There is no proof that AI can be controlled

Acknowledging the critical issue of AI control as one of humanity's most significant challenges, it remains largely misunderstood, undefined, and under-researched, according to Dr Roman V. Yampolskiy.

In his forthcoming publication, ‘AI: Unexplainable, Unpredictable, Uncontrollable’, Dr Yampolskiy, an expert in AI safety, examines how AI possesses the capacity to significantly alter society, potentially not always for the better.

He states: “We are facing an almost guaranteed event with potential to cause an existential catastrophe. No wonder many consider this to be the most important problem humanity has ever faced. The outcome could be prosperity or extinction, and the fate of the universe hangs in the balance.”

The challenge of uncontrollable superintelligence

Through a thorough examination of AI scientific literature, Dr Yampolskiy concludes there is no evidence to suggest that AI can be controlled safely. Even theoretical control measures are deemed insufficient.

“Why do so many researchers assume that AI control problem is solvable? To the best of our knowledge, there is no evidence for that, no proof. Before embarking on a quest to build a controlled AI, it is important to show that the problem is solvable.

“This, combined with statistics that show the development of AI superintelligence is an almost guaranteed event, show we should be supporting a significant AI safety effort.”

Dr Yampolskiy highlights that our capability to create intelligent software surpasses our ability to control or verify it. His review suggests that advanced intelligent systems may never be fully controllable, presenting a persistent risk despite their benefits. Minimising this risk while maximising potential benefits should be a priority for the AI community.

Identifying the challenges

AI, especially superintelligence, differs from conventional programs in its ability to learn, adapt, and operate semi-autonomously. Ensuring safety becomes infinitely complex as a superintelligent entity's decision-making capabilities evolve, making prediction and mitigation of safety issues nearly impossible.

Moreover, AI's inability to explain its decisions or our inability to comprehend them further complicates safety efforts. This lack of transparency could prevent understanding and mitigation of future accidents.

For instance, AI's role in critical decision-making sectors like healthcare and finance necessitates a system that can elucidate its reasoning, ensuring decisions are unbiased and justifiable.

Controlling the uncontrollable

As AI's autonomy expands, our control diminishes, associating increased autonomy with decreased safety. The notion that less intelligent entities (humans) cannot indefinitely control more intelligent entities (AI) is not due to a lack of safe design options; it's because such a design may inherently be impossible.

“Humanity is facing a choice, do we become like babies, taken care of but not in control or do we reject having a helpful guardian but remain in charge and free,” Yampolskiy explains.

He suggests finding a balance where some autonomy is exchanged for control could be a solution, albeit at the expense of limiting the system's capabilities.

Aligning human values with AI

The challenge of ensuring AI acts in humanity's best interest involves navigating the potential for conflicting orders and the risk of misuse. While having AI act as an advisor could mitigate some risks, this approach requires the AI to possess superior values.

The quest for value-aligned AI, inherently biased towards pro-human values, presents a paradox where AI might refuse direct orders to fulfil the underlying intent, posing a dilemma between protection and respect.

Minimising risk

Dr Yampolskiy advocates for AI systems that are modifiable, limitable, transparent, and understandable. He calls for a clear categorisation of AI as controllable or uncontrollable, suggesting moratoriums or partial bans on certain AI technologies.

Instead of being discouraged, he says: “Rather it is a reason, for more people, to dig deeper and to increase effort, and funding for AI Safety and Security research. We may not ever get to 100% safe AI, but we can make AI safer in proportion to our efforts, which is a lot better than doing nothing. We need to use this opportunity wisely.”