Investigating Responsible AI for Scientific Research: An Empirical Study

I. INTRODUCTION

• Responsible and ethical AI development and deployment should uphold fairness, accountability, transparency, and inclusivity while minimizing risks and negative consequences.The paper links these practices to trust among users and stakeholders and to sustainable, equitable AI advancement.Source: §I · p. 1
• Responsible AI guidelines face practical implementation gaps, inconsistent interpretation, and a lack of diverse perspectives, proven methods, shared norms, and robust legal accountability.Guidelines may also overlook business practices and political economies surrounding AI systems.Source: §I · p. 1
• Operationalizing ethical principles means translating them into actions, processes, and practices integrated into organizational systems, tools, and workflows.The passage associates effective operationalization with greater transparency and trust among users and stakeholders.Source: §I · p. 1
• The exploratory study examines responsible and ethical AI practices at CSIRO, Australia's national science research agency.CSIRO works across fields including agriculture, climate science, energy, health, and information technology.Source: §I · pp. 1–2
• CSIRO research teams face rising ethical expectations and pressure from emerging regulations and societal expectations to deliver increasingly ethical AI.The introduction says the study is timely because these pressures are expected to continue.Source: §I · p. 2
• The study uses a survey and follow-up interviews to identify awareness and understanding gaps and review current practices for responsible AI development and deployment.Participants included scientists, project and executive managers, software engineers, and user experience designers; selected AI for Missions projects address large-scale challenges such as plastic waste, drought resilience, and biosecurity.Source: §I · p. 2

II. BACKGROUND AND MOTIVATION

• Research on organizational responsible AI highlights practical implementation challenges, the influence of organizational culture and structures, and a need for empirical insights and workable methods.Prior work also debates approaches that may be too flexible or too strict for differing contexts.Source: §II · p. 2
• The literature identifies limited understanding of how organizations grasp and apply ethical AI concepts, alongside concerns about isolated AI-risk management and decisions made without practical AI knowledge.It also notes the challenges of operationalizing principles amid weak norms and accountability mechanisms and overlooked business and political-economic factors.Source: §II · p. 2
• Surveying organizational understanding can inform tailored strategies, governance models, and training programs for responsible and ethical AI initiatives.The passage names bias, fairness, transparency, and explainable and responsible AI as areas the approach may address.Source: §II · p. 3

III. RESEARCH METHODOLOGY

• The exploratory mixed-method study surveyed CSIRO scientists, managers, engineers, and designers and interviewed participants involved in AI for Missions projects.The six Missions challenges have a national focus, and data for both studies were collected in the second half of 2022.Source: §III · p. 3

A. SURVEY

• The 39-question survey examined project context, inclusive and ethical AI, risk assessment, and awareness of existing AI ethics frameworks; responses were analyzed with frequency and thematic analysis.Question-level response counts ranged from 23 to 35 because participation was voluntary.Source: A. SURVEY · p. 3
• The survey results are a snapshot from a small group of consenting participants, not a definitive study of AI use across CSIRO.The authors identify voluntary participation and a low response rate as limitations.Source: A. SURVEY · p. 3

B. Interview

• Interviews sought project context, potential ethical risks, and guidance for integrating responsible, ethical, and inclusive AI into design, development, deployment, and governance..Source: B. Interview · p. 3
• The 28 purposively selected interviewees represented AI projects across scientific domains including manufacturing, drug discovery, health monitoring, earth observation, and energy management.Participants were selected for expertise and involvement in designated projects.Source: B. Interview · pp. 3–4
• Interviews lasted 60 to 90 minutes, followed a structured 20-question guide, were transcribed verbatim, and were de-identified before analysis.The de-identification removed references to participants and specific project details.Source: B. Interview · p. 4
• Interview analysis compared themes manually coded by a researcher with themes generated by GPT-4, then refined them through discussions among three researchers.The comparison considered alignment and divergence to converge on themes grounded in project contexts.Source: B. Interview · p. 4
• Cross-project analysis identified eight themes concerning ethical and responsible AI in scientific deployments.The analysis examined recurring patterns and variations across project contexts.Source: B. Interview · p. 4

A. Survey

• More than 60% of 25 project responses were under development, and over 50% of those projects used historical data.Customer interviews were the dominant activity for building shared understanding, and people were predominantly involved in advisory roles.Source: A. Survey · p. 4
• 40% of respondents reported no understanding of Inclusive AI, 48% did not think their systems were inclusive, and 68% did not know or think their data needed greater representation.The passage suggests some scientific data may concern plants and animals rather than humans, which could make teams less likely to prioritize ethical AI approaches.Source: A. Survey · p. 4
• 70% of respondents had not used an AI Ethics Framework, 17% were unaware of its existence, and 83% had not used additional ethical-AI design resources.A small percentage conflated AI ethics frameworks with Human & Animal Research Ethics policies.Source: A. Survey · pp. 4–5
• Close to 40% of respondents were unaware of project ethical risks, 40% used no risk-assessment techniques, and 20% were unaware of any project assessment.Almost 48% of projects had undergone no risk assessment; privacy protection and information security were the top reported risks, while fairness, bias, accountability, and explainable AI were not top risks.Source: A. Survey · p. 5
• 68% did not know or think their data needed greater representation; respondents also seldom worked on missing data, and bias detection, explainable AI, and accountability were not main priorities.The passage identifies representative data as relevant to fairness and reducing bias, and missing data as a source of uncertainty that can affect decision validity.Source: A. Survey · p. 5
• Of 25 responses, 76% of projects did not replace human activity; among human-AI collaborations, 32% were advisory and 24% assistive.Advisory systems present options for human decisions; assistive systems perform basic processes while requiring significant human input.Source: A. Survey · p. 5
• Model training and new ML algorithm development dominated project processes, and black-box approaches were more prevalent than explainable approaches.The authors note that explainable AI techniques could provide greater transparency into model predictions or decisions.Source: A. Survey · p. 5
• Respondents reported gaps in inclusive-AI understanding, ethics-framework and resource use, risk assessment, and attention to missing data and bias detection.The survey summary describes areas for improvement across ethical and inclusive AI practices and development processes.Source: A. Survey · pp. 4–5

1) Ethical and Responsible AI

• Participants’ reluctance to conduct risk assessments reflected perceived lack of necessity and concern that assessments could slow research progress.Teams were encouraged to frame ethical considerations as value-adding opportunities rather than liabilities.Source: B. Interview · p. 5
• Participants reported limited understanding of broad AI ethics principles and wanted guidance tailored more directly to their projects.Concerns included balancing explainability with reliability and safety, as well as fairness, privacy, and societal impacts.Source: B. Interview · p. 5
• In agriculture and food projects, teams linked fairness to diverse training data and equitable distribution of benefits across the supply chain.One team sought to include seed growers from varied settings; another used AI Ethical Principles to evaluate traceability systems.Source: B. Interview · pp. 5–6
• Livestock disease testing teams emphasized legal needs for comprehensible AI, while water forecasting teams highlighted disclosing forecast uncertainty.Open data repositories also limited control over data use and potential misuse.Source: B. Interview · p. 6
• Teams found AI systems harder to monitor and fine-tune than physics-based models because of AI’s black-box nature.Participants emphasized identifiable human oversight of system outcomes.Source: B. Interview · p. 6

2) Risk Assessment and Strategic Management:

• Participants described risk registers and continuous system monitoring, but some viewed existing assessment procedures as administrative burdens.Monitoring was intended to identify inaccuracies, biases, or unforeseen consequences.Source: 2) Risk Assessment and Strategic… · p. 6
• A farming project balanced environmental benefits against potential biodiversity loss, illustrating competing risks within an AI application.The water-quality forecasting project emphasized distinguishing project objectives from ethical AI practices to avoid false security and unanticipated consequences.Source: 2) Risk Assessment and Strategic… · p. 6
• Participants said strategic communication and tangible ethical actions can shape public perception and support responsible AI deployment.They also noted AI’s potential contributions to societal and environmental well-being when used responsibly.Source: 2) Risk Assessment and Strategic… · p. 6

3) Managing Data and Addressing Uncertainty:

• Participants stressed that missing or incomplete data and measurement uncertainty can affect AI reliability and model effectiveness.Data work included collection, cleansing, pre-processing, and integration across sources such as historical records, satellite observations, and hyperspectral imaging.Source: 3) Managing Data and Addressing… · p. 6

4) Collaboration and Engaging Stakeholders:

• Participants emphasized collaboration among domain experts, industry partners, and stakeholders, including managing stakeholder apprehensions about new technologies.Shared knowledge from collaboration was described as a factor in project success; teams also valued diverse stakeholder perspectives.Source: 4) Collaboration and Engaging… · p. 6
• Participants called for adaptable strategies as AI projects face changing conditions and broader societal consequences.They highlighted risks, stakeholder perspectives, and possible socioeconomic effects, with attention to safety, consistency, and public confidence.Source: 4) Collaboration and Engaging… · p. 7

6) Regulatory Compliance and Ethical Guidelines:

• Participants emphasized aligning AI initiatives with ethical principles and regulatory requirements.They described ethical benchmarks as central to responsible AI work.Source: 6) Regulatory Compliance and Ethical… · p. 7

7) Democratization of AI Technology

• Participants warned that initial AI adoption costs could favor major entities and called for equitable technology dissemination.The interviews’ broader themes also included ethical responsibility, risk assessment, data management, collaboration, and societal effects.Source: 7) Democratization of AI Technology · p. 7

A. Synthesis of results from Survey and Interview

• The survey and interviews identify challenges in responsible and inclusive AI development and deployment at CSIRO.The findings are synthesized as insights into current practices and challenges.Source: A. Synthesis of results from Survey… · p. 7

B. Challenges

• Survey findings point to gaps in inclusive AI understanding, risk assessment, and use of ethical resources, alongside emphasis on model training over missing-data and bias detection.Interview findings call for comprehensive assessment and attention to human values, fairness, privacy, security, reliability, transparency, contestability, and accountability.Source: B. Challenges · p. 7

C. Comparison to existing literature

• Low awareness and use of ethical frameworks and risk-assessment techniques at CSIRO align with prior literature on limited practical guidance and weak enactment of AI principles [10].The discussion also connects limited rigorous risk assessment with calls for strong governance controls.Source: C. Comparison to existing literature · p. 7

D. Areas of improvements

• Proposed improvements include greater framework awareness, risk assessment, regulatory attention, stakeholder collaboration, and a human-centric approach to AI ethics.The discussion also recommends treating ethics as a value proposition and communicating ethical AI’s societal and environmental contributions.Source: D. Areas of improvements · pp. 7–8

E. Trade-off analysis

• Infrequent use of ethical frameworks and risk assessments was linked to competing resource demands or pressure for rapid project progress; limited inclusive-AI awareness may reflect gaps in expertise or training time.Interview findings concurred with the survey and point to treating ethical considerations as potential value additions rather than only costs of time and resources.Source: E. Trade-off analysis · p. 8

F. Future of AI for Science

• AI co-pilots, including Generative AI and Large Language Models, are expanding in scientific research, heightening the need for organizational ethics frameworks and risk assessments.The passage names fairness, transparency, contestability, and accountability as part of this need.Source: F. Future of AI for Science · p. 8
• Organizations are urged to tailor ethical AI principles to their circumstances, including project nature, stakeholders, and broader societal implications.The passage describes this adaptation as requiring changes to business-process modeling and reengineering.Source: F. Future of AI for Science · p. 8

VI. CONCLUSION AND FUTURE WORK

• The study concludes that scientific research needs greater emphasis on awareness, understanding, and implementation of ethical, responsible, and inclusive AI principles, and calls for further research on operationalization.Its findings contribute to ongoing discussion of AI’s role in shaping the future.Source: §VI · p. 8

A. Participant’s Demographic Questions

• Participants were asked how many CSIRO AI/ML projects they had worked on, their project strategy involvement, role, and level.Source: A. Participant’s Demographic Questions · p. 9

B. Project Context Questions

• Project-context questions cover project purpose, funding partners or customers, their sectors and business areas, and respondents’ technology-development involvement.Source: B. Project Context Questions · p. 9

C. Project related Questions

• Project-related questions ask what problem the AI addresses, who experiences it, the system type, stakeholder understanding, users, human-AI relationship, and data used.Source: C. Project related Questions · p. 9

D. Inclusive AI

• Inclusive-AI questions ask respondents to explain the concept, assess their own system’s inclusivity, and consider dataset representation and representativeness needs.Source: D. Inclusive AI · p. 9
• Uncertainty questions cover data sources of uncertainty, possible human replacement, reporting uncertainty, and effects on decision quality, reliability, and trust.Source: D. Inclusive AI · p. 9

F. Humans

• The questionnaire cross-references human-relationship and replacement questions with its risk-assessment section, which asks about awareness of ethical risks and techniques for managing them.Source: [Question 15 & 23] G. Risk Assessment · p. 9
• Data questions ask about management processes and techniques used to reduce ethical risk in AI systems or components.Source: [Question 15 & 23] G. Risk Assessment · p. 9

I. Process

• Respondents are asked which algorithm and model development processes they participate in and what machine-learning approach their project uses.Source: §I · p. 9

J. Ethical AI

• Ethical-AI questions cover output monitoring, use of ethics frameworks or guidelines, risk assessment, responsible-AI approaches, and barriers to adoption.Source: J. Ethical AI · p. 9

K. Follow up Question

• Survey participants are asked whether they would agree to be contacted for a follow-up interview.Source: K. Follow up Question · p. 9

• Interview questions begin with respondents’ project roles and target industry sectors, then identify business areas and AI system types.Source: Appendix B · pp. 9–10
• The questionnaire asks respondents to describe the problem, affected people, stakeholders, end users, and the human-AI relationship.Source: Appendix B · p. 10
• Respondents classify systems as assistive, advisory, autonomous, or discovery-oriented, then discuss responsible-AI practices, ethical-risk assessment, and data types.Source: Appendix B · p. 10
• Data questions cover historical or synthetic data and respondents’ involvement in collection, cleaning, processing, governance, and other management activities.Source: Appendix B · p. 10
• Follow-up prompts ask about uncertainty sources, output monitoring, data-planning and enterprise-risk tools, understanding ethical principles, and how responsible AI may support project objectives.Source: Appendix B · p. 10