Research Group Frauke Kreuter

AI-driven decision-making systems are becoming instrumental in the public sector, with applications spanning areas like criminal justice, social welfare, financial fraud detection, and public health. While these systems offer great potential benefits to institutional decision-making processes, such as improved efficiency and reliability, these systems face the challenge of aligning machine learning (ML) models with the complex realities of public sector decision-making. In this paper, we examine five key challenges where misalignment can occur, including distribution shifts, label bias, the influence of past decision-making on the data side, as well as competing objectives and human-in-the-loop on the model output side. Our findings suggest that standard ML methods often rely on assumptions that do not fully account for these complexities, potentially leading to unreliable and harmful predictions. To address this, we propose a shift in modeling efforts from focusing solely on predictive accuracy to improving decision-making outcomes. We offer guidance for selecting appropriate modeling frameworks, including counterfactual prediction and policy learning, by considering how the model estimand connects to the decision-maker’s utility. Additionally, we outline technical methods that address specific challenges within each modeling approach. Finally, we argue for the importance of external input from domain experts and stakeholders to ensure that model assumptions and design choices align with real-world policy objectives, taking a step towards harmonizing AI and public sector objectives.

Recent advances in Large Language Models (LLMs) have sparked wide interest in validating and comprehending the human-like cognitive-behavioral traits LLMs may capture and convey. These cognitive-behavioral traits include typically Attitudes, Opinions, Values (AOVs). However, measuring AOVs embedded within LLMs remains opaque, and different evaluation methods may yield different results. This has led to a lack of clarity on how different studies are related to each other and how they can be interpreted. This paper aims to bridge this gap by providing a comprehensive overview of recent works on the evaluation of AOVs in LLMs. Moreover, we survey related approaches in different stages of the evaluation pipeline in these works. By doing so, we address the potential and challenges with respect to understanding the model, human-AI alignment, and downstream application in social sciences. Finally, we provide practical insights into evaluation methods, model enhancement, and interdisciplinary collaboration, thereby contributing to the evolving landscape of evaluating AOVs in LLMs.

Algorithmic profiling is increasingly used in the public sector with the hope of allocating limited public resources more effectively and objectively. One example is the prediction-based profiling of job seekers to guide the allocation of support measures by public employment services. However, empirical evaluations of potential side-effects such as unintended discrimination and fairness concerns are rare in this context. We systematically compare and evaluate statistical models for predicting job seekers’ risk of becoming long-term unemployed concerning subgroup prediction performance, fairness metrics, and vulnerabilities to data analysis decisions. Focusing on Germany as a use case, we evaluate profiling models under realistic conditions using large-scale administrative data. We show that despite achieving high prediction performance on average, profiling models can be considerably less accurate for vulnerable social subgroups. In this setting, different classification policies can have very different fairness implications. We therefore call for rigorous auditing processes before such models are put to practice.

Large language models (LLMs) are perceived by some as having the potential to revolutionize social science research, considering their training data includes information on human attitudes and behavior. If these attitudes are reflected in LLM output, LLM-generated ‘synthetic samples’ could be used as a viable and efficient alternative to surveys of real humans. However, LLM-synthetic samples might exhibit coverage bias due to training data and fine-tuning processes being unrepresentative of diverse linguistic, social, political, and digital contexts. In this study, we examine to what extent LLM-based predictions of public opinion exhibit context-dependent biases by predicting voting behavior in the 2024 European Parliament elections using a state-of-the-art LLM. We prompt GPT-4-Turbo with anonymized individual-level background information, varying prompt content and language, ask the LLM to predict each person’s voting behavior, and compare the weighted aggregates to the real election results. Our findings emphasize the limited applicability of LLM-synthetic samples to public opinion prediction. We show that (1) the LLM-based prediction of future voting behavior largely fails, (2) prediction accuracy is unequally distributed across national and linguistic contexts, and (3) improving LLM predictions requires detailed attitudinal information about individuals for prompting. In investigating the contextual differences of LLM-based predictions of public opinion, our research contributes to the understanding and mitigation of biases and inequalities in the development of LLMs and their applications in computational social science.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model’s diverse response styles such as starting with ‘Sure’ or refusing to answer. Consequently, first-token evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

In this study, we explore the proficiency of large language models (LLMs) in understanding two key lexical aspects: duration (durative/stative) and telicity (telic/atelic). Through experiments on datasets featuring sentences, verbs, and verb positions, we prompt the LLMs to identify aspectual features of verbs in sentences. Our findings reveal that certain LLMs, particularly those closed-source ones, are able to capture information on duration and telicity, albeit with some performance variations and weaker results compared to the baseline. By employing prompts at three levels (sentence-only, sentence with verb, and sentence with verb and its position), we demonstrate that integrating verb information generally enhances performance in aspectual feature recognition, though it introduces instability. We call for future research to look deeper into methods aimed at optimizing LLMs for aspectual feature comprehension.

We present a research agenda focused on efficiently extracting, assuring quality, and consolidating textual company sustainability information to address urgent climate change decision-making needs. Starting from the goal to create integrated FAIR (Findable, Accessible, Interoperable, Reusable) climate-related data, we identify research needs pertaining to the technical aspects of information extraction as well as to the design of the integrated sustainability datasets that we seek to compile. Regarding extraction, we leverage technological advancements, particularly in large language models (LLMs) and Retrieval-Augmented Generation (RAG) pipelines, to unlock the underutilized potential of unstructured textual information contained in corporate sustainability reports. In applying these techniques, we review key challenges, which include the retrieval and extraction of CO2 emission values from PDF documents, especially from unstructured tables and graphs therein, and the validation of automatically extracted data through comparisons with human-annotated values. We also review how existing use cases and practices in climate risk analytics relate to choices of what textual information should be extracted and how it could be linked to existing structured data.

Whether future AI models are fair, trustworthy, and aligned with the public’s interests rests in part on our ability to collect accurate data about what we want the models to do. However, collecting high-quality data is difficult, and few AI/ML researchers are trained in data collection methods. Recent research in data-centric AI has show that higher quality training data leads to better performing models, making this the right moment to introduce AI/ML researchers to the field of survey methodology, the science of data collection. We summarize insights from the survey methodology literature and discuss how they can improve the quality of training and feedback data. We also suggest collaborative research ideas into how biases in data collection can be mitigated, making models more accurate and human-centric.

Automated decision-making (ADM) systems are being deployed across a diverse range of critical problem areas such as social welfare and healthcare. Recent work highlights the importance of causal ML models in ADM systems, but implementing them in complex social environments poses significant challenges. Research on how these challenges impact the performance in specific downstream decision-making tasks is limited. Addressing this gap, we make use of a comprehensive real-world dataset of jobseekers to illustrate how the performance of a single CATE model can vary significantly across different decision-making scenarios and highlight the differential influence of challenges such as distribution shifts on predictions and allocations.

The recent development of large language models (LLMs) has spurred discussions about whether LLM-generated ‘synthetic samples’ could complement or replace traditional surveys, considering their training data potentially reflects attitudes and behaviors prevalent in the population. A number of mostly US-based studies have prompted LLMs to mimic survey respondents, with some of them finding that the responses closely match the survey data. However, several contextual factors related to the relationship between the respective target population and LLM training data might affect the generalizability of such findings. In this study, we investigate the extent to which LLMs can estimate public opinion in Germany, using the example of vote choice. We generate a synthetic sample of personas matching the individual characteristics of the 2017 German Longitudinal Election Study respondents. We ask the LLM GPT-3.5 to predict each respondent’s vote choice and compare these predictions to the survey-based estimates on the aggregate and subgroup levels. We find that GPT-3.5 does not predict citizens’ vote choice accurately, exhibiting a bias towards the Green and Left parties. While the LLM captures the tendencies of ’typical’ voter subgroups, such as partisans, it misses the multifaceted factors swaying individual voter choices. By examining the LLM-based prediction of voting behavior in a new context, our study contributes to the growing body of research about the conditions under which LLMs can be leveraged for studying public opinion. The findings point to disparities in opinion representation in LLMs and underscore the limitations in applying them for public opinion estimation.

Identifying constructs in text data is a labor-intensive task in social science research. Despite the potential richness of open-ended survey responses, the complexity of analyzing them often leads researchers to underutilize or ignore them entirely. While topic modeling offers a technological solution, qualitative researchers may remain skeptical of its rigor. In this paper, we introduce TOPCAT: Topic-Oriented Protocol for Content Analysis of Text, a systematic approach that integrates off-the-shelf topic modeling with human decisionmaking and curation. Our method aims to provide a viable solution for topicalizing open-ended responses in survey research, ensuring both efficiency and trustworthiness. We present the TOPCAT protocol, define an evaluation process, and demonstrate its effectiveness using open-ended responses from a U.S. survey on COVID-19 impact. Our findings suggest that TOPCAT enables efficient and rigorous qualitative analysis, offering a promising avenue for future research in this domain. Furthermore, our findings challenge the adequacy of expert coding schemes as ‘‘gold’’ standards, emphasizing the subjectivity inherent in qualitative content interpretation.

Conversational large language models are trained to refuse to answer harmful questions. However, emergent jailbreaking techniques can still elicit unsafe outputs, presenting an ongoing challenge for model alignment. To better understand how different jailbreak types circumvent safeguards, this paper analyses model activations on different jailbreak inputs. We find that it is possible to extract a jailbreak vector from a single class of jailbreaks that works to mitigate jailbreak effectiveness from other semantically-dissimilar classes. This may indicate that different kinds of effective jailbreaks operate via a similar internal mechanism. We investigate a potential common mechanism of harmfulness feature suppression, and find evidence that effective jailbreaks noticeably reduce a model’s perception of prompt harmfulness. These findings offer actionable insights for developing more robust jailbreak countermeasures and lay the groundwork for a deeper, mechanistic understanding of jailbreak dynamics in language models.

Prompt-based methods have been successfully applied to multilingual pretrained language models for zero-shot cross-lingual understanding. However, most previous studies primarily focused on sentence-level classification tasks, and only a few considered token-level labeling tasks such as Named Entity Recognition (NER) and Part-of-Speech (POS) tagging. In this paper, we propose Token-Level Prompt Decomposition (ToPro), which facilitates the prompt-based method for token-level sequence labeling tasks. The ToPro method decomposes an input sentence into single tokens and applies one prompt template to each token. Our experiments on multilingual NER and POS tagging datasets demonstrate that ToPro-based fine-tuning outperforms Vanilla fine-tuning and Prompt-Tuning in zero-shot cross-lingual transfer, especially for languages that are typologically different from the source language English. Our method also attains state-of-the-art performance when employed with the mT5 model. Besides, our exploratory study in multilingual large language models shows that ToPro performs much better than the current in-context learning method. Overall, the performance improvements show that ToPro could potentially serve as a novel and simple benchmarking method for sequence labeling tasks.

The data-centric revolution in AI has revealed the importance of high-quality training data for developing successful AI models. However, annotations are sensitive to annotator characteristics, training materials, and to the design and wording of the data collection instrument. This paper explores the impact of observation order on annotations. We find that annotators’ judgments change based on the order in which they see observations. We use ideas from social psychology to motivate hypotheses about why this order effect occurs. We believe that insights from social science can help AI researchers improve data and model quality.

Despite the predominance of English in their training data, English-centric Large Language Models (LLMs) like GPT-3 and LLaMA display a remarkable ability to perform multilingual tasks, raising questions about the depth and nature of their cross-lingual capabilities. This paper introduces the decomposed prompting approach to probe the linguistic structure understanding of these LLMs in sequence labeling tasks. Diverging from the single text-to-text prompt, our method generates for each token of the input sentence an individual prompt which asks for its linguistic label. We assess our method on the Universal Dependencies part-of-speech tagging dataset for 38 languages, utilizing both English-centric and multilingual LLMs. Our findings show that decomposed prompting surpasses the iterative prompting baseline in efficacy and efficiency under zero- and few-shot settings. Further analysis reveals the influence of evaluation methods and the use of instructions in prompts. Our multilingual investigation shows that English-centric language models perform better on average than multilingual models. Our study offers insights into the multilingual transferability of English-centric LLMs, contributing to the understanding of their multilingual linguistic knowledge.

Linking digital trace data to existing panel survey data may increase the overall analysis potential of the data. However, producing linked products often requires additional engagement from survey participants through consent or participation in additional tasks. Panel operators may worry that such additional requests may backfire and lead to lower panel retention, reducing the analysis potential of the data. To examine these concerns, we conducted an experiment in the German PASS panel survey after wave 11. Three quarters of panelists (n = 4,293) were invited to install a research app and to provide sensor data over a period of 6 months, while one quarter (n = 1,428) did not receive an invitation. We find that the request to install a smartphone app and share data significantly decreases panel retention in the wave immediately following the invitation by 3.3 percentage points. However, this effect wears off and is no longer significant in the second and third waves after the invitation. We conclude that researchers who run panel surveys have to take moderate negative effects on retention into account but that the potential gain likely outweighs these moderate losses.

Objectives: To examine the association of non-pharmaceutical interventions (NPIs) with anxiety and depressive symptoms among adults and determine if these associations varied by gender and age.
Methods: We combined survey data from 16,177,184 adults from 43 countries who participated in the daily COVID-19 Trends and Impact Survey via Facebook with time-varying NPI data from the Oxford COVID-19 Government Response Tracker between 24 April 2020 and 20 December 2020. Using logistic regression models, we examined the association of [1] overall NPI stringency and [2] seven individual NPIs (school closures, workplace closures, cancellation of public events, restrictions on the size of gatherings, stay-at-home requirements, restrictions on internal movement, and international travel controls) with anxiety and depressive symptoms.
Results: More stringent implementation of NPIs was associated with a higher odds of anxiety and depressive symptoms, albeit with very small effect sizes. Individual NPIs had heterogeneous associations with anxiety and depressive symptoms by gender and age.
Conclusion: Governments worldwide should be prepared to address the possible mental health consequences of stringent NPI implementation with both universal and targeted interventions for vulnerable groups.

Functions and datasets to support Valliant, Dever, and Kreuter (2018), doi:10.1007/978-3-319-93632-1, ‘Practical Tools for Designing and Weighting Survey Samples’. Contains functions for sample size calculation for survey samples using stratified or clustered one-, two-, and three-stage sample designs, and single-stage audit sample designs. Functions are included that will group geographic units accounting for distances apart and measures of size. Other functions compute variance components for multistage designs and sample sizes in two-phase designs. A number of example data sets are included.