AI safety: Beyond manifest harm, towards fair and ethical AI

policy circle image

The discourse surrounding artificial intelligence is laying a lot of emphasis on AI safety. This was evidenced by the recent high-profile international AI safety summit organised by the United Kingdom, where India was a participant. Following this, the UK government announced the creation of an AI safety institute, and across the Pacific, the United States mirrored this effort by establishing a similar institute and consortium. These events, all withing a month, highlight the newfound importance of AI safety on the international agenda.

The need for AI safety is self-evident — who would want unsafe AI? Unsafe AI inherently implies negative consequences, and everyone naturally gravitates towards positive, safe outcomes. However, defining what constitutes ‘safe’ AI is more complex. Safety is a broad term that encompasses a wide spectrum of meanings. In everyday contexts, safety means the absence of harm or risk—like a safe street where one can walk without fear of harm, a secure bank account protected from fraud, or safe toys without choking hazards. Similarly, safe AI may be understood as AI that poses no manifest harm or risk of such harm.

The AI safety summit in the UK discussed concerns about the misuse of AI by rogue elements, and focused on cybersecurity, biotechnology, and disinformation. This focus on manifest harm is also evident in the US initiative, which highlights the importance of rigorous testing to eliminate potential risks. Prime Minister Narendra Modi’s focus on the issue, particularly in the context of deepfakes, also aligns with the broader theme around risks of manifest harm. US Vice President Kamala Harris further stressed on this aspect by citing examples where AI has caused harm in healthcare and legal settings, such as faulty AI leading to healthcare mishaps or wrongful convictions.

READGoogle antitrust case: A battle for the soul of the internet

AI safety: Beyond the fear of unknown

The internet is abuzz about a letter written by staff researchers of OpenAI to the board of directors which warned that a powerful artificial intelligence (AI) discovery named Q* could threaten humanity. The letter is believed to be one of the factors that led to Open AI chief executive Sam Altman´s firing on November 17. Several OpenAI staffers believe that Q* is a breakthrough in artificial general intelligence, says a Reuters report. This revelation is expected to lend further urgency to the efforts towards AI Safety.

The ongoing debate on AI safety, with its emphasis on avoiding manifest harm, carries an implicit message. On the surface, it seems to advocate shunning unsafe AI and, by extension, embracing AI that is free from such harm. But this perspective might be overly simplistic and potentially misleading.

Consider a fictional example involving a ride-hailing app named Suber. In this scenario, Amir, a tech-savvy user, selects a ride through Suber, which employs AI algorithms to choose the best driver based on reliability ratings. While this process seems safe and efficient, avoiding any direct harm to the customer, it raises several ethical and fairness concerns.

For instance, Alex, a driver with lower ratings due to circumstances beyond his control, is unfairly disadvantaged by the rating system. Ravi, the chosen driver, loses a significant part of his earnings to the app, despite doing most of the work. Also, the surveillance and performance pressures on drivers by the AI system raise questions about fair work conditions. These issues are not just for the gig economy but are indicative of patterns across various sectors where AI is used.

Such concerns extend to social media and the broader tech industry. In ‘Algorithms of Oppression’, the author discusses how AI in search engines can perpetuate biases and stereotypes, particularly against marginalised groups. ‘Move Fast and Break Things’ delves into how creative industries are exploited by AI-driven recommendation systems, often to the detriment of artists and creators. These examples illustrate how the current use of AI is riddled with structural issues, making its operation far from reasonable when viewed through a political economy lens.

The debate on AI safety, focusing mainly on the avoidance of direct harm, falls short in addressing these broader, more nuanced ethical and socioeconomic issues. While the AI safety movement is crucial and provides a necessary foundation for discussion, it must not be seen as a comprehensive solution. If global enthusiasm for AI safety leads to standards that only address manifest harm, it could inadvertently endorse AI systems that are exploitative or unfair under the guise of being ‘safe’.

Such a scenario would disproportionately benefit Big Tech, allowing them to use ‘Safe AI’ labels for competitive advantage, while potentially sidelining smaller entities that might struggle to meet these standards. The discourse around AI safety should ideally be inclusive, multi-dimensional, and vigilant against being commandeered by vested interests. While AI safety is an important goal, it is crucial to recognise its limitations and strive for a more holistic approach that encompasses fairness, ethics, and broader societal impact.

(Dr Deepak P is an Associate Professor of Computer Science at Queen’s University Belfast. His research interests are in AI ethics, and the politics of artificial intelligence.)