Some people believe AI is all good for humanity while others warn that this technology is dangerous and poses a serious threat to society. While we don’t know who is right about AI yet, some scientists believe we should err on the side of caution.
Roman Yampolskiy, Director of the Cybersecurity Laboratory at the University of Louisville recently published a study that suggests that although artificial intelligence is a great technology that is already benefitting humans in numerous ways, nobody is addressing the problems with safe AI control.
We shouldn’t underestimate the fact that a superintelligent general AI, capable of outperforming human intelligence, can also cause an existential catastrophe if it gets out of human control, according to Yampolskiy.
Partial vs fully controlled AI
The study highlights that, currently, humans have partial control over AI. Human operators can influence but not fully dictate AI behaviors. This might include setting broad guidelines or objectives but there’s no real way to ensure compliance with those objectives under all circumstances.
“An example of partial control is current AI systems where safety measures and oversight mechanisms can guide AI behavior but cannot eliminate the risk of unintended consequences due to the AI’s limited understanding of complex human values,” Yampolskiy told ZME Science.
For instance, when Microsoft developed Bing AI, the company set some guidelines regarding the use of language and tone. However, when Bing AI was first released, it acted erratically and even threatened users. That must have spooked some folks.
We don’t know what was wrong with the program but one possibility could be that it was under partial control and didn’t adhere to its human creators’ guidelines. At least, such AI programs should be able to explain the factors that lead them to act or make decisions in a certain way.
This is why Yampolskiy advocates for full safe control over AI. This would guarantee — at least in theory — that AI systems will always act in ways that are beneficial to humanity and aligned with our values, regardless of the situation.
“One of the key findings from our research in the field of AI safety is the concept of ‘alignment problem,’ which underscores the challenge of ensuring AI systems’ goals are perfectly aligned with human values and intentions. This misalignment, especially in powerful AI systems, demonstrates a form of being out of human control,” Yampolskiy added.
The current study couldn’t find any evidence that the current generation of AI programs is under full safe human control.
AI can only threaten our jobs, not us. Right?
The main concern most people have regarding AI is whether it would replace their jobs. However, what they don’t understand is that the concern about AI causing existential catastrophe goes beyond job displacement.
It involves scenarios where highly advanced AI systems start to act in ways that are (un)intentionally harmful to humanity on a global scale.
“The risk is particularly pronounced with the development of superintelligent AI, which could outperform human capabilities in all domains, including strategic planning and manipulation, potentially leading to scenarios where humans could not control or counteract their actions,” Yampolskiy told ZME Science.
For instance, a superintelligent AI tasked with a seemingly benign goal could adopt catastrophic means to achieve it if those means were not explicitly prohibited. But can you ever really cover all scenarios?
“We are facing an almost guaranteed event with the 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,” Yampolskiy warned in a press release.
Challenged with AI safety and control
Studies (including the current research) on AI safety and control face several limitations, including the theoretical nature of many risks associated with advanced AI, making empirical validation challenging.
Additionally, the rapid pace of AI development can outstrip the speed at which safety measures are conceptualized and implemented.
Also, there’s the challenge of ensuring global cooperation in AI safety efforts, as different stakeholders may have divergent interests and levels of commitment to safety protocols.
All these factors make safe AI control a very complicated issue.
In fact, a study published in 2021 by researchers from the Max Planck Institute for Human Development suggests even if we want to, technically, it is almost impossible to fully control a super-intelligent AI.
“It is impossible, due to fundamental limits inherent to computing itself. Assuming that a superintelligence will contain a program that includes all the programs that can be executed by a universal Turing machine on input potentially as complex as the state of the world, strict containment requires simulations of such a program, something theoretically (and practically) impossible,” the authors of the 2021 study notes.
However, that doesn’t mean the problem is completely unsolvable.
There are still some ways to ensure safety
When we asked Yampolskiy about some good steps he thinks AI companies and regulators can take to mitigate the risk with superintelligent AI and make it safe and controllable, he replied with the following suggestions:
- Developing robust AI safety standards that are universally adopted and enforced.
- Investing in research to better understand and mitigate potential risks associated with AI technologies.
- Implementing transparency measures to ensure the workings and decision-making processes of AI systems are understandable to humans.
- Establishing oversight mechanisms that include both technical safeguards and regulatory frameworks to monitor and guide AI development and deployment.
- Fostering international cooperation to ensure global alignment on AI safety standards and practices.
“The goal is to ensure that as AI technologies continue to evolve, they do so in a way that benefits humanity while minimizing potential harm,” Yampolskiy told ZME Science.
The study has been published by Taylor & Francis Group.