Beyond Principles: Examining the Conflicts in Implementing Responsible AI in Enterprise Structures

Introduction

The rapid evolution of artificial intelligence (AI) technologies has outpaced conventional regulatory mechanisms, leading to a significant disparity between ethical principles and practical implementation. As AI systems increasingly influence content creation and decision-making processes, concerns about algorithmic bias, privacy infringement, intellectual property, and lack of transparency have intensified. While academic institutions, think tanks, governmental bodies, and technology companies have proposed numerous ethical frameworks and guidelines, a significant gap remains between articulating these high-level principles and effectively applying them within organisations exploring AI.

In this context, non-governmental agreements and standards have emerged as potential conduits between ethical ideals and enterprise incentives. This essay posits that multi-stakeholder communities, such as trade associations and professional societies whose members include companies or professionals that develop or deploy AI, can play a pivotal role in closing this divide between responsible AI principles and practices. By leveraging their collective expertise, credibility, and influence, these organizations can help translate responsible AI into tangible practices, potentially addressing governance challenges more rapidly and flexibly than traditional governmental approaches. This essay will critically examine two papers that approach the challenge of operationalising AI ethics from reciprocal perspectives: Rakova et al.’s (2021) empirical study of organisational practices in implementing responsible AI principles and Marchant’s (2021) exploration of professional societies’ potential in AI governance.

This analysis will be further augmented by examining case studies from other industries where trade associations implemented voluntary standards with varying degrees of success. These cases offer valuable perspectives on the challenges and limitations of industry self-governance, providing cautionary advice as well as potential solutions. By extracting lessons from these examples, we can identify possible strategies and pitfalls in industry self-governance relevant to the AI sector and other emerging industries facing similar ethical challenges.

By juxtaposing these two levels of analysis — organisations and industry associations — and incorporating insights from other industries, this essay aims to provide a nuanced understanding of the mechanisms through which AI ethics principles can be effectively translated into practice. The subsequent sections will summarize and evaluate the findings of Rakova et al. (2021) and Marchant (2021). Subsequently, this paper explores case studies from other industries, extracting valuable lessons applicable to AI governance. Finally, we will synthesize these insights, offering recommendations for future research and practice in responsible AI.

Summary of “Where Responsible AI Meets Reality”

This section provides a concise overview of the key findings from Rakova et al.’s (2021) study, “Where Responsible AI meets Reality,” on the challenges and opportunities of implementing responsible AI in enterprise organizations. Through in-depth interviews with 26 AI practitioners across various roles in major AI-developing companies, the authors provide valuable insights into the challenges and opportunities associated with implementing ethical AI principles in corporate environments.

The study reveals an ecosystem characterised predominantly by reactive rather than proactive, responsible AI efforts. According to the authors, many organisations tend to react to AI ethics concerns prompted by external pressures such as media scrutiny, regulatory threats, or public outcry rather than proactively committing to responsible AI development.

A key challenge identified by Rakova et al. is the difficulty in measuring and quantifying the success and impact of responsible AI initiatives. Traditional business metrics, such as revenue growth or user engagement, fail to adequately capture the value of ethical AI practices. While there are no universally accepted responsible AI metrics, the situation is analogous to cybersecurity, where success is often measured by the absence of negative outcomes. However, tracking ethical lapses in AI is even more challenging, as the potential harms can be diffuse or long-term. This misalignment between traditional performance indicators in business and ethical considerations poses a challenge. It hinders organisations in justifying the allocation of resources to responsible AI efforts.

The study highlights the ambiguity surrounding roles and responsibilities for responsible AI within organisational structures. This lack of clear accountability can lead to fragmentation of responsibility, where no single individual or team is empowered to drive responsible AI practices forward. While some organisations have created dedicated roles such as AI ethicists or responsible AI leads, these roles often lack the authority or resources to effect significant organisational change.

Rakova et al. emphasise the misalignment between responsible AI aspirations and existing organisational incentives and culture. In many enterprise organisations, the pursuit of short-term business objectives, such as beating competitors to market or maximising user engagement, often conflicts with the longer-term goals of developing AI systems that are safe, fair, and socially beneficial. This tension is compounded by the difficulty in quantifying the business value of ethical AI practices, creating a cycle where responsible AI initiatives struggle to gain traction and demonstrate results within traditional corporate structures.

The authors also note a concerning trend of overreliance on individual champions advocating for change rather than comprehensive, organization-wide support. While passionate advocates for responsible AI can play a crucial role in raising awareness and driving initial progress, the study argues that sustainable change in enterprise environments requires the backing of leadership and the integration of ethical considerations into core business processes.

Despite these challenges, the study identifies several emerging practices that suggest a path forward for enterprise organisations seeking to translate AI principles into action. These include implementing proactive risk assessment processes, embedding ethical considerations into AI systems’ design and development lifecycle, and creating dedicated roles and teams focused on responsible AI implementation.

The study mentions that some organisations are beginning to engage with external stakeholders, including professional societies and industry associations, to develop and implement responsible AI standards. However, these collaborations are still in their early stages, and there is potential for more robust multi-stakeholder engagement to drive systemic change.

Looking to the future, Rakova et al. envision a state where responsible AI practices are anticipatory rather than reactive, fully aligned with organisational values, and redefine success to include societal welfare alongside traditional business metrics. The authors emphasise that realising this vision will require significant organisational reforms, including structural support for responsible AI initiatives, broad ethical training for employees involved in AI development and deployment, and continuous feedback loops to ensure that ethical considerations are adaptively integrated into evolving AI practices.

Evaluation of “Where Responsible AI Meets Reality”

We now turn to a critical evaluation of Rakova et al.’s (2021) study, assessing its strengths, limitations, and implications for future research. The paper’s primary strength lies in its focus on real-world experiences of AI professionals. By interviewing practitioners, the authors shed light on the challenges of implementing AI ethics in practice. This approach uncovers the structural, cultural, and operational barriers practitioners face. This qualitative approach captures nuances and contextual factors that quantitative studies might miss, offering valuable insights for organizations navigating ethical AI implementation.

The study’s analysis of prevalent, emerging, and aspirational practices offers organizations a practical roadmap for advancing their responsible AI initiatives. Particularly noteworthy is the paper’s emphasis on the need for new metrics and performance indicators for responsible AI. By highlighting the inadequacy of traditional business metrics in capturing the value of ethical AI practices, the authors make a compelling case for the development of new evaluation frameworks.

The study’s most consequential limitations are three. First, the sampling: 21 of 26 respondents are based in the United States, and most others are in English-speaking countries — this Western-centric frame undercuts the universality of the findings, particularly given that the most consequential AI deployments are increasingly happening outside the West. Second, the institutional positioning: two of the four authors are affiliated with the Partnership on AI, an organization founded and funded by Google, Facebook, Amazon, and Microsoft. This is not a disqualifying conflict, but it does call into question whether the study’s notably charitable framing of regulatory ambiguity — described as enabling for practitioners — would have survived a more independent reading. Third, the implicit consensus: the paper presupposes that “responsible AI” is a settled concept whose definition is shared across practitioners. Mittelstadt (2019) and others have made the opposite case forcefully, arguing that the discipline is too principled to fail precisely because nothing is at stake when the principles conflict. The Rakova et al. study would have been stronger with a clearer position on whose responsible AI is being implemented.

These limitations are worth dwelling on rather than enumerating exhaustively, because they shape the policy implications. A US-centric, industry-adjacent, consensus-presupposing account of responsible AI is going to systematically underweight the structural reforms that would actually shift the equilibrium — and overweight the practitioner-level interventions that are easier to fund and publish.

Despite these limitations, Rakova et al. contribute valuable insights into the practical obstacles of implementing responsible AI in enterprise settings. The study’s focus on real-world experiences provides a nuanced understanding of the barriers to ethical AI practices, which is often missing in theoretical discussions. Additionally, this research takes on added significance given that effecting change within influential AI companies may be more impactful and expedient than waiting for new legislation. Changes in the internal practices of major AI players can have far-reaching global impacts.

To address the limitations of this study and build on its findings, future research could incorporate quantitative methods, explore industry-specific variations, and analyze the evolution of responsible AI practices over time. Comparative studies of organisational structures and legal contexts could provide further insights into the factors that enable or hinder responsible AI implementation. Such research would contribute to a more comprehensive understanding of the challenges in translating AI ethics principles into organisational practice.

While Rakova et al.’s study offers valuable insights into the challenges of implementing responsible AI, it also highlights the need for further research to address its limitations. In the next section, we will examine Marchant’s analysis of the role of professional associations in AI governance.

Summary of “Professional Societies as Adopters and Enforcers of AI Soft Law”

Gary E. Marchant’s (2021) paper, “Professional Societies as Adopters and Enforcers of AI Soft Law,” examines the potential for professional organisations to promote responsible AI through mechanisms that are not legally binding. This analysis is particularly relevant to AI technologies, where regulatory mechanisms often struggle to keep pace with rapid advancements in capability.

Marchant focuses on three major organisations: the Association for Computing Machinery (ACM), the Institute of Electrical and Electronics Engineers (IEEE), and the Association for the Advancement of Artificial Intelligence (AAAI). Each of these societies has developed ethical guidelines and codes of conduct for their members, including AI practitioners.

The ACM’s Code of Ethics and Professional Conduct, updated in 2018 to address AI-specific issues, serves as a foundation of professional ethics in computing. The IEEE’s Ethically Aligned Design initiative represents a more comprehensive approach, producing a series of standards and guidelines for responsible AI development. The AAAI’s adoption of the ACM code with minor revisions illustrates the potential for collaboration across professional societies.

However, Marchant acknowledges a crucial limitation: the current lack of empirical evidence demonstrating the impact of these professional ethics codes on AI practitioners’ behaviour. This gap in evidence significantly hinders assessing the effectiveness of soft law approaches in AI.

Marchant proposes three key strategies to enhance the impact of professional society initiatives. Firstly, he advocates for more transparent enforcement of ethical codes by professional societies. This includes procedures for handling ethical violations and potentially publicising the outcomes of these processes. Secondly, Marchant suggests that employers could play a crucial role in enforcing professional society codes within their organisations. This approach recognises the significant influence that workplace environments have on professional conduct. Thirdly, the paper explores the possibility of increased professionalisation in AI. Drawing parallels with established professions like medicine and law, Marchant considers the potential benefits and challenges of implementing mechanisms such as mandatory ethics training, certification, and licensing for AI practitioners.

Marchant’s analysis emphasizes the interplay between professional and trade associations, individual practitioners, and their employers. He argues that effective governance of AI requires a multi-faceted approach that engages all these stakeholders. The paper also addresses the difficulties associated with AI governance, including the rapid pace of technological change and the global nature of AI development. These factors, he contends, create significant obstacles in establishing and enforcing consistent ethical standards across the industry.

Marchant’s work provides a comprehensive overview of the current state of professional society involvement in AI ethics and governance. By identifying the strengths and limitations of current initiatives and suggesting potential paths for improvement, the paper lays the groundwork for a more effective and impactful role for professional societies in shaping the future of AI ethics. This analysis suggests that professional societies could be pivotal in building critical mass for ethical AI adoption, particularly when working with other stakeholders such as employers and regulatory bodies.

Evaluation of “Professional Societies as Adopters and Enforcers of AI Soft Law”

Having summarized Marchant’s (2021) perspective, this section evaluates his analysis, highlighting its strengths, limitations, and implications for future research and practice. The paper offers a comprehensive overview of professional societies’ role in AI governance, providing valuable insights alongside some limitations.

The paper’s primary strength lies in its thorough analysis of recent initiatives by professional societies in the AI ethics space. Marchant’s detailed examination of ethical codes and guidelines developed by the ACM, IEEE, and AAAI provides a basis for further research and policy discussions. His assessment of the effectiveness of ethical codes raises questions about current approaches to AI ethics enforcement.

The paper’s proposals for enhancing professional societies’ role in AI governance demonstrate innovative thinking. Suggestions such as transparent enforcement mechanisms, collaboration with employers, and increased professionalisation of the AI field represent potential pathways for improving soft law approaches. However, these proposals — particularly those related to the licensing and certification of AI practitioners — require situating in the broader literature on industry self-governance that Marchant largely sidesteps. Gunningham and Rees (1997) provide the canonical institutional analysis of when industry self-regulation works and when it fails, identifying three structural conditions: a “regulatory shadow” (credible threat of state intervention), enforceable membership conditions, and transparent monitoring with public reporting. Hemphill (1992) extends this with the observation that voluntary industry codes typically reproduce the competitive interests of their largest members and tend to function as anticompetitive barriers to entry rather than as genuine ethical floors. The AI profession’s institutional structure — dominated by employer-funded membership in the ACM and IEEE, with no licensing requirement and no enforcement track record — fails on all three of Gunningham and Rees’s conditions. Marchant’s proposals would address some of this, but the paper would benefit from explicit engagement with why the same proposals (transparent enforcement, professionalisation) have produced mixed-to-poor results in adjacent fields like accounting (Sarbanes-Oxley happened because the AICPA failed to self-police) and securities analysis (the post-2002 reforms exist because the National Association of Securities Dealers failed).

While Marchant’s proposals offer potential pathways for improvement, it is crucial to also examine their limitations. The paper’s focus on U.S.-based professional societies is one such limitation, restricting its generalisability to other cultural and regulatory contexts. Given that AI governance is a global challenge, a more comprehensive examination of professional associations outside the U.S. and potential cross-cultural differences in AI ethics implementation would have strengthened the analysis. This Western-centric approach overlooks the diverse perspectives and values that should inform global AI governance.

Marchant’s paper does not fully explore the complex interactions between intraorganizational and interorganizational contexts. The paper lacks a nuanced analysis of how proposed changes in professional society practices might be implemented within companies and across industry networks. Additionally, it fails to address potential conflicts of interest when societies funded by corporate membership fees seek to enforce ethical standards affecting their members, leaving a complex power dynamic in AI governance unexplored.

Despite these limitations, Marchant’s work contributes significantly to the discourse on AI governance by highlighting the potential of professional societies as changemakers in promoting responsible AI. Future research should prioritise cross-cultural studies on AI ethics implementation and governance, examining how professional societies can effectively navigate diverse global perspectives to develop more universally applicable ethical guidelines. Additionally, empirical studies on the impact of professional society initiatives and comparative analyses of governance models across regions and industries would further enrich our understanding of effective AI governance and the efficacy of soft law approaches.

Marchant’s work, while insightful, underscores the complexities of AI governance and the need for further research. In the following section, we integrate the insights from both Rakova et al. (2021) and Marchant (2021) and draw upon lessons from other industries to explore potential paths forward.

Integrating Insights and Next Steps

This section synthesizes the key findings from Rakova et al. (2021) and Marchant (2021), exploring their implications for responsible AI governance. We also examine case studies from other industries to extract valuable lessons applicable to AI.

The papers by Rakova et al. (2021) and Marchant (2021) offer complementary perspectives on the challenges and role of professional societies in implementing responsible AI practices in corporations. Rakova et al.’s study illuminates the organisational barriers to responsible AI implementation, including difficulty measuring success, ambiguous accountability, and misaligned incentives. Marchant’s work highlights the potential of trade associations to influence ethical norms and practices through soft law mechanisms. These papers suggest that an integrated approach involving cross-organizational collaboration between AI practitioners, professional societies, and other stakeholders may offer a promising path for actualizing AI ethics principles.

Examining case studies from other industries contending with similar challenges can help us understand the potential and pitfalls of such multi-stakeholder approaches. The experiences of the Fair Labor Association (FLA) and the Interactive Advertising Bureau (IAB) in championing industry self-regulation offer valuable lessons about the potential and the constraints of multi-stakeholder governance.

The Fair Labor Association (FLA), formed in 1999 to address labour rights issues in global supply chains, stresses independent monitoring, verification processes, and capacity building rather than disciplinary measures. This approach has led to improvements in working conditions across the apparel industry (Bartley, 2007). However, it has also faced criticism for insufficient enforcement and potential conflicts of interest due to its corporate funding structure (Macdonald, 2011). The FLA’s experience underscores the importance of independent monitoring mechanisms, which could translate to the creation of third-party auditing processes for AI systems and practices. This aligns with Marchant’s proposals for more transparent enforcement of ethical codes by professional societies.

The Interactive Advertising Bureau (IAB) offers another instructive case study in industry self-regulation. The IAB achieved significant industry-wide change through non-governmental regulation by developing comprehensive ad format and user experience standards and linking compliance with access to premium ad inventory (IAB, 2014). This demonstrates the potential for industry standards to drive change when tied to commercial incentives, aligning with Rakova et al.’s (2021) findings on the importance of connecting responsible AI practices with corporate incentives.

However, the IAB case also highlights potential limitations of self-regulation. Critics argue that the standards have been heavily influenced by the IAB’s largest members (Sharma et al., 2010), mirroring Rakova et al.’s observations about the impediments to AI ethics integration in the face of financial imperatives. Moreover, Rotfeld (1992) argues that the effectiveness of advertising self-regulation is directly proportional to the threat of government regulation, suggesting that without external pressure, such programs may lack teeth.

While these case studies demonstrate the potential of multi-stakeholder approaches, they also highlight the constraints in driving self-governance in AI ethics. These include potential conflicts of interest when industry players dominate decision-making processes, the risk of creating ineffective standards without adequate enforcement mechanisms, and the challenge of balancing member interests while promoting responsible AI development. The global nature of AI development further complicates efforts to establish universally applicable standards, as different regions may have varying priorities and regulatory environments. The voluntary nature of many multi-stakeholder initiatives means that unethical actors can choose not to participate, potentially limiting the effectiveness of such approaches in addressing ethical concerns in AI.

Marchant’s (2021) proposal for increased professionalisation in AI, including potential licensing or certification requirements, could provide a stronger enforcement mechanism than voluntary standards. Such an approach could help address the limitations of voluntary participation while still maintaining the flexibility and industry-specific knowledge that self-regulation can provide. However, these measures may face significant challenges and potential resistance. They could create barriers to entry, potentially stifling innovation and diversity in the field. Companies may oppose such measures due to concerns about competitiveness in what many perceive as an AI capabilities race.

Professional societies’ dual role as promoters of ethical standards and representatives of their members’ interests presents a challenge in AI governance. This tension becomes particularly acute when trade associations are funded by the companies they aim to steer. To address this, AI professional societies could consider adopting governance structures that separate their standard-setting and enforcement functions from their value-creation roles. For instance, they could establish independent ethics boards with diverse representation from academia, civil society, and government, alongside industry members. These boards could be empowered to develop and enforce ethical standards while the broader organisation continues to provide value to members through networking, education, and advocacy. Transparent decision-making processes could enhance the societies’ credibility. However, balancing these competing interests requires consistent attention and innovative solutions.

Synthesising these insights reveals several considerations for advancing responsible AI governance. First, balancing diverse stakeholder representation with effectiveness in decision-making processes is crucial. This might involve creating tiered engagement structures in AI, with broader consultation processes feeding into more focused decision-making bodies. This approach could help address the challenges of ambiguous accountability identified by Rakova et al. while leveraging the potential of professional societies to shape ethical norms, as Marchant suggested.

Second, robust monitoring and enforcement mechanisms are essential. For AI governance, this could translate into the development of independent auditing processes for AI systems and practices, potentially overseen by multi-stakeholder bodies. However, as seen in the FLA case, these mechanisms must be designed to avoid conflicts of interest and ensure effective enforcement.

Third, aligning standards with business incentives is vital for driving widespread adoption. Future AI governance efforts should explore ways to create tangible benefits for organisations that adhere to ethical AI practices.

Fourth, governance frameworks should prioritise adaptability and learning, given AI technologies’ progressing capabilities. This approach aligns with Marchant’s suggestions for enhancing the role of professional societies in shaping ethical norms and practices in AI.

Finally, developing more sophisticated methods for measuring the impact and effectiveness of responsible AI practices remains a central challenge. Future research should focus on creating comprehensive impact measurement frameworks beyond market metrics to assess broader societal effects. This could involve developing standardised ethical impact assessments for AI systems and creating industry-wide benchmarks for responsible AI practices. Such Responsible AI metrics could include fairness scores for AI outputs across different demographic groups, the proportion of AI decisions accompanied by clear explanations, compliance rates with data protection regulations, and alignment scores with company values and industry guidelines. Such quantifiable measures can provide tangible goals for companies, facilitate meaningful comparisons across the industry, and help track progress in implementing responsible AI practices over time.

Conclusion

This essay has explored the challenges of implementing responsible AI principles by examining Rakova et al.’s (2021) study on organizational barriers and Marchant’s (2021) analysis of professional societies’ role in AI governance. By synthesizing these perspectives with case studies from other industries, we’ve illuminated the complexities of operationalizing AI ethics and the roles various stakeholders can play.

Rakova et al.’s work highlights key obstacles in implementing responsible AI practices, including difficulties in quantifying ethical impact and misalignment between ethical aspirations and organizational incentives. Marchant’s analysis complements this by exploring the potential of soft law approaches through professional societies, proposing more transparent enforcement and employer collaboration to bridge the gap between principles and practice.

Brief case studies from organizations like the Fair Labor Association and Interactive Advertising Bureau underscore the challenges of balancing diverse interests and maintaining independent oversight in multi-stakeholder frameworks. They emphasize the need for comprehensive metrics that evaluate broader societal impacts beyond market performance.

The path to responsible AI demands continuous learning, adaptation, and collaboration. By learning from past successes and failures, building diverse coalitions, and maintaining rigorous research and testing, we can work towards AI technologies that respect human rights, promote fairness, and ensure accountability. The insights offered here serve as a foundation for this ongoing effort. Through collective action, informed by both history and aspiration, we can shape a more equitable AI-enabled future.

References

• Bartley, T. (2007). Institutional Emergence in an Era of Globalization. American Journal of Sociology, 113(2), 297-351.
• Gunningham, N., & Rees, J. (1997). Industry self-regulation: An institutional perspective. Law & Policy, 19(4), 363-414.
• Hagendorff, T. (2020). The ethics of AI ethics: An evaluation of guidelines. Minds and Machines, 30(1), 99-120.
• Hemphill, T. A. (1992). Self-regulating industry behavior: Antitrust limitations and trade association codes of conduct. Journal of Business Ethics, 11(12), 915-920.
• IAB. (2014). Self-Regulatory Program for Online Behavioral Advertising.
• Kaye, K. (2019, April 30). Big Tech’s AI Ethics Nonprofit Signals Business-Aligned Motives. RedTail.
• Macdonald, K. (2011). Fair Labor Association Structure and activities. In D. Held & T. Hale (Eds.), Handbook of Transnational Governance Innovation (pp. 243-251). Polity Press.
• Marchant, G. E. (2021). Professional societies as adopters and enforcers of AI soft law. IEEE Transactions on Technology and Society, 2(4), 183-191.
• Mittelstadt, B. (2019). AI Ethics — Too principled to fail? SSRN Electronic Journal.
• Rakova, B., Yang, J., Cramer, H., & Chowdhury, R. (2021). Where responsible AI meets reality: Practitioner perspectives on enablers for shifting organizational practices. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 1-23.
• Rotfeld, H. J. (1992). Power and limitations of media clearance practices and advertising self-regulation. Journal of Public Policy & Marketing, 11(1), 87-95.
• Sharma, L. L., Teret, S. P., & Brownell, K. D. (2010). The food industry and self-regulation: standards to promote success and to avoid public health failures. American Journal of Public Health, 100(2), 240-246.