What We Know about Responsible AI Practices in Industry: A Half Decade of Empirical Research

This paper synthesizes current knowledge about Responsible AI (RAI) practices in industry through a literature review of 161 empirical studies published between 2019 and August 2025. The review focuses on research that engages industry practitioners via methods such as interviews, surveys, workshops, and ethnographies. The synthesis reveals both meaningful progress and persistent challenges in industry RAI practice.

Progress and Positive Trends:

• Increased Awareness: Practitioner awareness of RAI's existence and importance has grown over time. More recent studies show practitioners, even those not in dedicated RAI roles, are more aware of and able to articulate the importance of RAI. This is supported by quantitative survey evidence and qualitative interview quotes. Factors shaping awareness include regional regulations (e.g., GDPR), company policies, cultural norms, and demographic factors.

• Professionalization of RAI Roles and Practices: RAI activities have become more formalized and standardized, shifting from ad-hoc advocacy by individuals to dedicated roles such as RAI software engineers, AI ethics specialists, and RAI policy analysts. This is illustrated in three areas: (1) AI fairness testing has moved from being neglected or reactive to being more proactive and systematic, with dedicated pipelines; (2) model explainability and interpretability work is becoming a deliberate and repeatable organizational practice; (3) RAI documentation has become more standardized and institutionalized, sometimes even tied to KPIs.

• Adoption of RAI Interventions: Adoption of RAI policies, processes, and tools has steadily increased. Companies are adopting formal internal policies, practitioners are using guidelines and documentation (e.g., model cards, datasheets), and computational toolkits (e.g., Fairlearn, AIF360) are being actively adopted and adapted into daily operations.

• External Stakeholder Engagement: There are attempts to engage external stakeholders, including end users in testing and auditing, external domain experts (e.g., clinicians, policy experts), and progress in managing third-party data annotators with more awareness of diversity and working conditions.

Persistent Challenges:

• Lack of RAI Knowledge and Training: Despite increased awareness, practitioners often lack sufficient technical and socio-technical expertise. This leads to misuse of RAI tools, misinterpretation of results, and hinders cross-role communication. The rise of generative AI has exacerbated these gaps.

• Organizational Dynamics and Misalignment: RAI work is often not prioritized compared to AI development, leading to significant human resource, time, and budget constraints. This is particularly challenging in startups. These constraints can demotivate practitioners and lead to check-the-box compliance exercises.

• Lack of Tailored RAI Interventions: Many RAI interventions are too abstract and not tailored to specific domains, applications, or stages of the AI development pipeline. They often focus on evaluating model outputs, leaving other stages (e.g., data annotation, problem formulation) unsupported.

• Resource and Structural Barriers to External Engagement: Meaningful external engagement is hindered by resource limitations, procedural and organizational constraints (e.g., privacy compliance, recruitment difficulties), and structural issues like the supply-chain-like structure of generative AI development, which creates an accountability horizon with blurred lines of responsibility.

Opportunities and Implications:

• For Researchers: The paper calls for moving beyond cataloging known problems toward designing, deploying, and evaluating RAI interventions as end-to-end sociotechnical systems embedded in real organizational workflows. It highlights understudied areas like generative AI, frontier AI labs, and the need for greater demographic and geographic diversity in research. It also advocates for more longitudinal and ethnographic approaches and co-designing interventions with practitioners.

• For Practitioners: Organizations should treat RAI capability as core organizational infrastructure, investing in ongoing training and embedding RAI into everyday work practices. They should also institutionalize external engagement with formal structures like expert panels and algorithm review boards.

• For Policymakers: Regulation should prioritize implementability, focusing on concrete lifecycle stages and organizational workflows. It should support the buildout of organizational capacity and prioritize substantive accountability over procedural compliance to avoid symbolic compliance. The paper also notes that the anticipation of regulation can shape organizational behavior, and regulation should be designed to evolve alongside the technology.

Improvements to AI Systems Based on This Paper:

1. Embedded RAI Training and Contextual Help: Build AI systems with built-in, role-specific training modules and just-in-time explanations. The system can detect when a practitioner misuses a fairness tool or misinterprets an explainability output, then automatically provide corrective guidance, domain-specific examples, and cross-role translation (e.g., explaining technical results to policy teams). This addresses the persistent lack of RAI knowledge and reduces misuse.

2. Lifecycle-Stage-Aware RAI Interventions: Design AI systems that provide tailored RAI support at every stage—not just model evaluation. The system can prompt users during problem formulation, data annotation, and deployment with stage-specific checklists, risk assessments, and mitigation templates. For example, it can flag potential bias in training data collection before modeling begins, or suggest external expert review during problem scoping.

3. Adaptive Resource-Aware RAI Workflows: Create AI systems that dynamically adjust RAI processes based on organizational constraints (e.g., startup size, budget, timeline). The system can prioritize the most impactful RAI actions given limited resources, automate repetitive documentation, and generate lightweight audit trails that satisfy regulatory needs without overwhelming small teams. This prevents check-the-box compliance by focusing on substantive, feasible actions.

4. Structured External Stakeholder Integration: Build AI systems with built-in mechanisms for external engagement, such as interfaces for end-user feedback during testing, secure channels for third-party auditors, and templates for clinician or policy-expert review. The system can manage consent, anonymize data, and schedule structured feedback loops, overcoming privacy and recruitment barriers. It can also clarify accountability by tracking which decisions were influenced by which external inputs.

5. Proactive Regulatory Alignment and Reporting: Implement AI systems that continuously monitor evolving regulations (e.g., GDPR, upcoming AI acts) and automatically flag compliance gaps in workflows. The system can generate substantive, evidence-based reports that demonstrate real RAI effort (e.g., logs of fairness tests, documentation updates, stakeholder consultations) rather than symbolic compliance. It can also simulate regulatory scenarios to help organizations prepare for future rules.

6. Generative AI-Specific RAI Guardrails: For generative AI systems, embed safeguards that address the accountability horizon problem. The system can trace data provenance, annotate third-party content, and provide clear responsibility boundaries for each component (e.g., foundation model vs. fine-tuned layer). It can also include specialized training for practitioners on generative AI risks (e.g., hallucination, bias amplification) and offer tools for continuous monitoring of outputs in production.

7. Longitudinal and Ethnographic Feedback Loops: Design AI systems that support long-term RAI evaluation by automatically logging practitioner decisions, tool usage, and outcomes over time. The system can generate longitudinal reports that reveal how RAI practices evolve, identify recurring failure points, and suggest iterative improvements. This enables organizations to move beyond one-time audits and continuously refine their RAI capabilities.

What the Improved AI System Can Do:

• Reduce RAI tool misuse and improve cross-team communication through adaptive, in-context training.

• Ensure RAI is applied across the entire AI lifecycle, not just at evaluation, by prompting and guiding at each stage.

• Function effectively in resource-constrained environments by prioritizing high-impact, feasible RAI actions.

• Facilitate meaningful external audits and stakeholder input while maintaining privacy and accountability.

• Stay ahead of regulatory changes and produce substantive compliance evidence, avoiding symbolic gestures.

• Manage the unique risks of generative AI by clarifying accountability and monitoring outputs in real time.

• Provide organizations with longitudinal insights to continuously improve their RAI practices and adapt to new challenges.

Responsible AI (RAI) has become a central concern for technology companies, regulators, and the public. How industry practitioners interpret, implement, and sustain RAI work directly shapes the design and deployment of AI systems. As empirical scholarship examining RAI practices in industry has rapidly expanded, findings are dispersed across studies that focus on different roles, organizational contexts, and interventions. This work synthesizes current knowledge through a literature review of 161 empirical studies spanning six years, each engaging industry practitioners via interviews, surveys, workshops, ethnographies, and other methods. Our synthesis reveals both meaningful progress and persistent challenges in industry RAI practice. Practitioner awareness has increased, RAI activities have become more professionalized, and interventions such as toolkits and guidelines are more widely adopted. At the same time, practitioners continue to face substantial barriers, including limited training, uneven organizational support, and a lack of interventions tailored to day-to-day work practices. By consolidating and organizing these findings, we provide a more complete account of industry RAI than any single study to date. We conclude by discussing implications for RAI researchers, practitioners seeking to adopt effective practices, and policymakers aiming to ground governance efforts in the realities of industry contexts.