PaperlayerView exact revision on arXiv

Source-linked AI summary

Risk or Chance? Large Language Models and Reproducibility in HCI Research

Thomas Kosch, Sebastian Feger

arXiv:2404.15782v3cs.HC

TL;DR

HCI’s methodological diversity makes reproducibility a concern as LLM use expands across UX research. This article reviews risks and opportunities and offers initial recommendations; it highlights both potential benefits and reproducibility concerns.

  • Problem

    HCI’s diverse research methodologies present reproducibility challenges, and the effects of increasing LLM adoption across UX research activities warrant examination.

    Source: §1 · p. 1

  • Method

    The article reviews LLM-related reproducibility risks and opportunities across UX research, drawing on analogies, bias, data analysis, documentation, education, and publication pressure.

    Source: §1 · p. 1, §3 · p. 2

  • Results

    The article identifies potential for LLMs to accelerate data analysis and disseminate results, alongside concerns about inconsistent outputs, bias, and publication pressure.

    Source: §4 · p. 4

  • Takeaways & Limitations

    The authors call for focused HCI community discussions to explore best practices for valid and reproducible LLM use.

    Source: §4 · p. 4

Abstract

from arXiv · show

Reproducibility is a major concern across scientific fields. Human-Computer Interaction (HCI), in particular, is subject to diverse reproducibility challenges due to the wide range of research methodologies employed. In this article, we explore how the increasing adoption of Large Language Models (LLMs) across all user experience (UX) design and research activities impacts reproducibility in HCI. In particular, we review upcoming reproducibility challenges through the lenses of analogies from past to future (mis)practices like p-hacking and prompt-hacking, general bias, support in data analysis, documentation and education requirements, and possible pressure on the community. We discuss the risks and chances for each of these lenses with the expectation that a more comprehensive discussion will help shape best practices and contribute to valid and reproducible practices around using LLMs in HCI research.

1 INTRODUCTION

As LLMs become accessible tools for HCI research, their use raises questions about reproducibility across a field with diverse methodologies. This article reviews risks and opportunities and aims to inform reproducible practices.

1 INTRODUCTION

  • LLMs are anticipated to become commonplace in HCI data analysis and have been explored for textual analysis, user-centered design, and simulating participants or replicating studies [2].Source: §1 · p. 1
  • HCI’s methodological diversity creates reproducibility challenges, prompting a need for systematic understanding of how LLM use affects reproducibility.Source: §1 · p. 1
  • The article reviews reproducibility risks and opportunities involving p- and prompt-hacking, bias, data analysis, documentation, education, and publication pressure.Source: §1 · p. 1

2 LARGE LANGUAGE MODELS AS RESEARCH TOOL

LLMs’ architecture and use of training data can make outputs difficult to reproduce or validate. Their use also raises concerns about value lock-in, cultural bias, and hallucinations in HCI research.

2 LARGE LANGUAGE MODELS AS RESEARCH TOOL

  • Autoregressive LLMs generate each token from preceding tokens and cannot validate outputs during generation; prompt engineering can make repeated outputs less equal.The models cannot holistically self-adjust, and users cannot bring human resources into generation to improve output precision or validation.Source: §2 · pp. 1–2

2.1 Value Lock-In

  • Because models may retain attitudes from their initial training, analyses can appear meaningful without reflecting participants’ beliefs, while model updates may unpredictably change results.A repository documenting model version, parameters, prompts, and outputs may circumvent some issues.Source: §2.1 · p. 2

2.2 Training Bias

  • Training data drawn from internet users may reflect Western, educated, industrialized, affluent, and democratic perspectives, reinforcing cultural bias in research analysis.Human reviewers may amplify such bias through confirmation bias.Source: §2.2 · p. 2

2.3 Hallucination

  • Unnoticed hallucinations in LLM-supported analyses or simulated participation may enter published results, motivating validation methods for HCI research.The paper asks how qualitative and quantitative analysis results should be tested.Source: §2.3 · p. 2

3 IMPLICATIONS FOR REPRODUCIBILITY IN HCI RESEARCH

The section maps reproducibility risks and opportunities across HCI research activities as LLM use expands, and proposes practices to address them. It considers prompt manipulation, bias, analysis support, reporting and education, and pressure on researchers and reviewers.

3.1 Learning from Today’s Reproducibility Challenges: Let’s not Repeat Mistakes

  • Prompt hacking can skew LLM responses by adjusting prompts, paralleling p-hacking; the authors propose identifying analogous risks and evaluating pre-registration and transparent prompt protocols.They recommend learning from existing best practices rather than repeating past mistakes.Source: §3.1 · pp. 2–3

3.2 Bias Across User Experience Research

  • LLMs may amplify narrow training-data perspectives, so the authors call for transparent training data, multiple models, and scrutiny of how model combinations and human reports affect each use case.Combining models oriented toward different regions or cultures is offered as one example.Source: §3.2 · p. 3

3.3 LLMs for Cross-Validation and Analysis Support

  • LLMs could support cross-validation and analysis across HCI methods, but should remain assistive rather than become a single source of analysis.For qualitative work, they might complement manual analysis; for quantitative work, they can cross-check statistical-test applicability and calculations.Source: §3.3 · p. 3

3.4 Defining New Reporting Requirements and Educating the Community

  • The authors call for venue-level reporting requirements covering LLM use, prompts, models, and versions, alongside education about bias, hallucination, and value lock-in.They also recommend transparent, accessible models that disclose training data and remain available for reproduction.Source: §3.4 · p. 3

3.5 The Risk of Increased Research Pressure on HCI Reproducibility

  • Premature LLM adoption may bring mispractices before best practices are established, while increased submissions could strain peer review; the authors propose managing expectations, requiring practices, and educating reviewers.They also suggest evaluating how LLMs might support peer review while considering bias, hallucination, and value lock-in.Source: §3.5 · pp. 3–4

4 RISK OR CHANCE FOR REPRODUCIBILITY IN HCI?

LLMs offer HCI opportunities to accelerate data analysis and disseminate results, but their design and frequent updates can complicate reproducibility. The article calls for focused community discussion about their impact and specialized platforms for reproducible AI-supported analysis.

  • LLMs may accelerate HCI data analysis and dissemination, but their structure, training, and frequent updates are not designed for consistent results and may affect reproducibility, manifest biases, and increase publication pressure.The article aims to prompt focused discussions, including on specialized platforms for disseminating reproducible AI-supported HCI data analysis.Source: §4 · p. 4
Loading 2404.15782v3…