1. Relevance & Research Question
This paper explores the potential of large-language-model–based “digital twins” as a novel instrument for empirical social research. The Last Interview investigates whether an AI agent, trained on an extended multi-stage quantitative survey complemented by a single qualitative session, can faithfully reproduce the attitudes and reasoning patterns of the original respondent. We address three central questions:
• Which types of data are most suitable for constructing such a digital twin?
• How accurately can a data-driven synthetic twin mirror a participant’s measurable opinions across multiple quantitative dimensions?
• What methodological and ethical implications arise when such an agent is used beyond its initial research scope?

2. Methods & Data
The study was conducted at HTW Berlin with volunteer students who consented to provide comprehensive personal data. Each participant completed eight sequential quantitative interview stages—covering demographics, personality traits, value orientations, decision preferences, and behavioral indicators—followed by one in-depth qualitative interview to contextualize the numeric responses. These combined data sets served as the training corpus for an advanced large-language-model architecture, producing individualized conversational agents (“digital clones”).

3. Results
We present the first practical implementation of this concept. Each digital twin was able to answer survey questions in a manner closely aligned with its human counterpart, demonstrating high agreement across quantitative measures. Notably, the clones occasionally produced reasoned extrapolations beyond the original questionnaire, suggesting emergent interpretative capacities despite being trained primarily on structured survey data.

4. Added Value
This project introduces a hybrid methodology that fuses extensive quantitative measurement with state-of-the-art generative modeling to create person-specific research agents. It illustrates how digital twins can support longitudinal analysis without repeated data collection while foregrounding critical issues of informed consent, data sovereignty, and identity representation. For the General Online Research community, The Last Interview provides a replicable framework and empirical evidence for extending quantitative social science into the era of personalized AI. Although we consider the current approach to be in an early developmental stage, the project has already prompted reflections far beyond its initial market-research intent, revealing broader scientific and ethical implications than originally anticipated.

Relevance & Research Question

Advances in computational methods, in particular recent advances in Artificial Intelligence (AI), have vastly reduced the manual effort required to derive results from any given survey dataset. This equally applies to structured, quantitative, interview-level data, but also to qualitative data. For the former, statistical and visualization methods may now be applied automatically; for the latter, sentiments, topics and codes are now easily extracted. As an industry, we’re thus experiencing the commoditization of results. This gives rise to the new questions of how to efficiently work with this overly abundant set of results, how to focus on what matters, how to tell signal from noise.
Methods & Data

In this talk, we introduce the concept of the result space, which we define as the set of all possible results that can be derived from a given dataset of interview-level raw survey data by (automatically) applying current analytical methods. Based on our practical work across dozens of surveys over the past years, we propose alternatives to structuring this space, e.g., by variable, by methodology, or by significance of result; we look into alternatives for sorting and ranking results; and we discuss ways for measuring relations between results.
Results

To strongly anchor the rather theoretical aspects of our work to every-day practical use, we illustrate the specific applicability of these concepts on real-world survey datasets, and on specific questions that we can now answer: What are the Top 3 things to know among all results relating a given sub-demographic? Of all the regression analyses, which ones stand out and why? Is there anything I overlooked in this summary that I wrote? We further demonstrate practicality by showcasing the system used to automatically derive the result spaces.
Added Value

Certain slices through the result space of a survey have already proven their practical value: Personas, for example, allow zeroing in on the particular wants, needs, and opinions of a sub-demographic of particular interest. With the toolkit presented in this talk, we generalize this concept, thereby providing the means to more deeply and more effectively investigate increasingly abundant survey results.

Relevance & Research Question:

Digital transformation, Evidence-based policymaking

Amid rapid technological innovation and digital transformation, EU-ALMPO addresses the need for more agile, inclusive, and responsive to evolving skill mismatches labour market interventions. Anchoring policy design in data, machine-learning analytics, and stakeholder co-creation, the project objective is the creation of the EU Active Labour Market Policy Observatory – an AI-enabled digital hub that enhances the design, implementation, and evaluation of ALMPs across Member States. By integrating advanced AI tools into a centralised digital platform, the Observatory supports evidence-based decisions and fosters knowledge exchange among policymakers, researchers, and labour-market actors.

Methods & Data

Analytical framework, Participatory validation, Comparative policy evaluation

Funded under Horizon Europe, EU-ALMPO has completed the WP1, which developed the analytical framework underpinning the Observatory. The framework analysed existing ALMP systems, identified structural gaps, and assessed the effectiveness of policies in addressing skills mismatches. It also provides a conceptual bridge - a translation layer - for its integration into the project’s AI-supported system for policy-makers across the EU and beyond. Methodologically, it combines a literature review, a meta-evaluation of ALMPs, and participatory validation with stakeholders from several EU countries.

Results

Skills mismatch analysis, Servuction model

Serving both diagnostic and prescriptive functions for skill mismatch analysis, the framework also deepens reflection on the implications of generative technologies for policy innovation. Inspired by the Servuction Model - bridging together front-end and back-end - it develops a user-oriented perspective and translate the knowledge and content developed in the project into actionable items that are valuable and useful for the policy-makers involved.

Added Value

AI-policy integration, Adaptive governance

EU-ALMPO represents a pioneering intersection between labour-market policy and AI. Its analytical framework and reflections around the ways to bridge policy design and AI tools, offers a data-driven, adaptive, and inclusive approach that strengthens Europe’s capacity to respond to future labour-market transformations. While the focus of the project is related to policy making in the area of skills and labour market, the project represents an innovative ground to bridge policy and technology and is thus relevant for potentially other policy areas as well.