AI Researcher & Technologist (PhD)
Paula
Szewach
Computational Social Scientist.
Working at the intersection of democracy, technology, and civil society. My understanding of technology is critical and analytical, grounded in how tools are actually created and used by people, not just whether they work. I am committed to democratic advocacy and responsible technology, with the technical range to build and assess solutions across the full research pipeline.
- Aug 2026 Awarded a Generación de Conocimiento grant, co-PI with Max Pellert, alongside team members Alejandro de la Fuente Cuesta & Albert Martinez Serra, to study moral and sociopolitical bias in AI.
01 / About
Understanding how technology is built, and who it's built for.
I am a computational social scientist with 10+ years working at the intersection of democracy, technology, and civil society. My PhD examined how people interpret and respond to digital information, combining computational, experimental, and survey-based methods to model complex human attitudes online.
My recent work sits at the intersection of LLMs, human behavior, and measurement: designing supervised fine-tuning pipelines for transformer-based models, building human-in-the-loop annotation workflows, and developing structured evaluation frameworks for classification systems. My core expertise is turning abstract human constructs into reproducible, defensible metrics.
I also collaborate with international organizations, NGOs, and multi-country teams on technology, democracy, and public-interest research, work that keeps my AI research technically rigorous, socially aware, and grounded in real human interpretation.
- Affiliation
- Barcelona Supercomputing Center
- Focus
- Democracy, Technology & Society
- Methods
- Causal Inference, Experiments, Surveys, LLM Evaluation
- Consulting
- Responsible AI & Digital Rights
02 / Publications
Selected work
Academic Work
Privacy and Security Matter in AI Chatbot Selection: Evidence from a Two-Stage Experimental Test
Working paper (SSRN)
Leveraging Large Language Models to Enhance Supervised Classification of Political Incivility on Social Media
EPSS conference paper
Information on public opinion has lasting effects on second-order climate beliefs, but minimal and ephemeral effects on first-order beliefs
Journal of Environmental Psychology
The Effects of Forecasts on the Accuracy and Precision of Expectations
Public Opinion Quarterly
Wishful Thinking in Response to Events: Evidence from the 2021 German Federal Election
Electoral Studies
Toolbox of Individual-Level Interventions Against Online Misinformation
Nature Human Behaviour
Public Reactions to Communication of Uncertainty: How Long-Term Benefits Can Outweigh Short-Term Costs
Public Opinion Quarterly
Is Resistance Futile? Citizen Knowledge, Motivated Reasoning, and Fact-Checking
in Knowledge Resistance in High-Choice Information Environments (book chapter)
Professional & Applied Work
Urban Digital Twins: Guidelines for Local Policymakers
Barcelona Supercomputing Center & Eurocities
Media Market Risk Ratings: Argentina
Global Disinformation Index
03 / Projects
Ongoing projects
All of these are underway alongside collaborators and teams, not solo work.
STAGE: Coordinating Multi-Agent Experiments
STAGE is a multi-agent framework that turns a researcher's plain-language experimental conditions into a live chatroom built around a real participant.
Cybersecurity Incident Timeline
An open-source documentation project tracking cybersecurity incidents in Argentina, where breaches are rarely disclosed and never systematically recorded. Built from civil society using public sources, it treats that missing evidence as a policy problem worth naming, not just a data gap.
Do AI Guardrails Protect Everyone Equally?
A framework for stress-testing AI guardrails against real user behavior rather than regulatory text alone. Using large-scale survey data, it builds a typology of user risk profiles and shows that compliance failures fall hardest on the least digitally equipped.
Attitudes Towards Data Privacy and Security
Studying how people weigh privacy and security when choosing AI tools, and what interventions actually shift those preferences.
Is Our Political Leaning More at Risk with LLMs?
Comparing how well large language models can read political leaning from browsing history against classic statistical models trained on labeled data. What's at stake isn't just accuracy, but who can now run this kind of profiling, and how easily.
04 / Contact
Working on something at the intersection of AI, democracy, and society?
Don't hesitate to reach out.