Human Compatible

The book "Human Compatible: Artificial Intelligence and the Problem of Control" presents a comprehensive look at the question of how we can maintain control over a technology that is becoming increasingly complex and powerful. The author, a renowned expert on artificial intelligence, explores both the potential and the risks of AI and suggests that the key to safe and effective AI is to connect it to human values. At a time when AI is beginning to permeate all areas of life - from healthcare to transportation to law - questions about how to ensure ethical and safe AI are becoming urgent. The book offers the reader a detailed look at what happens if AI development spirals out of human control and how technology can be prevented from serving interests that conflict with society's values. Russell stresses that the simple pursuit of technological progress is not enough. The future of AI should be designed with human needs in mind, to serve as a tool to help solve global problems - be it the climate crisis, healthcare or social inequalities. The author draws on examples from history and current technologies, highlighting that the answers to key questions about the safety and ethics of AI are not just a matter of technology, but also of philosophy and the social sciences.

This idea leads us to consider how the design of artificial intelligence (AI) needs to reflect human values and goals to avoid unpredictable and possibly catastrophic consequences. For example, suppose we design AI to manage the environment as efficiently as possible - but without a focus on human needs. The result could be a decision to drastically restrict access to natural resources, causing the living conditions of millions of people to deteriorate. There is an example of this from 2016, when an algorithm running a social network created content with the sole aim of maximising views. It soon became clear that it was spreading false news, because that was what attracted the most attention. This case reminds us that if we optimise AI based on numerical goals alone, it can ignore the wider moral and societal implications. Another example might be the autonomous vehicle. Without proper consideration of human priorities, such a system could interpret 'safe driving' as necessarily protecting passengers, even if it means risking the lives of pedestrians. These situations highlight the need for AI systems that understand not only commands, but also the context and values of society. However, design compatible with human values does not mean that AI should have "human morality". Rather, it is about systems being able to adapt their goals according to the current social and cultural context. Researchers such as philosopher Nick Bostrom have proposed so-called "value corrections" that can ensure that algorithms reflect not only efficiency but also ethics. Such an AI design setting requires complex and multidisciplinary collaboration. Philosophical questions about right and wrong must meet with technical practices for AI development to truly serve humanity. Deciding how to use AI and what values to put at its core is not only a technical choice, but also an ethical one that will affect future generations.

The author highlights the need to create AI that can learn and adapt according to people's current needs and preferences. What does this mean in practice? For example, let's imagine we have an AI system designed to manage a city's public lighting. If it focuses solely on saving energy, it might decide to turn off lights on less busy streets, which, while saving electricity, will increase the risk of crime. Such an example highlights that AI must respond not only to predefined goals, but also to changes and new needs that may come unexpectedly. In 2020, healthcare algorithm developers realized that preconfigured models designed to predict health risk did not adequately account for the needs of minority patients. A learning AI that was able to flexibly adjust its models would be able to respond more flexibly to similar challenges, which would mean more equitable healthcare for all. An important aspect of adaptive AI is the ability to take in new information and learn from it. An example is autonomous systems in the manufacturing sector that have adopted adaptive control methods. In this way, the system can continuously adapt to changes in production, reducing the need for manual intervention and minimising losses. However, AI adaptation alone is not enough. Systems must be able to understand the context of decisions. Without this, they run the risk of adapting in a direction that does not reflect the actual needs of humans, which can lead to undesirable outcomes. This is a fundamental difference between adapting to blindly defined rules and adapting based on a comprehensive understanding of the environment. To avoid these problems, the researchers recommend implementing mechanisms that allow AI to perceive and interpret feedback from humans. This would prepare AI to maintain long-term consistency with human needs and preferences. This approach to AI design helps minimize risks while paving the way for systems that can become truly beneficial helpers in everyday life.

The book also addresses the issue of transparency and interpretability of AI algorithms, which means that humans need to understand how and why AI makes certain decisions. Let's imagine that AI is making decisions about approving loans at a bank. If the process is opaque, mistakes or inequities may happen that clients will not be able to understand or address. If an algorithm automatically rejects an application based solely on an obscure formula, the question arises as to whether its intent was truly fair. In 2018, there was a well-publicized case in the United States where an algorithm used in court systems to assess offender risk disadvantaged certain populations based on biases hidden in the data. The consequence was unfair access to sentencing, and pressure has increased to develop more transparent algorithms that can be fair to all. This case highlights why interpretability is key. AI transparency is also important in healthcare. Imagine an algorithm that recommends treatments to doctors. If the doctor doesn't understand why the AI suggested a particular course of action, uncertainty can arise and put the patient's health at risk. Being able to explain its decisions would allow AI to be a real asset to professionals, instead of limiting them. To achieve transparency, researchers are proposing techniques such as "black box debugging," which allow insight into the internal processes of a system. These techniques help identify which factors influence an AI decision, making it more understandable. At the same time, there is a trend to replace classical algorithms with "white box" systems that work on principles more easily accessible to the public. Last but not least, transparency and interpretability contribute to public trust in AI technologies. When people know that they can understand and verify AI decisions, they are more open to using it and more willing to accept it in their daily lives.

Russell warns of the serious risks associated with autonomous systems that could evolve beyond human control. Imagine, for example, a military drone equipped with an algorithm to identify and eliminate threats. Without clearly defined constraints or proper security measures, this drone could mistakenly regard civilians as enemies. In such cases, autonomous systems could become dangerous not only for security but also for the very principles of humanity. A real-life example warns us - in 2020, it was confirmed that a military autonomous drone carried out an attack during an operation in Libya on its own accord, without direct human intervention. This incident shocked the public and reopened discussions about the risks associated with autonomous weapons. It shows that the unpredictable behaviour of AI can lead to potentially tragic consequences if we fail to maintain full control over it. However, similar challenges are not limited to the military domain. Autonomous systems in industry or transport, such as robotic machines in factories or self-driving cars, can cause damage and endanger lives if they fail. For example, if an autonomous car fails to correctly recognise a pedestrian, a fatal accident could result. It is therefore absolutely essential to put safety mechanisms in place to prevent these technologies from exceeding the limits of their scope. Experts warn that clear rules need to be established that set out when and how AI can operate autonomously. So-called 'kill switch' mechanisms or automatic shutdown functions would allow humans to intervene and disable the system in the event of danger. Indeed, without these precautions, autonomous systems could actually start to act beyond human control. The overall goal is to develop AI that can both support us and respect the limits we set. Only in this way can we safely reap the benefits of AI without risking unforeseen threats.

Multidisciplinary collaboration is important in developing ethical and safe AI. Technological advances and the rapid development of AI require not only technology experts, but also philosophers, ethicists, lawyers and social scientists to ensure that AI serves all people and not just a narrow group or a single purpose. For example, imagine that a development team composed entirely of engineers designs an algorithm for the healthcare sector. Without consulting doctors or ethical specialists, they could focus only on the effectiveness of the treatment, ignoring complex moral issues such as the protection of patients' personal data. In the past, it has been shown that such one-sided solutions often lead to situations where the technology does not reflect the real needs or concerns of users. In 2019, several universities teamed up with tech giants to form an interdisciplinary research team to investigate the impacts of AI on society. They found that it was the collaboration of experts from different disciplines that was able to uncover potential risks and weaknesses in the system that would have remained hidden to technical experts. This approach became an inspiration for other companies to bet on multidisciplinary teams when developing AI. Philosophers also make an important contribution, bringing the perspective of values and moral principles. Philosopher Shannon Vallor, for example, has proposed the concept of 'ethical design', which encompasses not only the technological but also the values aspects of AI development. This concept assumes that AI designers must consider, in the design process, what consequences the technology will have on people's lives. Lawyers and policy makers have an equally essential role to play in setting the legal framework for the responsible use of AI. Without their oversight, technology could develop in a way that disrespects fundamental human rights or privileges certain groups at the expense of others. This multidisciplinary approach allows for the creation of AI that is not only technologically advanced, but also ethically acceptable and safe for the general public. It is a way to ensure that advances in AI will be of positive benefit to society as a whole.

The book also focuses on AI's ability to reflect not only pragmatic goals, but also the moral and emotional aspects of human decisions. Nowadays, it is no longer enough for AI to simply carry out given tasks - it also needs to understand that some decisions have deeper meaning and affect people on an emotional level. Let's imagine that AI helps in the selection of job candidates. If it focuses only on numbers and performance, it may overlook values such as compassion, teamwork skills or creativity, which are key in many positions. An employer may find themselves with a team of admittedly high-performing but emotionally distant employees, which will affect both the work environment and morale. One example is a project to develop AI therapy assistants to help patients with mental health issues. These assistants need to distinguish between when it is appropriate to give advice and when it is appropriate to just listen. If the AI is unable to take into account the emotional needs of the client, patients may feel misunderstood or even harmed. This project highlights the importance of AI that understands human emotions and can adapt its behaviour. It is equally important for AI to distinguish between different cultural values that may influence decision-making. In Asia, for example, there may be a greater emphasis on collectivism and respect for authority, whereas in Western societies, individuality and openness are more valued. AI that understands these differences would be better able to serve people from different backgrounds. In practice, this ability to reflect emotional aspects manifests itself in customer service, for example, where AI assesses the tone of a customer's voice to better recognise their satisfaction or frustration. In this way, it can personalise its response and show that it 'understands' and adapts to the customer's needs. To achieve this level of interaction, AI needs to be equipped not only with technical but also ethical standards that can account for these complex aspects of human behaviour. To be truly beneficial to society, AI needs to be able to do more than just perform tasks - it needs to understand why things are important to humans that often cannot be easily quantified.

The book "Human Compatible" points out that AI must be designed in accordance with human values in order to serve safely and responsibly. The author emphasizes the need for adaptive AI that understands human needs and can make decisions in a moral and emotional context. At the same time, he warns of the risks of autonomous systems that could act beyond human control if they are not equipped with safety mechanisms. Multidisciplinary collaboration is key to the development of ethical AI that respects the wider social context. Artificial intelligence should not only be efficient, but above all humane, opening the way to a better and more responsible future for all. Do you have feedback for us? We'd love it if you let us know your opinion on our content! Just send an email to feedback@rewio.app with the title of this book in the subject line and share your impressions!