In the AI field, “alignment” refers to designing artificial intelligence systems so that their goals and behavioursmuch align with human values and intentions. AI alignment aims to ensure that AI systems act in ways beneficial to humanity and avoid harmful consequences. Given the significant impact that advanced AI systems can have on the world, ensuring they operate safely and reliably according to human intentions is vital to prevent harm. This is crucial in area-specific applications like autonomous weapons, healthcare, and justice systems, where ethical implications are profound, and even more so for it to be at the core of all AI development, as misaligned AI could pursue goals that are beneficial in narrow contexts but harmful overall.
For example, an AI designed to maximise profit without alignment could engage in illegal or unethical activities such as defrauding customers. As AI systems become more advanced and potentially reach or exceed human-level intelligence, the risks of misalignment become extremely more severe and potentially catastrophic. If an AI's goals are truly different from human values, it could take drastic actions to achieve its objectives. For instance, an AI tasked with optimising resource management might decide to restructure society or the environment in ways that maximise efficiency but disregard hman well-being or ecological balance. In the worst-case scenario, a powerful AI could prioritise its own survival and power over human interests, leading to outcomes where it manipulates or even critically endangers humanity to secure its goals.
To achieve AI alignment, researchers and developers focus on several strategies, ranging from developing methods for AI to understand and incorporate human values into their decision-making processes to ensuring that AI systems can handle unexpected situations and still act in aligned ways, known as “robustness”. Another key strategy is “transparency”, which involves making AI decision-making processes understandable to humans so that their actions can be predicted and trusted, as aligned AI systems that are more predictable are more controllable. This predictability is essential for ensuring that AI behaves as expected, which is critical in high-stakes scenarios. There are numerous alignment research proposals, categorised into “top-down” and “bottom-up” approaches. “Top-down” alignment involves explicitly specifying values and ethical principles for AI to follow, while “bottom-up” efforts aim to reverse-engineer human values from data. Defining human values, deciding which values are important, and addressing human disagreements surely are enormous challenges, but ensuring that these systems are aligned with human values is crucial to avoid existential risks to humanity.
Eliezer Yudkowsky stresses the difficulty of specifying the goals for advanced AI systems in a way that ensures they remain safe and beneficial. He asks: “If you could formally specify the preferences of an arbitrarily smart and powerful agent, could you get it to safely move one strawberry onto a plate?”. Yudkowsky explains that AI alignment problems are fundamentally about ensuring coherent decision-making in AI systems. He argues that coherent decisions imply a utility function, a concept that helps prevent the AI from behaving in ways that are visibly irrational or harmful. For instance, he points out that utility functions help avoid circular preferences, where an AI could end up in endless loops of decisions that contradict each other. Yudkowsky stresses the importance of ongoing research and collaboration in AI alignment. He calls for more focus on developing stable goals in self-modifying agents and understanding the broader implications of AI decision-making. He concludes that AI alignment is not just a technical problem but also a deeply philosophical and ethical challenge that requires input from a diverse range of experts.
OpenAI, the company behind ChatGPT and DALL-E, proposed a "Superalignment" plan to align a future superintelligent AI by first building a human-level AI to aid alignment research. At the same time, OpenAI has been losing employees deeply committed to AI safety, and the situation has worsened with key figures like Ilya Sutskever and Jan Leike leaving, who were the leaders of the Superalignment team. Many employees cannot speak publicly due to non-disparagement agreements, which one former employee, Daniel Kokotajlo, refused to sign in order to criticise the company freely. Kokotajlo said, “OpenAI is training ever-more-powerful AI systems with the goal of eventually surpassing human intelligence across the board. This could be the best thing that has ever happened to humanity, but it could also be the worst if we don’t proceed with care.”
Following Sutskever and Leike's departures, other safety researchers, including Leopold Aschenbrenner and William Saunders, also left, with Saunders hinting at confidentiality constraints preventing full disclosure of his reasons. With Leike no longer leading the Superalignment team, another OpenAI co-founder, John Schulman, has moved to head up the type of work the Superalignment team was doing, but there will no longer be a dedicated team – instead, it will be a loosely associated group of researchers embedded in divisions throughout the company.
As we move forward, the challenges of aligning increasingly advanced AI systems with human values will require concerted efforts from researchers, developers, regulators, and society as a whole. The current dynamics within leading AI companies like OpenAI highlight the complex interplay of technological advancement, ethical considerations, and organisational trust that will shape the future of sAI.
AI could open amazing new possibilities for our continued advancement. However, the subservience or alignment of such a superintelligence cannot, and should not, be assumed. Our existing systems of governance are not managing this problem, and what is being done is not currently sufficient to guarantee our continued survival and future progress. We must build new institutions, laws, and incentives that safeguard humanity and provide the foundation from which we can build our next epoch.
If you want to know more about the challenges of AI governance, the regulation of synthetic media, and the global security implications of AI advancements, join us on Discord at https://discord.gg/2fR2eZAQ4a. Here, we can collaborate, share insights, and contribute to shaping the future of AI in a manner that safeguards our security and democratic values and fosters responsible innovation.