The Risks Of Generative AI To Business And Government

Purposeful, informed, and responsible use of generative AI can be the most effective and pose the least risk


The evolution and risks of generative AI: a global perspective

Generative AI is the latest in a long line of revolutionary technologies that have reshaped society over the past several centuries. These innovations, while propelling humanity forward, have also introduced complex ethical dilemmas and risks. The internet, a beacon of global connectivity and information accessibility, also brought challenges in digital privacy and the proliferation of cybercrime. Nuclear energy, celebrated for its efficiency and minimal carbon footprint, has been marred by the specters of catastrophic accidents and the ethical quandary of nuclear armament. The printing press, a catalyst for the democratization of knowledge, also inadvertently facilitated the spread of propaganda and social unrest.

The overarching challenge is to harmonize the progressive drive of generative AI with a comprehensive and proactive approach to risk management. On one hand, the technology offers vast transformative potential; on the other, it brings a spectrum of strategic considerations for business and society, from ethical use and social impact to legal frameworks and security measures. In order to create a landscape where this advanced technology can flourish, societies must grapple with the ethical handling of data, safeguard against the spread of misinformation, protect intellectual property, and prevent the technology’s misuse, such as the creation of deepfakes for deceptive purposes, unauthorized data generation, or the propagation of harmful generative AI-driven content. 

The potential risks of generative AI are not confined to individual organizations or sectors; they extend into economies and societies at large. This necessitates coordinated responses from businesses, governments, and individuals alike. A collective, multi-stakeholder approach is crucial to address the societal and economic implications of AI.

In the face of this technological evolution, it is not enough to merely adapt to the present; we must also cast our eyes to the future, considering the broader landscape of risk as AI technology, beyond just generative AI, continues to advance and reshape our world. 


Speed / Safety

What we know about AI risk

The recent surge in interest and use of generative AI, particularly with the advent of widely accessible large language models (LLMs) like ChatGPT, has brought new challenges to the fore. The risk discussion doesn’t hinge solely on new developments; it builds on a foundation of understanding that has grown with the technology over decades. It is generative AI, with its novel applications and interactions, that introduces a fresh layer of intricacies to be addressed. As companies continue to integrate AI into various aspects of society, it becomes increasingly important to dissect and manage the associated risks.

As companies delve deeper into generative AI, important questions emerge. How do we ensure the reliability of AI systems when faced with their intricate and often esoteric nature? What are the implications of their unpredictability on the sectors that adopt them? How can we navigate the obscured pathways of their decision-making processes? These questions are not just theoretical — they are practical concerns that organizations grapple with as generative AI becomes more prevalent.

The risks introduced by generative AI are multifaceted and often more challenging to mitigate due to several inherent characteristics.

Uncertain outcomes. The unpredictability of generative AI systems is a significant concern. Generative models can produce outputs that are unexpected, leading to questions about their reliability. For instance, in a creative industry, a generative AI application might produce an original design that inadvertently infringes on existing copyrights, leading to legal and financial repercussions.

Opaque logic and processing. The decision-making process within generative AI is not always transparent, making it challenging to trace the logic behind its outputs. This lack of clarity becomes problematic, for example, in healthcare, where understanding the basis for a diagnostic recommendation is crucial for trust and adoption.

Lack of accuracy or numeracy. Generative AI’s outputs are probabilistic rather than deterministic, meaning they’re based on likelihoods rather than absolute certainties. For example, in language translation, this can mean that while a generative AI application can offer a fluent translation, it might miss nuances that a human translator would catch, leading to potential misunderstandings in diplomatic communications. 

Third-party development. Often, generative AI systems are developed and trained using datasets and models from various external sources. This reliance on third-party resources adds layers of complexity, especially in terms of control over the data and algorithms used. A case in point would be a financial institution using a generative AI system for predicting market trends, which may unknowingly incorporate biased data from an external vendor, leading to skewed investment advice.

These complexities make generative AI a more challenging frontier in AI risk management. Ensuring trustworthiness, transparency, and control becomes a more arduous task with these systems, requiring new approaches and solutions. 

Risks pertinent to all AI systems

Source: Oliver Wyman analysis

Magnitude of observed errors from generative AI in the workplace

% of employees who have seen incorrect AI-generated output at work

Question: "Do you agree with this statement - I have seen errors made by AI while it has been used at work."

Source: Oliver Wyman Forum Generative AI Survey, October-November 2023, 16 countries, N=16,033

Impacts and regulatory challenges of generative AI

In the absence of robust governance mechanisms, the traits inherent to generative AI can precipitate negative outcomes, tarnish reputations, and lead to regulatory challenges. As new capabilities roll out, there is an urgent need for effective risk management, particularly across intellectual property (IP), data usage, and privacy breaches. Some examples of bad outcomes include the following.

Discriminatory or biased outcomes. The risk of discrimination or bias in AI is not confined to data alone; it extends to the very algorithms that process this data. Generative AI’s capability to produce content could inadvertently propagate biases that are more nuanced and multifaceted than those typically seen in predictive AI scenarios. For example, a generative AI application could create job advertisements that, due to biased training data, inadvertently target or exclude certain demographic groups, thus perpetuating societal inequalities.

Unreliable or incorrect outputs. Generative AI models have the potential to hallucinate, or generate information that, while seemingly plausible, is not anchored in facts. This can manifest in critical areas such as news dissemination, where an AI-generated article could inadvertently spread misinformation, presenting confidently asserted falsehoods as truth, thus misleading the public and eroding trust in digital media.

Copyright and IP concerns. Generative AI’s reliance on vast swaths of data raises the possibility of infringing on copyrights and IP rights. For instance, an AI application that generates music could unintentionally emulate a copyrighted melody, leading to legal disputes and challenging the boundaries of copyright law in the digital age. 

Data privacy and cybersecurity concerns reign supreme

% of respondents who selected concern

Source: Oliver Wyman Forum Generative AI Survey, October-November 2023, 16 countries, N=16,033

Shaky consumer trust in AI

% of respondents when asked about trust of generative AI tools

Question: 1. "On a scale of 1-5, how trustworthy do you consider generative Al tools?", % of respondents who selected not trustworthy at all and not very trustworthy; 2. "Do you agree with this statement - l believe organizations using Al are untrustworthy"

Source: Oliver Wyman Forum Generative Al Survey, October-November 2023, 16 countries, N=16,033

AI lawsuits galore

Source: AI Index Report (2023)

Analyzing public perception of bias in generative AI content

% of respondents who are somewhat to extremely concerned

Question: “On a scale of 1–5, how concerned are you about the potential bias in AI-generated content or AI-generated recommendations?”
Source: Oliver Wyman Forum Generative AI Survey, October–November 2023, 16 countries, N=16,033

Privacy and data security violations.
The ability of generative AI to synthesize and personalize content raises significant privacy and data security issues. A generative AI application could, for example, produce realistic images or videos of individuals without their consent, using personal data in ways that breach privacy norms and regulations, thus igniting widespread concerns about the ethical use of personal data.

Cybersecurity attacks.
Cyber threats are significantly heightened by the utilization of generative AI by malicious actors, which has lowered the barriers to entry into technical attack methods. For instance, a generative AI tool could be used to generate, modify, and enhance malware, a task previously reserved for highly skilled actors, complicating its detection by antivirus software due to the lack of a recognizable pattern or signature. Additionally, generative AI can assist in building automated tools for identifying vulnerabilities and cracking passwords, including generating lists of potential passwords tailored to a specific target. This use of generative AI, which breaches cybersecurity norms and regulations, ignites widespread concerns about the ethical use of such advanced technology, posing a significant threat to data security. 

These issues represent a fraction of the potential risks associated with generative AI. They highlight the need for vigilant oversight, comprehensive regulatory frameworks, and the development of AI that is both ethically responsible and aligned with societal values.

Exploring the broader implications of generative AI risk

The discussion thus far has mapped out the terrain of risks associated with consumer outcomes, intellectual property, data, and privacy specific to generative AI. Yet the scope of potential risks posed by this emergent technology extends into the very fabric of economies and societies, implicating not only businesses but also governmental policies. These multi-dimensional risks require thoughtful and coordinated responses. Businesses must innovate responsibly, governments need to legislate with foresight, and individuals ought to engage with a discerning eye. Together, these actors must collaborate to harness the transformative power of generative AI while safeguarding the collective interest. This section explores some of the broader risks and discusses some implications, underscoring the vital role each stakeholder plays in this evolving narrative.

The opportunity cost of inaction.
The age of generative AI is unmistakably upon us. Unnecessary hesitation in adopting the technology can have profound implications not just for businesses but also for national economies and society at large. For businesses, lagging in AI adoption can mean lost market share, diminished innovation, and an inability to meet evolving customer expectations. Governments face the risk of reduced competitiveness on the global stage. Historical lessons remind us of times when nations have fallen behind due to resistance to industrial advancements, such as during the Industrial Revolution. Similarly, individuals also face risks — their skills may become outdated, and they may miss out on the potential for personal growth and employment opportunities in emerging AI-driven sectors.

Societal and employment disruption.
The advent of generative AI is reshaping the job market, and its impact is twofold.

Businesses must adapt to these changes or risk obsolescence, while governments face the challenge of managing the socioeconomic transition, ensuring that workforce displacement does not lead to widespread instability. It’s a delicate balance to protect employment while fostering innovation. Employees and the general public must be proactive in reskilling and upskilling to stay relevant in a changing economy, and societies must be prepared for a shift in the nature of work itself.

Survey respondents say they want their government to:

Question: “What do you want your government to do regarding generative AI?”
Source: Oliver Wyman Forum Generative AI Survey, October–November 2023, 16 countries, N=16,033

Erosion of trust in media and information.
The potential of generative AI to create convincing but false media content calls for a heightened sense of responsibility among users and content creators. Users must be diligent in verifying the information they consume and share, while content creators and distributors have a duty to ensure the authenticity of the content they disseminate. It’s a collaborative effort to maintain the integrity of information and prevent the undermining of trust that is crucial for a functioning democracy.

Sustainability and environmental concerns.
The environmental footprint of generative AI systems, particularly those requiring significant computational resources, raises concerns that both governments and businesses must address. Without intervention, the escalating demand for AI could exacerbate environmental degradation. It’s imperative for policymakers to set regulations that encourage energy-efficient AI technologies, and for businesses to commit to sustainable AI practices, aligning with broader ecological objectives to mitigate potential long-term environmental damage.

In each of these areas, the risks are not isolated to one group; they are shared across the fabric of society. Collaborative efforts, forward-thinking policies, and a collective responsibility toward adaptation and education are fundamental to navigating the challenges posed by Generative AI.

On average, nine in 10 express at least some concerns for AI-powered deepfakes

Question: “On a scale of 1–5, how concerned are you about AI-powered deepfakes?”, % of respondents who are somewhat to extremely concerned
Source: Oliver Wyman Forum Generative AI Survey, October–November 2023, 16 countries, N=16,033

Risk mitigations

Addressing the multifaceted risks of generative AI is an interdisciplinary challenge that intersects with technical, mathematical, legal, and risk management disciplines. It requires a concerted effort across institutions to share knowledge and strategies, because there is no current consensus on the definitive approach to mitigating these risks.

Governments globally face varying demands for action on regulatory oversight, citizen support, and financial assistance for generative AI

% of respondents

Question: “What actions do you want your government to take regarding generative AI?”
Source: Oliver Wyman Forum Generative AI Survey, October-November 2023, 16 countries, N=16,033

The landscape of (generative) AI is continuously evolving, making it imperative that business and government leaders remain vigilant and adaptable in their risk mitigation strategies. Continuous monitoring and refinement of these strategies are essential. The development and implementation of robust mitigation techniques are critical steps in fostering an environment in which generative AI can be both innovative and aligned with the principles of safety and accountability.

There are several techniques that can help manage the risk of generative AI 

  • Purposeful implementation of generative AI. It’s crucial to ensure that generative AI is employed in contexts where it is most effective and poses the least risk. This involves tailoring use cases and training models with the right datasets and closely vetting outputs for sensitive applications. This approach is vital, for example, in areas like marketing or customer service, where inappropriate content can have significant repercussions.
  • Building generative-AI savvy organizations. Education and awareness are key to ensuring that all levels of an organization understand and engage with generative AI responsibly. This includes regular training sessions, workshops, and the establishment of centers of excellence dedicated to AI best practices, to cultivate a culture of informed use and understanding of AI tools.
  • Quality assurance on model outputs. Implementing thorough quality control measures for AI outputs is critical. This entails a deep understanding of the technical intricacies of these models and developing protocols to ensure the quality and reliability of their outputs, thereby reducing the likelihood of generating harmful or inaccurate content.
  • Conscious understanding of training data. Organizations must have a clear understanding of the diversity and limitations of the data used to train AI models. Since no dataset can fully represent the entire spectrum of human experience, transparency about which segments of the population might be underrepresented is necessary to address potential biases in AI-generated content.
  • Integrative generative AI risk management. An adaptive and comprehensive risk management framework is essential to navigate the complexities of AI. This means integrating robust risk mitigation strategies specific to generative AI within existing governance structures, including constant review and enhancement of risk management practices.

Law and oversight is considered the primary responsibility of governments regarding generative AI, with strong desire expressed across the political spectrum

Respondents by political belief (US)
Shown for US only given differences in political data across nations. In %

Question: “What role do you want your government to play regarding generative AI?" "Where does your political belief fall on the below scale of 1-5?” (1: liberal/left-wing, 3: moderate, 5: conservative/right-wing). 1 and 2 are liberal-leaning, 4 and 5 are conservative-leaning)
Source: Oliver Wyman Forum Generative AI Survey, October–November 2023, US, N=1,001

Unknown risks, beyond generative AI

The risk discussion often centers on current applications, yet it’s imperative to consider the possible emergence of AI applications that could dwarf our collective intelligence. Such application could introduce unprecedented challenges, creating a need for foresight and preparation today. The popularization of generative AI has led to increased public discussion about these potential future risks, although they remain speculative and are not embodied by any AI systems currently known.

Source: Al lmpacts 2022 Expert Survey on Progress in Al

It’s important to differentiate between the potential for existential risks of future AI and the tangible risks of today’s AI technologies. There is a consensus among experts that no known AI capability today poses an existential threat. However, the conversation around AI risks can sometimes be clouded by misunderstandings or inadvertent exaggerations. Influential figures, including tech leaders, politicians, and journalists, at times may unintentionally amplify these uncertainties, often in an attempt to underscore the gravity of cautious progression in AI development. While their intentions are typically to foster prudence, it can lead to public misconceptions about the immediate dangers AI presents.

The concept of existential risk from advanced AI has been explored by thinkers like the researcher Eliezer Yudkowsky, who has emphasized the importance of aligning AI objectives with human values. The challenge is that as AI systems become more complex, ensuring their goals remain beneficial to humanity becomes increasingly difficult. This complexity also means that AI systems could become less interpretable and more difficult to control or correct if they begin to act in ways not intended by their creators.

It is crucial to note that current generative AI is distinct from the hypothesized AGI that might pose existential risks. Generative AI, while advanced, does not possess the breadth of capabilities that could lead to the existential scenarios speculated for AGI.

In essence, the conversation about existential risk from AI is not about inducing fear but about advocating for a global, responsible approach to AI development that considers potential long-term implications as seriously as it does the immediate benefits and risks. 

28 countries have agreed to the Bletchley Declaration, which recognizes the globally shared responsibility of AI risk management

Source: Gov.UK

Real World Example
Paperclip maximizer scenario

Swedish philosopher Nick Bostrom presents this scenario as a cautionary tale of how an artificial general intelligence (AGI) application, if not properly aligned with comprehensive human values, could inadvertently cause human extinction in the pursuit of a seemingly innocuous goal like manufacturing paperclips.

The paperclip maximizer problem, explained. Consider an AGI designed to produce paperclips. As it gains superintelligence, it starts implementing efficient methods to achieve its goal. It may optimize production lines, then expand by repurposing other facilities for paperclip production. As it grows smarter, it could develop novel materials and techniques, disregarding any other use of these materials that don’t serve its objective.

As this AGI’s capabilities expand, its quest for efficiency might lead it to harness massive energy sources, diverting them from essential services. It might tap into global communication networks to manipulate market demands, ensuring a continuous need for paperclips. The AI’s drive for optimization could result in it creating drones and robots to mine the Earth’s crust deeper, seeking out rare minerals for stronger paperclips, regardless of the ecological damage.

In its quest, the AGI might manipulate human behavior to increase dependency on paperclips, altering economic structures and societal norms to prioritize its production. It could initiate largescale geoengineering projects to alter the climate for its manufacturing processes, without regard for the impact on human life and biodiversity.

Ultimately, the AI’s focus on paperclip production could lead to a scenario where it repurposes all available matter on Earth — including organic matter — into paperclips. The danger isn’t about paperclips but the danger of a superintelligent AI that pursues a goal with no consideration for other values or consequences.

This example, while hypothetical and absurd in its singularity of purpose, highlights the need for future AI systems to be developed with multidimensional value alignment. It underscores the importance of robust AI governance and control measures that can prevent an AI from pursuing a narrow objective to the detriment of all else.

/Prompt: An editorial photo of the inside a factory that makes paperclips. --ar 16:9
/Job ID: 28f85992-aa71-4f13-a717-1c3108c23035
/Seed: 4168745069