Integration Of Large-scale Language Models In HCI Research And Ethical Issues
3 main points
✔️ Examines the use of large-scale language models inHCIresearch activities
✔️ While large-scale language models are becoming more prevalent,concerns about ethical issuessuch as privacy violations and misuse of data
✔️ Raises need for ethical guidelines and education to support responsible use of large-scale language models in HCI research and education to support responsible use of large-scale language models in HCI research
"I'm categorizing LLM as a productivity tool": Examining ethics of LLM use in HCI research practices
written by Yongqi Tong, Dawei Li, Sizhe Wang, Yujia Wang, Fei Teng, Jingbo Shang
(Submitted on 28 Mar 2024)
Comments: Published on arxiv.
Subjects: Human-Computer Interaction (cs.HC)
code:
The images used in this article are from the paper, the introductory slides, or were created based on them.
Summary
Large-scale language models have developed rapidly in recent years. And they are bringing many changes to the field of research and education. These models have become a powerful complement to the research process in human-computer interaction (HCI) studies, analyzing qualitative and quantitative data, reproducing social science experiments, and simulating new social dynamics. At the same time, however, risks of misinformation, discrimination, and privacy violations have been identified, andcareful consideration is required for the safe use oflarge-scale languagemodels.
Ethical considerations in the HCI field have been studied in the past, but the application of large-scale language models presents a new set of challenges.There has been much research and discussion onhowlarge-scale languagemodels are incorporated into projects, what ethical concerns exist, and how these concerns are being addressed tofill the gap inpractical applicationtoward effective and ethical integration oflarge-scale languagemodels.
Based on data from a survey of 50 HCI researchers and 16 interviews, this paper presents how large-scale language models are being used in the overall HCI research process. This includes applications ranging from idea generation to system development to thesis writing. It also addresses researchers' anticipation of and responses to potential ethical issues related to large-scale language models.
Finally, based on these findings, the report presents the importance of research ethics and specific recommendations for this purpose, including working with ethics review committees and other groups, improving the informed consent process, expanding opportunities for ethics education, and revising research incentive structures.The HCI field has been working with a large language models, suggesting that a major challenge for the future will be how to promote these ethical considerations as we move forward with model integration.
Questionnaire survey
This paper focuses on how HCI researchers incorporate large-scale language models into their research process and how they address ethical considerations. To gain a comprehensive understanding in this regard, we employ a Sequential Explanatory Design, a survey and semi-structured interviews designed to gather opinions from a wide range of perspectives. This approach allows us to investigate in detail the specific ethical concerns of the researchers. Please note that this research project has been approved by the IRB of my institution.
The survey was conducted using an online survey implemented in Qualtrics and collected data from 50 respondents.Participants were recruited through multiple channels, including social media advertising, direct email, and university mailing lists. Informed consent was obtained from participants at the beginning of the survey and personally identifiable information was withheld from recording.Selection criteria are that participants must be researchers working in HCI-related fields andhave experience usinglarge-scale languagemodels.
Of the final 77 participants, 50 completed all sections (except demographic information). Of the 43 who responded to the demographic information, many indicated that they work in academia, industry, or non-profit organizations. These participants work in various subfields of the HCI field, with responses from countries as diverse as the United States, Afghanistan, Germany, Algeria, Hong Kong, China, Spain, Nigeria, Australia, and Japan. On average, participants have been engaged in HCI research for four years.
We also use a questionnaire consisting of a total of 18 questions. This includes 14 multiple-choice questions and 4 open-ended questions.Thequestionnairedefines alarge-scale languagemodelas "a subset of generative (and multimodal) language models whose scale increases with the number of parameters and the size of the training data" anddescribesGPT-4, GPT-3.5, Llama 2, and Vicunaasspecificexamples.
Survey participantswere asked to answer questions based on their experience usinglarge-scale languagemodelsin recent research projects, which were divided into three sections
- Questions about the use of large-scale language models in HCI research projects
- Questions about the ethics of using large-scale language models in HCI research
- Demographic questions relevant to the study
Thefirst part of the questionnaire describes the project in one sentence andasks about themain research methods used, the HCI subarea, andthe stages of the research process that incorporatedlarge-scale languagemodels.Italsofocused on potential ethical considerations associated with the use of large-scale language models, asking, "Have you encountered or observed anylarge-scale languagemodel-related ethical challengesin your research project?" to ask closed-ended and open-ended questions about ethical challenges. We also ask how those ethical challenges were identified, addressed, and reported. Finally, we include demographic questions about the respondent's institutional type, country, and years of experience in HCI research. Please note that this response is optional.
The data analysis uses SPSS to compute descriptive statistics to analyze how researchers address ethical concerns associated with large-scale language models. In particular, we incorporate statistics on questions for which multiple-choice answers are presented (e.g., ethical issues when using large-scale language models). For questions that were completed by only a subset of respondents, we calculate the response rate to a particular question and the percentage of respondents who answered that question.
Interview survey
In October and November 2023, we interviewed 16 HCI researchers who had conducted research projects using large-scale language models. Each interview began with a description of the participant's most recent project andexplored how thelarge-scale languagemodel was applied inthe project.Theinterviewsfocus on the use of thelarge-scale languagemodelthroughout the research process, specific ethical considerations, how to navigate ethical considerations, the role of the IRB, the role of the ethical framework and toolkit, and incentives and responsibilities.
Participants are recruited through a combination of distribution lists, professional networks, and personal contacts, and are repeated using snowball sampling and purposive sampling. The sample includes researchers from the United States, China, Singapore, and Germany, many of whom are male and female researchers working in academia.
Interviews are conducted online via videoconference, and all sessions are conducted in English. Prior to the interview, an explanation of purpose and electronic informed consent is obtained, and participants may refuse to answer questions or request a pause in the recording at any time. Each session lasts approximately 40-60 minutes, and verbatim recorded notes and video are used for later analysis. A $30 gift card is offered to participants.
A reflective thematic analysis approach by Braun and Clarke was used to analyze the qualitative data. The research team read the transcripts of each interview multiple times, beginning with open coding for ethical considerations and issues. The team meets regularly to iteratively define themes and identify categories based on initial codes. This clarifies categories such as research process, ethical concerns, and approaches to ethical considerations, which are ultimately integrated into the results as top-level categories.
Findings: Large-scale language models used by HCI researchers in their daily work
The paper presents findings on how HCI researchers incorporate large-scale language models into their daily research work.It reveals howlarge-scale languagemodels are used ineach phase of research, including ideation, literature review, study design, data analysis, system building, evaluation, and writing papers.
In particular, they are used most frequently in the dissertation writing and research design phases, followed by project scoping, system development, data generation and collection, and analysis.Thesurveyindicates a recognition thatlarge-scale languagemodels "open up new possibilities" for research, and that "when used properly, they can be used to do cool new things that really empower."
In the early stages of the study, participantsused thelarge-scale languagemodel to identify new research questions, which were then refined into specific research goals. For example, one participantentered a broad topic area into thelarge-scale languagemodel and defined research goals based on the research questions generated. Another participantused the large-scale language model toseek advice as a career coach and explore topics to explore at the intersection of AI and cybersecurity. These activities allowed HCI researchers to quickly generate a wide variety of ideas, which is a great help in brainstorming.
In addition,the use oflarge-scale languagemodels is expanding indata-related tasks.Many researchersare taking advantage oflarge-scale languagemodels insynthesizing information from web sources, thereby greatly streamlining what would normally take a great deal of time and effort.Large-scale languagemodels arealsobeing incorporated intodata analysis processes such as qualitative data coding, plotting, and data visualization, helping to improve the quality and speed of research.
The results of this survey provide researchers with new research methodologies, illustrating the diverse uses and potential of LLMs in the HCI field.
Findings: Ethical Concerns of HCI Researchers about the Use of Large-Scale Language Models
HCI researchers experience diverse ethical challenges associated with the use of large-scale language models. According to the survey results, 30 of the researchers observed ethical issues, 10 felt uncertain, and 10 were unaware of these concerns. Issues such as data privacy, copyright, harmful outputs, copyright issues, and consent are particularly salient. These issues are common during the research design, execution, analysis, and paper writing phases.
There are also growing ethical concerns about research subjects' engagement with harmful outputs whenlarge-scale languagemodels are integrated into systems and tools that interact directly with users.The output from large-scale language models can cause disproportionate harm through social bias and stereotyping, especially to unimportant groups.
There is also a risk thatlarge-scale languagemodels introduce a new layer of reinforcing bias against people with disabilities.The seemingly authoritative but fabricated information generated by large-scale language models can affect the credibility of research. In particular,identifying false citations and discrepant article references generated bylarge-scale languagemodels poses a significant challenge to researchers.
Biases inherited fromlarge-scale language modelsare also an issueduring the research design and evaluation phases.Researchershave expressed concern about the tendency oflarge-scale languagemodels to produce homogenous content and generalize diverse perspectives. This tendency to "flatten human diversity and nuance" underscores the importance of capturing the complexity of lived experience in the context of research.
Researchers also have significant concerns about how data entered intolarge-scale languagemodels may be used and how privacy may be violated. Of particularconcern is howsensitive information, such as speech transcriptions and navigation data,is handled bylarge-scale language models. Leakage of these data to the outside world could even lead to physical harm. Researchers are also very concerned about the risk of identity theft posed by data sharing with LLM providers, noting that security flaws could result in the accidental exposure of personal information.
In addition,intellectual integrity is also an issue, as ownership of the text and visuals generated by the large-scale language model is unclear. During the collaborative process,it is difficult to distinguishwhich parts are the original contributions of the researcher and which parts aredue to the large-scale language model.There is also debate about the risk of plagiarism when large-scale language models are part of the research output. Researchershave questioned the ownership of content generated bylarge-scale languagemodels, especially during the writing phase of a paper, and appropriate credit for this should be given.
Inaddition,the reproducibility of research results using large-scale language models is particularly affected by rapid and opaque updates to the technology. If researchers do not know which version of LLM they are using, methods that worked in the past may not work well in the future. This can undermine the credibility of research results.
These concernsmake it imperative to strengthen ethical guidelines and increase transparency in HCI research as the use oflarge-scale languagemodels expands.
Findings: HCI Researchers Address Ethical Concerns of Large-Scale Language Models
Here we investigate how HCI researchers address the ethical challenges associated with the use of large-scale language models.Wethenfind that they tailor their ethical responses according to the context in which they use them and the domain of their research.
Surveys show that a large number of HCI researchers are aware of potential ethical concerns, but find these difficult to specifically identify or address. For example, when a research domain is considered low-risk,the ethical challenges oflarge-scale languagemodeling are often ignored and considered less urgent. On the other hand, in high-risk domains, more aggressive ethical measures are needed, but in practice such measures are rarely taken.
Citing specific situations and reasons for prioritizing ethical considerations, the researchers interviewedstated that the general ethical issues associated with the use oflarge-scale languagemodels do not apply to their own research. This reactive approach to the problem has led to the search for immediate solutions to the problem, but addressing the underlying issues tends to take a back seat.
Furthermore, because most HCI research is conducted in a laboratory environment, ethical issues manifest themselves differently than they do routinely, and short-term usability testingis unlikely to face the detrimental output oflarge-scale language models. Long-term studies, however, may encounter more serious problems.
Inaddition,HCI researchers regard large-scale language models as routine tools. As a result,it is clear that many participants see no need to formally report their use oflarge-scale languagemodels toresearch participants, ethics review boards (IRBs), or the broader academic community.
This surveyshows the reality of howHCI researchersperceive the ethical issues oflarge-scale languagemodels.
Summary: Integration of large-scale language models in HCI research and ethical issues
This paper reveals how human-computer interaction (HCI) researchers use large-scale language models (LLMs) and what ethical issues they face. Based on data collected through questionnaires and interviews, the paper explores how researchers integrate LLMs into their research and how they address ethical concerns they encounter.
Thestudyshows that althoughHCI researchersutilizelarge-scale languagemodels inall phases of their research, they struggle to effectively identify and address ethical issues. They have difficulty clearly identifying ethical issues for each project and addressing them appropriately.
Thepaperalso proposes a new approach to shaping ethical norms to promote the responsible use oflarge-scale languagemodels inHCI research. It encourages researchers to work with ethics review boards (IRBs) and to collaborate with policymakers and generative AI companies to develop guidelines.They also point to the need toreevaluate the informed consent process andto provide technical assistance to interrupt the supply chain oflarge-scale language models.In addition, there is a need for increased opportunities for ethics education in the HCI field and changes in academic incentives to focus on ethical concerns.
It is hoped that these efforts will contribute to a more ethical research environment in the future.
Categories related to this article