasible cost.
Celebrity messages reduce online hate and limit its spread
We run a randomized trial in Nigeria and show that preventive messages from celebrities reduce online hate and limit its spread.
Can social media reliably estimate unemployment?
We show that detecting self reported job loss on Twitter, adjusted for demographics, improves forecasts of US unemployment claims up to two weeks before official data.
HateDay: Insights from a Global Hate Speech Dataset Representative of a Day on Twitter
We introduce the first global hate speech dataset representative of a single day on Twitter, across eight languages and four English speaking countries.
NaijaHate: Evaluating Hate Speech Detection on Nigerian Twitter
We introduce the first hate speech dataset built on a representative sample of Nigerian tweets, and show that models tested on convenient data greatly overestimate their real world performance. This work led to NaijaXLM-T.
EconBERTa: Towards Robust Extraction of Named Entities in Economics
We introduce a language model pretrained on 1.5 million economics articles and a new dataset for extracting entities from economics research. This work led to ImpactAI.
Long ties, disruptive life events, and economic prosperity
We show that long distance social ties play a causal role in economic mobility, using Facebook network data at population scale.
Predicting food crises using news streams
We show that news can anticipate food crises up to 12 months ahead, using 11 million articles. This work led to ZeroHungerAI.
Multilingual Detection of Personal Employment Status on Twitter
We develop an active learning method to detect disclosures of employment status on social media in three languages, with applications to job matching, social protection, and labor market measurement.
Mobile phone data for informing public health actions across the COVID-19 pandemic life cycle
We set out how mobile phone data can guide public health action at each stage of a pandemic.
Identifying Predictive Causal Factors from News Streams
We develop a framework for finding the news events that predict real world outcomes.
Quantifying reputation and success in art
We map the careers of half a million artists through the global network of galleries and museums.
Honors
I was named one of the 100 most influential people working on AI in government, two years running.
Our paper HateDay, a global hate speech dataset representative of a day on Twitter, received this award at the leading conference in computational linguistics.
ImpactAI was selected for the inaugural cohort of Google.org's $20M accelerator for high impact AI.
The award recognized "Quantifying reputation and success in art," published in Science.
Selected Press
AI and Global Food Security: A Focus on Early Warning Systems
This analysis of AI early warning systems for food security features ZeroHungerAI.
This feature looks at how researchers are testing AI to help the world's poorest people.
Why Big Changes Early in Life Can Help Later On
This feature covers the PNAS study of long ties and economic prosperity.
AI Can Now Forecast the Next Food Crisis
This story covers the Science Advances study predicting food crises from news.
News You Can Use to Better Predict Food Crisis Outbreaks
This story explains how news coverage can flag food crises before they break.
Tyler Cowen and Alex Tabarrok highlighted the food crisis study on their blog.
Machine Learning Uses News to Predict Food Insecurity
This story explains how machine learning can anticipate food crises.
This documentary explores how reputation and success spread through networks, featuring the research on art careers.
To Get to the Top of the Art World, Start There in the First Place
This article looks at how early access to prestigious institutions shapes an artist's career.
The Surprising Formula for Becoming an Art Star
This article explores what the data reveals about which artists become stars.
How to Be a Successful Artist, According to Science
This article covers the science of artistic careers and reputation networks.
The Surprising (?) Formula for Becoming an Art Star
Tyler Cowen and Alex Tabarrok highlighted the art careers study on their blog.
The Secret to Making It as an Artist
This article explains why an artist's first years matter most.
Success in Prestigious Galleries Is Key to Lifelong Artistic Success
This story covers the Science study of art careers.
Facebook Thinks You're Gay or a Drug User Based on Three 'Likes,' Study Says
This article covers what a handful of Facebook likes can reveal about a person.
The Real Promise of the Sharing Economy Is What It Could Do for the Poor
This article looks at how peer to peer rental markets could benefit lower income households.
Selected Talks
Global Impact Evaluation Forum
I gave an invited talk on AI and impact evaluation.
I gave an invited talk on deploying AI in international institutions.
Workshop on Food and Humanitarian Crises
I gave an invited talk on using AI to anticipate food crises.
Can We Predict Artistic Success?
I gave a talk on what data reveals about creative success.
Service
I serve on the journal's editorial board.