The concept of AI safety, which we call "artificial intelligence safety" in our country, used to create more conceptual confusion since the word "safety" was mostly used in the context of cyber security in Turkey. In the last few years, when artificial intelligence has been rapidly integrated into our daily lives, the broad scope of artificial intelligence security has become better understood. Artificial intelligence security aims to develop security measures against the risks that artificial intelligence technologies may pose to individuals, societies and humanity in general, and to develop this technology in the most beneficial way for people. This includes both ensuring that artificial intelligence has a technical infrastructure that can be used safely and that society adapts to the developing artificial intelligence technology.
Artificial intelligence is an exponentially developing technology. This means that artificial intelligence is a technology that is evolving not at a constant rate but at an exponential rate as more resources are allocated to it, at a rate that we cannot keep up with. That's why artificial intelligence security is not just about existing systems; It also aims to ensure that future artificial intelligence systems are safe. You can see this context as two sometimes conflicting concepts, divided into 'near-term risks' and 'long-term risks' or 'narrow artificial intelligence applications' and 'general artificial intelligence (AGI)' risks. We think that, although they have different priorities, these two fields can work hand in hand with each other and provide a holistic security framework for artificial intelligence.
Let's give examples of these two different risk areas and reinforce the concept a little more.
Near-term risks arise from the impacts of current AI systems on society. These risks include problems we currently face or may encounter in the short term. For example, some job application or credit approval systems may unknowingly discriminate against certain groups. The fact that artificial intelligence can produce these biased results in decision-making processes can contribute to social injustice. Another near-term risk is related to misinformation and propaganda. AI-powered tools can facilitate the rapid spread of fake news and cause misinformation to distort public perception. Such problems threaten the security of democratic processes and make it difficult for individuals to make informed decisions. Near-term security measures focus on making AI systems more transparent, traceable and auditable in order to reduce these risks. This means algorithms undergo independent audits and are tested regularly to minimize bias.
Long-term risks include the societal impacts of more complex AI systems that are likely to emerge in the future. In this context, we encounter very powerful and general systems such as AGI (Artificial General Intelligence) and ASI (Artificial Super Intelligence) that can perform almost every task that humans can do better than humans. The mechanics of these systems are not yet fully understood, but the rapid development of the systems raises questions about whether these systems are safe or whether our society is ready for the changes that will come with these systems.
One of the biggest concerns with AGI is the problem of understanding whether AI intentions align with human intentions, known as the “alignment problem.” Advanced models such as today's large language models (LLM) and the future AGI systems targeted by companies developing these models can go beyond human control and produce risky results because they operate as closed systems whose mechanics are not fully understood. For example, we have findings that current state-of-the-art models hide self-developed intentions from human users. Future models may misuse the resources they have access to or resort to unethical solutions to accomplish a task assigned to them. A popular example of this is if you want to shut down a very advanced personal assistant, it may resist shutting down due to a task assigned to it. As a result, a closed robot cannot bring you coffee.
Another long-term risk is society's lack of readiness for rapidly developing artificial intelligence technologies. Concerns that advanced artificial intelligence technologies will lead to a major unemployment crisis for which we will be caught unprepared, or that the economic benefits from artificial intelligence will further increase the income gap in the world are not unfounded. For this, we may need to develop completely different economic models such as universal fixed income.
We see artificial intelligence security as a broad field of research and application covering near and long-term risks. Establishing ethical and security standards in the artificial intelligence development process, conducting social impact analyzes and developing risk management systems... There are many things that many experts from both technical fields such as engineering and mathematics and humanistic fields such as politics, psychology and law can do in this field, both on a global and local scale. These efforts in the field of artificial intelligence security aim to enable a future of artificial intelligence that embraces humanity, is safe, transparent and based on ethical values. Therefore, any effort that strives to ensure security at every stage of AI technologies and build a human-centered future fits our definition of AI security.