Political Economy + AI =

(Ex–Public Governance Working Group)

An online research seminar organised jointly by King's College London and Université Paris-Dauphine – PSL.

Contact the organisers: Vladimir Avetian or Edgar Jimenez Bedolla.

About: Political economy research using AI and computational methods. Work in progress and methods talks welcome.

Autumn 2026 season

Mondays, 16:30–17:30 London / 17:30–18:30 Paris, online.
45-minute presentation + 15-minute discussion.

Philine Widmer (UniDistance Suisse, PSE)

The Slant of Large Language Models: Evidence from a Persuasion Experiment

(with Tiffany Kreutschy)

Abstract

Large language models (LLMs) promise personalized persuasion at near-zero cost, but their output has a slant of its own. In an experiment, U.S. respondents viewed an LLM-generated post advocating a policy stance, randomly personalized to up to six demographics. Messages persuade on average. Personalization raises persuasion only when matching on political party, almost exclusively among Democrats. Benchmarked against Congress members' language, messages slant Democratic even when prompted to target Republicans. Paradoxically, personalization with Democratic-slanted LLMs can benefit Republican positions by finding language to persuade Democrats. Our findings speak to debates about persuasion and slant in the age of generative AI.

Sultan Mehmood (New Economic School)

Concealed Repression and Conflict

(with Nico Voigtländer)

Abstract

States repress insurgency through public force and hidden coercion. We study Pakistan’s 2007–2008 campaign against Taliban-linked militancy, combining administrative records on enforced disappearances with sub-district panels on political violence and protest. The campaign generated visible military operations in Pashtun areas and covert spillovers against Baloch individuals outside the main theater of Baloch insurgency. This variation allows us to compare political responses to open and concealed repression. Visible repression is followed by armed backlash, with little protest response. Covert repression is associated instead with muted mobilization and a fall in militant violence. Overt and covert repressions produce sharply different political responses.

Andreu Casas Salleras (University of Oxford)

Large Language Models as Lab Agents: Transfer Learning for Building Generalizable Automatic In-House Annotators

Abstract

Automatic data annotation is a common task in social science research, such as labeling social media posts for the presence of hate speech, or new articles for their ideological leaning. When fine-tuned with some manually annotated examples, Large Language Models (LLMs) consistently outperform traditional machine learning models. However, researchers often fine-tune a pre-trained LLM from scratch for each new task at hand. Building on the notion of transfer learning, in this paper we leverage several datasets annotated in the same research lab, for different tasks and projects, to develop a Lab Agent that gradually performs better at zero-shot and few-shot classification on new unseen tasks. We explore the performance of several options in the development of such Lab Agents, with the goal of facilitating and speeding up accurate data annotation for social science research.

Christian Arnold (University of Birmingham)

Invisibilization as a Source of Probability Bias in Language Models

Abstract

Declining to judge a person by their race looks like a fair choice. But for a system that must decide how much to learn from what people tell it, the choice has a price. Language models are aligned to avoid using protected attributes, and they increasingly refuse or hedge when such attributes appear. We measure what that costs in a two-urn game with an exact Bayesian answer, in which a model must decide how far to trust an unverifiable report from a participant whose description we randomize.

Across ten configurations from two developers (6,600 sessions), models often decline to assess the reporter, treating the report as worthless, at rates set by the description rather than by the evidence. At its widest, a Black-described reporter is declined 46% of the time and a white-described reporter 9%, yet when assessed, the Black-described reporter receives slightly more trust. The bias is invisible to audits that compare average trust or look only at assessed cases.

It grows with reasoning effort, follows words rather than categories (models invoke race for “Black” but not for its Spanish translation negro, and never for “Indigenous”), and worsens under the two instructions a practitioner would try first. Changing the rules works where instructions fail.

When a model’s trust is scored against revealed truth over repeated rounds, it credits identical records equally: in four of five configurations from both developers, the racial gap in trust falls below 0.02 by the last rounds. Invisibilization rests on presumptions about strangers that nobody checks, and accountability to evidence removes it.

Spring 2027 season

Programme to be announced soon.

Archive

Access previous working group sessions