DEMOS-AI

02

AI systems are not politically empty.

Large language models absorb ideological and cultural regularities from their training data. These representations can affect how models classify, generate and recommend content — with real consequences for scientific knowledge and public perception.

01

Where are political values represented inside language models?

02

Can those representations be manipulated?

03

Do they affect what models classify, recommend or generate?

Three key research questions

03

Our research programme

We combine representation analysis, model intervention and behavioural evaluation to understand the nature and consequences of sociopolitical biases in AI.

01

MAP

Political values inside LLMs

Linear probing · Representation analysis

02

INTERVENE

Change internal representations

Activation steering

03

TEST

Observe behavioural consequences

Classification · Polarization · Incivility

04

AUDIT

Question the gold standard

Human–LLM auditing

05

UNDERSTAND

Political and societal consequences

Public opinion · AI-mediated information

04

Models methods evidence

Political values across model layers

We use linear probing to detect political and moral value representations in model activations across layers, and test how they can be modified through activation steering.

Fig. 2 — Heatmap of linear probe signal for authority, egalitarianism, achievement, benevolence, purity and care across model layers 0 to 32

Illustrative figure — pattern the project will test, not yet a result.

05

Selected outputs

All outputs →

Paper · 2025Only a Little to the Left: A Theory-grounded Measure of Political Bias in LLMsACL →
Paper · 2025Extracting Affect Aggregates from Social Media Data with Temporal Adapters for LLMsICWSM →
Preprint · 2025LLMs replicate and predict human cooperation across experiments in game theoryarXiv →