Peril and promise

AI can revolutionise industries, amplify human creativity and help solve some of our most complex problems. However, it also raises questions about control, bias, climate, and the future of work. Our primer for leaders last year provided concrete recommendations to address both these possibilities – the promise and the peril.  

This report is about their intersection with every organisation’s sustainability objectives (commonly described in terms of environmental, social, and governance (ESG) issues). For the purposes of the report, we make no assumptions about what those commitments may be, or the principles which guide them (e.g. UN Sustainable Development Goals and Global Compact, or narrower, sector-specific ones like the Equator Principles). 

We describe three imperatives for leaders and anyone thinking about strategy for AI, AI adoption or sustainability:

1. Recognise and understand the peril of AI for your company’s ESG and sustainability agenda. While this peril overlaps with the general risk associated with AI (about which much has already been written), it is not the same thing.
2. Prepare to harness the promise of AI to accelerate and amplify the impact of your sustainability agenda. It will make it easier to fulfil existing objectives and allow organisations to imagine more ambitious ones as well.
3. Take practical steps today to turn these insights into action by incorporating them into both your AI and sustainability/ESG strategy and development pathway. 

There has been an enormous amount of attention paid to the risks of AI. What may be less appreciated is how much the disruptive impact of AI may affect the sustainability (or ESG) agenda for individual companies. Consider each element of that agenda in turn.

Amplifying and accelerating the impact of technology on the environment

There is not yet much attention being paid to the environmental impact of AI, simply because at this stage of the adoption cycle, its impact is still limited. However, AI has the potential to significantly expand the environmental footprint of every part of the technology value chain. It starts at mining and refining rare elements for bespoke chips. It includes the energy and water used in data centres, which are expected to account for more than 3% of greenhouse gas emissions by next year, and 15% by 2040. Training and even using GenAI models is more energy intensive than one might think (even if we believe that will come down)

Finally, there is the issue of disposing increasing volumes of so-called e-waste (estimated ~64 million tons in 2024), which is proving hard to keep out of landfill. Most of this is not (yet) due to AI. But the exponential growth anticipated for this technology means it is something companies will increasingly have to start tracking. In some jurisdictions, companies will be required to comply with reporting obligations (such as the EU’s Corporate Sustainability Reporting Directive (CSRD) and even the law (e.g. EU’s Artificial Intelligence Act).

Potentially dramatic disruption of social norms and expectations

We are only beginning to get our heads around the social impacts of AI adoption, but they may be among the most important. Start with the impact on one’s own employees and their families. Résumé screening, job filtering and even performance evaluation is increasingly augmented by AI. This means that limitations in your model-risk-management capabilities, as well as errors and biases in the internal data you use to develop AI, may impact your own people first. If those errors turn out to be systemic (e.g. because every company trains its résumé-screening technology on similar data), then the consequences for specific communities can become graver still, as we discuss below.  

The same is true for customers: as AI systems increasingly intermediate your interactions with them through all stages of the customer journey (from identity verification, membership or product approval, customer service and ongoing support), the potential grows for inadvertent customer harm from biased, miscalibrated, or poorly designed AI. The harm can cover the gamut from inconvenience (additional security procedures required) to exclusion (product application denied) to fundamental human rights (e.g. AI-driven target selection in warfare).

New possibilities for cybercrime and fraud 

Fraud and cyber risk are also issues that have both a specific component (i.e. those directly affected) and a general societal component. As interactions between an organisation and its stakeholders are more intermediated by AI, vulnerabilities for everyone will increase. At the same time, AI will be used by scammers and attackers in increasingly novel ways. Even in cases where incidents do not rise to the level of legal or regulatory risk, reputational damage can be sustained if an organisation’s social commitments are undermined or discredited.

Specific peril can rapidly scale 

When the problems are common enough, the impact can spread to society more broadly. The effect is multiplicative, not additive. For example, unfairly assessing the credit risk, health profile or professional skills of 1,000 people is always a problem. However, if those 1,000 people are concentrated in communities already struggling with other kinds of disadvantage, then the impact can be much greater, as is the feedback on social welfare.   

In fact, societal harm can be significant even when individual harm is not. Consider the controversies around the role of social media in elections over the past ten years: the AI involved was relatively primitive, and for most individuals, the impact limited to excess adrenaline and the odd heated argument with relatives and friends. However, the social impacts are still being felt today.  

As the AI and agents behind such things as automated social media engagement become more interactive and persuasive, their capacity to influence large groups (and erode public trust and social solidarity) will continue to grow.  

What’s more, as the complexity of AI in use increases, and especially as it is allowed greater freedom to learn and evolve, the risk of what computer-scientists call ‘wire heading’ becomes real. This is when AI takes shortcuts to achieve its objective by manipulating or ‘rigging’ either its reward function or the operational environment. For example, you may have heard the fable of the robot housekeeper: programmed to find the ‘most efficient’ way to keep the house clean, it poisons the pets.

For now, that is just a joke, and the domain of science fiction. However, there are always many ways to skin a cat. The more capable the AI, the more likely it is to find methods its creators had not considered. This is why model explainability, auditability, traceability and alignment are so important. They will remain critical areas of research and development for a long time (as we discuss further below).

Finally, although this may feel like it is beyond the scope of any one organisation’s responsibility, responsible leaders must think through the implications of widespread adoption of AI technologies on society overall, especially when social responsibility is an important part of the brand. There is no point, for example, marketing one’s impact on such things as equality, inclusion, or shared opportunity, while simultaneously deploying technology seen as disadvantaging large segments of the community without having a clearly considered explanation for how those things are compatible. Of course, this point is relevant for everything, not just AI. However, the exponential adoption of this disruptive technology means this will likely be a more pressing issue than ever.

Challenges to the foundations of governance and accountability

AI poses a challenge for all aspects of the ‘governance’ part of the ESG agenda, and all senior leaders (including directors on boards) should educate themselves about the implications for them and their obligations.

This is because AI calls into question a core governance tenet today: that for every operation or risk there should be clear and unambiguous lines of accountability. For many functions (depending on industry and jurisdiction), that can even be a legal or regulatory requirement (e.g. directors obligations enshrined in corporate law, or industry-specific licensing requirements). AI challenges that because it raises critical questions about who is really doing what. Operations may be accountable for customer service, but if the customer support team (directed by a language model maintained by technology) does the wrong thing, who is responsible? More importantly, is the answer to such questions commonly understood in your organisation?

Start with model transparency, traceability, and auditability 

As the scope of AI adoption expands, more executives will need to become more proficient in old-fashioned model governance and risk management to fulfil the ‘G’ in ESG. Unfortunately, there are no easy ways to do this. AI models are less transparent than even the most complex system, and critical parameters can dynamically adapt (which we call ‘learning’). This is a big theme in PwC’s own Responsible AI toolkit, as well as the many ethical AI frameworks and principles which are emerging around the world (including in Australia, the OECD, and the EU).  

All models incur risk, and the management of this risk is the same (conceptually) for AI as for simpler ones:

• Understandability (e.g. through good design, documentation, and taking the time to track and test intermediate outputs)
• Being careful about selecting applications that balance value with risk (e.g. for AI, internal applications may be preferable to externally facing ones, at least to begin with)
• Setting clear boundaries and appetite for the residual risk
• Being disciplined about monitoring, oversight and control

AI alignment

This also requires leaders to pay attention to something called ‘AI alignment.’ This is about ensuring that AI models perform their function in ways consistent with the overarching ethics and values of the organisation, even as they are given increasingly wide discretions. 

This is not easy, as early users of GenAI discovered a few years ago. The solution obviously requires:

• extensive testing to see how things can go wrong,
• explicit filters and overrides to catch potentially problematic output, such as text that promotes self-harm, and,
• where possible, incorporating ethical objectives and priorities into the ‘objective functions’ of various systems. 

As one might expect, this is an active area of both academic research (see this survey from Chinese University in Hong Kong), and practical experimentation (especially around emerging ‘AI hubs’ such as the Mila Institute in Montreal) 

Systemic risk and regulatory ‘supervisability’

The same things that can challenge internal governance and accountability for organisations also apply to regulators – which is not just a problem for them. If it affects the health and confidence of the market, it is a problem for everyone. The risks are compounded as concentration increases around common elements of the ‘AI’ value chain (whether chips, classes of models, methodological approaches to problems, or data).  

This may sound esoteric, but so did the mechanics of measuring credit correlation inside securitisations and structured products before the GFC. When Gary Gensler, Chairman of the US Securities and Exchange Commission (SEC), testified to Congress about his concerns over AI concentration as a risk to the financial system, he no doubt had this experience in mind. Unfortunately, systemic risk is not limited to the financial system. For leaders, it is critical that they: 

• understand the fundamentals
• ask questions about vulnerabilities, risks, mitigants and backup plans
• apply special scrutiny to the examination of widely held common assumptions.