ESG Echoes in AI Ethics: Drawing Lessons from Corporate Sustainability's Missteps for Effective Algorithmic Accountability

The widespread adoption of Environmental, Social, and Corporate Governance (ESG) principles since the mid-2000s, followed by the burgeoning field of modern AI ethics, marks a critical juncture in corporate responsibility and technological governance. While both fields propound transformative benefits, there is an inherent tension when balancing their ethical aspirations with corporate interests. The article explores the similarities between these movements and the insights that can be gleaned from ESG disillusionment to understand the dynamics of AI ethics and algorithmic accountability.

ESG and AI Ethics: Parallels and Pitfalls

The struggle of the ESG and AI ethics industrial complexes to balance ethical aspirations with profit motives is increasingly evident. ESG critics derisively label the movement as ‘greenwashing,’ a marketing gimmick pandering to environmentally-conscious investors (Yu et al., 2020). The evidence overwhelmingly indicates that the ESG movement has not slowed climate change: despite 98.8% of S&P 500 companies publishing ESG reports as of June 2023, the last ten years have ranked as the ten warmest years on record (Center for Audit Quality, 2023; NOAA, 2023). A similar dichotomy is present in AI ethics, where corporate narratives often prioritize public perception and profitability over consequential changes.

Elettra Bietti’s analysis of ‘ethics washing’ in AI critiques the tech industry’s use of ethical language as a tool for self-regulation and self-promotion, often yielding superficial commitments that favor the interests of industry stakeholders more than societal good (Bietti, 2021). This concept calls attention to the instrumentalization of ethics language, conflating the marketing of ethics with genuine ethical practice by corporations. Bietti argues for a more nuanced approach to tech policy, arguing that ethics should go beyond performative rhetoric about ‘fixing’ algorithms. Instead, she suggests that we should consider the ways in which algorithms can further existing social inequities and the settings when we should avoid using them altogether. The status quo of ethics as performance in private industry often results in commitments adopted more for appearance and misdirection than actual stewardship. Bietti’s insights are critical in examining the authenticity and impact of initiatives in AI ethics and ESG.

The tendency to rely on market-driven solutions to market-driven problems has proven inadequate in addressing complex social-ecological and socio-technical challenges. The case of American Airlines exemplifies this discrepancy: the world’s largest airline made it to the Dow Jones Sustainability Index despite emitting 49 million metric tons of carbon dioxide last year (American Airlines, 2023). This inclusion raises questions about the efficacy of ESG metrics in reflecting environmental stewardship.

Similarly, many decisions that OpenAI’s leadership team made as it evolved from a nonprofit without financial obligations to a for-profit subsidiary worth $86 billion mirror this contradiction. Despite safety concerns within OpenAI, the release of generative AI products like the chatbot ChatGPT in November 2022 and the GPT Store in January 2024 illustrates the difficulty in maintaining foundational ethical principles when faced with huge financial incentives (Weise et al., 2023). Together, these examples underscore how market-driven dynamics can lead firms to reveal large quantities of ethics data to seem transparent but perform poorly in these aspects.

The influence of commercial pressures is visible in Google and Meta’s rushed deployments of AI products following ChatGPT’s release. For instance, Google fast-tracked AI chat and cloud offerings despite internal acknowledgment of “stereotypes, toxicity and hate speech in outputs” from their generative AI chatbot (Weise et al., 2023). Similarly, Meta accelerated the release of its LLaMA chatbot despite internal debates over its open-source nature and potential misuse (Weise et al., 2023).

In October 2020, Tesla released its Full Self-Driving (FSD) beta to a select group of consumers. This represented a significant step in the development of autonomous vehicle technology, and Tesla rushed to take it. Their marketing portrayed a near-autonomous system, despite it being an SAE Level 2 driver-assist feature requiring constant supervision. Their website describes Full Self-Driving capability as “The system is designed to be able to conduct short and long distance trips with no action required by the person in the driver’s seat”. This gap between company messaging and customer expectations also was impacted by actual performance. The FSD capability failed to operate appropriately in common driving situations, resulting in 273 autopilot-related crashes in 2021 (SAE, 2023). Drivers using the FSD capability admitted to increasingly unsafe changes in driving behavior over time such as falling asleep at the wheel with weights to keep FSD operating (Nordhoff et al., 2023). On February 16, 2023, Tesla recalled 362,758 vehicles with the FSD beta software.

These safety issues and recalls associated with FSD have eroded public trust in Tesla’s autonomous driving capability roadmap. Many consumers feel misled by the company’s marketing, which promised more advanced and reliable features than what was actually delivered (SAE, 2023). Tesla’s case illustrates the dangers of prioritizing market dominance over thorough ethical evaluation.

This trend of rapid deployment in the AI arms race echoes Volkswagen’s 2015 emissions scandal, in which the German automaker deceived US environmental regulators in order to inflate their performance metrics and falsify eligibility for energy-efficient subsidies. Between 2008 and 2015, Volkswagen programmed their diesel engines to activate emissions controls only during laboratory emissions testing. However, just one week before the US Environmental Protection Agency announced their investigation confirming that Volkswagen willingly violated the Clean Air Act, Volkswagen was named the “world’s most sustainable automotive group” by the Dow Jones Sustainability Index (Volkswagen of America, 2015). Volkswagen’s emissions deception led to an estimated 1,200 premature deaths in Europe (Chu, 2017), nearly 1 million tons of extra nitrogen oxides emitted per year (Mathiesen & Neslen, 2015), and its lauding as the leader of automotive environmental stewardship eroded trust in corporate sustainability commitments.

The parallel to current AI development is direct, not analogical. Volkswagen built one technical system for benchmarks and another for the road; AI labs deploy models that perform safely in red-team evaluations and unpredictably at scale. Volkswagen’s executives announced “clean diesel” while their engineers knew the cars were polluting; OpenAI’s leadership announced safety reviews while shipping GPT-4 in months. In both cases the gap is the same: a marketing claim attached to a system the company knows behaves differently in production. Volkswagen’s deception was prosecuted only because emissions testing made the gap measurable. AI’s equivalent gap — between published evals and real-world output — has no standardized test, no regulator, and no penalty. The case of Volkswagen’s marketing of ‘clean diesel’ cars, reveals a rush to market that mirrors the recent actions of AI companies. Across AI ethics and ESG, such examples demonstrate a troubling pattern: companies often sideline ethical obligations (despite claiming otherwise) in pursuing market dominance, resulting in significant societal repercussions. Both sectors demonstrate a tendency to prioritize technological advancement and market presence over thorough ethical evaluation, raising concerns about the integrity of their initiatives. The similarity in these patterns across both sectors highlights the critical need for effective ethical oversight. It becomes evident that the belated lessons learned from the automotive industry’s ethical lapses are imperative guides for the responsible development of AI technologies.

The Imperative for Algorithmic Accountability

The ESG movement’s shortcomings, rooted in its dependence on trivial rather than enforceable measures, along with insights from advocates of algorithmic accountability, provide guidance to avert similar pitfalls in the AI industry. Frank Pasquale’s (2019) notion of a ‘second wave’ of algorithmic accountability offers a framework for reorienting the AI ethics discussion. Moving beyond the limited scope of the AI ethics’ incremental improvements to existing systems, algorithmic accountability calls for a comprehensive reassessment of AI systems’ deployment. Central to this framework is a reexamination of the necessity of specific AI systems, coupled with a push for structural reforms in their governance. This perspective aligns with an important lesson from the ESG movement — the insufficiency of self-regulation in industries driven by profit motives — and offers a proactive step towards embedding ethical considerations at the core of AI development.

Nissenbaum and colleagues’ exploration of accountability in algorithmic society is vital for understanding the ethical implications of AI. She identifies multiple barriers to accountability in the tech industry, including the diffuse responsibility among contributors, the inevitability of errors in AI systems, shifting blame onto technology, and ownership without liability (Cooper et al., 2022; Nissenbaum, 1996). This perspective is crucial for understanding how the barriers to accountability have evolved with technology. Cooper et al. (2022) also detail contemporary interventions that, while necessary for developing actionable notions of accountability, are insufficient on their own. They assert that we must go beyond post hoc impact assessments and transparency reports to build a culture of accountability. They suggest measures including designating multiple actors accountable, auditing requirements throughout the AI development pipeline, and strict liability frameworks with measures to identify accountable parties and address algorithmic harms directly (Cooper et al., 2022).

The insights from Pasquale (2019) and Cooper et al. (2022), combined with Nissenbaum’s (1996) foundational work on accountability barriers, shed light on potential regulatory and institutional mechanisms to address the societal impacts of AI. The shift to a lasting culture of accountability necessitates concrete actions set forth in the next section of this essay. By doing so, AI ethics can transcend theoretical discourse, becoming an integral part of AI governance.

Framework for Enforceable AI Regulation

ESG and AI ethics initiatives can effectively function as heat shields, enabling corporations to maintain their power and capital while creating an illusion of responsibility. Instead, this essay argues for a decisive shift towards enforceable legislative standards that directly confront and penalize dangerous corporate behaviors. There is evidence that strong regulations and disincentives can help mitigate a race to the bottom in the private sector. Criminal penalties for unethical AI practices, akin to those in the financial sector, can deter malicious use. For example, prosecuting two of the most influential crypto exchange founders has proven more effective in cleaning up the cryptocurrency space than any well-intentioned ESG initiative (Khalili, 2023). Clearview AI faced significant regulatory scrutiny and legal challenges in Europe and the United States for violating privacy laws, including the EU’s General Data Protection Regulation (GDPR). The company’s practices of scraping images without consent and lack of transparency were deemed non-compliant with privacy regulations. Litigation in the U.S. alone has been so financially burdensome for Clearview AI that the company is likely to go bankrupt before the case reaches trial (Hill, 2024). The U.K. has also fined an additional $9.4M and ordered that any data of residents be deleted (Hart, 2022). Learning from ESG history and Clearview’s experience, AI companies can emphasize the importance of genuine transparency, robust ethical guidelines, and meaningful stakeholder engagement to build trust and accountability while mitigating their shared challenges of incomplete disclosures, ethical ambiguities, and superficial compliance.

Consequently, enabling the dissemination of dangerous or hateful content by AI models should result in criminal penalties. Taxing AI’s extractive business practices — training models on copyrighted material without permission, scraping public data into private datasets — may also prove fruitful, modeled on carbon emissions taxes that price externalities the market would otherwise ignore. Only through such tangible measures can we steer AI development towards a beneficial and safe course.

The examples of recklessness in sustainability and AI ethics from Volkswagen and the developers of foundation models are akin to non-cooperative game theory, where the Nash equilibrium results in corporations prioritizing rapid innovation and market dominance to stay competitive. We need systematic external interventions that deter detrimental corporate behaviors to shift this equilibrium towards more responsible outcomes. Therefore, this framework must incorporate the following elements:

Criminal penalties for irresponsible or damaging AI practices. Modeled on Sarbanes-Oxley’s individual accountability for executives who certify materially misleading financial disclosures. Applied to AI, a CEO who signs off on a public safety claim contradicted by internal red-team results would face personal liability — not just a corporate fine that gets absorbed as a cost of doing business. The Clearview AI case shows that civil enforcement alone, even when financially crushing, arrives too late to undo the harm.

Licensing requirements based on model size and use case. Modeled on FDA tiering and FAA airworthiness certification. Below a compute threshold, no license needed. Above it, deployment requires demonstrated capability assessments, red-teaming results, and rollback procedures. Above a second threshold (frontier models with novel capabilities), a pre-deployment safety case. This is how aviation moved from frequent fatal accidents to one of the safest forms of transportation in the world — graduated, capability-tied requirements with mandatory disclosure.

Transparency requirements for the sources of model training data. Modeled on real-estate disclosure law and pharmaceutical labeling. A buyer of a house has the right to know what materials are in the walls; a patient has the right to know what is in their medication. A user of an AI system has the analogous right to know what corpus shaped its outputs — particularly when those outputs influence hiring, lending, healthcare access, or political speech. Voluntary disclosure has not produced this transparency. A regulatory mandate, with audit rights for affected parties, would.

Revenue sharing with creators when AI models are trained on their intellectual property. Modeled on YouTube’s Content ID system, which automatically identifies copyrighted material in user uploads and routes a share of advertising revenue to rights holders. The infrastructure for this already exists. The political will does not. Applied to AI, every model that demonstrably draws on copyrighted training data would pay into a pool distributed to creators whose work appears in attribution traces.

These four mechanisms, taken together, would shift the equilibrium. They take inspiration from consumer protection initiatives in regulated, high-risk industries. While not perfect, these sectors have seen the beneficial impact of structured regulatory frameworks in the US, such as criminal penalties (finance and banking), licensing requirements (energy, healthcare, and law), transparency requirements (real estate, aviation, and pharmaceuticals), and revenue sharing (online video sharing, notably YouTube’s Content ID system). This framework, informed by teachings from ESG, regulated industries, and AI ethics discourse, will guide AI towards a more socially equitable future.

Just as the ESG movement has grappled with the challenge of actualizing its ethical intentions amidst profit-driven corporate dynamics, AI ethics faces similar hurdles, often sidelining ethical commitments under competitive pressures. This analysis underscores the importance of integrating ethical considerations into AI’s operational and decision-making processes. It proposes a framework that addresses some of the immediate challenges in AI and broader societal implications, echoing the calls for a systemic overhaul by scholars like Powles and Nissenbaum (2018). In doing so, the essay aligns itself with a growing body of critical thought that seeks to push for a holistic reevaluation of how technology intersects with societal values.

The lessons from ESG’s shortcomings illuminate a path forward for AI ethics grounded in a commitment to societal welfare and ethical standards, addressing the needs and values of diverse stakeholders, including consumers, developers, and regulatory bodies. The urgency for a robust regulatory framework for AI, as illustrated through examples from the AI and transportation sectors, underscores the need to integrate ethical considerations at every stage of technological development. This essay advocates for robust external safeguards, challenges the current over-reliance on corporate self-regulation, and assesses the pervasive influence of corporate interests on regulatory frameworks. In doing so, it contributes to the discourse on AI ethics and champions a vision where algorithmic accountability becomes the cornerstone of a future where AI is a positive, ethical force in our lives.

Bibliography

• 2022 Sustainability Report. (2023). American Airlines. s202.q4cdn.com
• Bietti, E. (2021). From Ethics Washing to Ethics Bashing: A View on Tech Ethics from Within Moral Philosophy (SSRN Scholarly Paper 3914119). doi.org/10.2139/ssrn.3914119
• Buccino, J. (2023, November 3). Red Teams, watermarks and GPUs: The advantages and limitations of Biden’s AI executive order. Federal News Network.
• Chu, J. (2017, March 3). Study: Volkswagen’s excess emissions will lead to 1,200 premature deaths in Europe. MIT News.
• Cooper, A. F., Moss, E., Laufer, B., & Nissenbaum, H. (2022). Accountability in an Algorithmic Society: Relationality, Responsibility, and Robustness in Machine Learning. 2022 ACM Conference on Fairness, Accountability, and Transparency, 864–876.
• Climate at a Glance: Global Time Series. (2023). NOAA National Centers for Environmental Information.
• Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. (2023, October 30). The White House.
• Hart, R. (2022, May 23). Clearview AI Fined $9.4 Million In U.K. For Illegal Facial Recognition Database. Forbes.
• Khalili, J. (2023, November 22). CZ Has Left Binance, SBF Is in Jail. Crypto Is About to Get Boring. Wired UK.
• Mac, R., & Browning, K. (2024, June 13). Clearview AI Reaches Settlement in Landmark Facial Recognition Case. The New York Times.
• Mathiesen, K., & Neslen, A. (2015, September 23). VW scandal caused nearly 1m tonnes of extra pollution, analysis shows. The Guardian.
• Nissenbaum, H. (1996, March). Accountability in a computerized society. Science and Engineering Ethics, 2(1), 25–42.
• Nordhoff, S., Lee, J. D., Calvert, S. C., Berge, S., Hagenzieker, M., & Happee, R. (2023). (mis-)use of standard autopilot and full self-driving (FSD) beta: Results from interviews with users of Tesla’s FSD beta. Frontiers in Psychology, 14.
• Pasquale, F. (2019, November 25). The second wave of algorithmic accountability. LPE Project.
• Powles, J., & Nissenbaum, H. (2018, December 7). The seductive diversion of ‘solving’ bias in artificial intelligence. OneZero on Medium.
• Pucker, K. P., & King, A. (2022, August 1). ESG Investing Isn’t Designed to Save the Planet. Harvard Business Review.
• Shifflett, S. (2023, November 19). Wall Street’s ESG Craze Is Fading. The Wall Street Journal.
• Still the One: American Airlines Named to Dow Jones Sustainability North America Index. (2022, December 12). American Airlines.
• Tesla’s full self-driving recall impacts AV Industry. (2023, March 7). SAE International.
• Volkswagen is world’s most sustainable automotive group. (2015, September 11). Volkswagen of America.
• Weise, K., Metz, C., Grant, N., & Isaac, M. (2023, December 5). Inside the A.I. Arms Race That Changed Silicon Valley Forever. The New York Times.
• Yu, E. P., Luu, B. V., & Chen, C. H. (2020). Greenwashing in environmental, social and governance disclosures. Research in International Business and Finance, 52, 101192.