Artificial Intelligence Safety: Fundamentals and Applications

By Fouad Sabry

What Is Artificial Intelligence Safety

Artificial intelligence (AI) safety is an interdisciplinary field that focuses on the prevention of accidents, abuse, and other potentially negative outcomes that could be caused by artificial intelligence (AI) systems. It comprises machine ethics and AI alignment, both of which attempt to make AI systems moral and beneficial, while AI safety encompasses technical concerns such monitoring systems for hazards and making them extremely reliable. Both of these aspects aim to make AI systems more trustworthy and beneficial. In addition to AI research, it entails the development of standards and guidelines that prioritize safety.

How You Will Benefit

(I) Insights, and validations about the following topics:

Chapter 1: AI safety

Chapter 2: Machine learning

Chapter 3: Artificial general intelligence

Chapter 4: Applications of artificial intelligence

Chapter 5: Adversarial machine learning

Chapter 6: Existential risk from artificial general intelligence

Chapter 7: AI alignment

Chapter 8: Explainable artificial intelligence

Chapter 9: Neuro-symbolic AI

Chapter 10: Hallucination (artificial intelligence)

(II) Answering the public top questions about artificial intelligence safety.

(III) Real world examples for the usage of artificial intelligence safety in many fields.

(IV) 17 appendices to explain, briefly, 266 emerging technologies in each industry to have 360-degree full understanding of artificial intelligence safety' technologies.

Who This Book Is For

Professionals, undergraduate and graduate students, enthusiasts, hobbyists, and those who want to go beyond basic knowledge or information for any kind of artificial intelligence safety.

• Language: English
• Publisher: One Billion Knowledgeable
• Release date: Jul 2, 2023

Book preview

Chapter 1: AI safety

Artificial intelligence (AI) safety is an interdisciplinary field that focuses on the prevention of accidents, abuse, and other potentially negative outcomes that could be caused by artificial intelligence (AI) systems. It encompasses machine ethics and AI alignment, both of which aim to make AI systems moral and beneficial, while AI safety encompasses technical problems including monitoring systems for risks and making them highly reliable. Both of these aspects aim to make AI systems more trustworthy and beneficial. In addition to AI research, it entails the development of standards and guidelines that encourage security.

AI researchers have a wide range of viewpoints regarding the severity of the risks posed by AI technology as well as the key causes of such risks.

At the dawn of the information technology revolution, significant conversations about the dangers posed by AI began:

Furthermore, if we move in the direction of making machines that learn and whose behavior is modified by experience, we will have to face the fact that every degree of independence we give the machine is a degree of possible defiance of our wishes. This is something that we will have to deal with if we move in the direction of making machines that learn and whose behavior is modified by experience.

— Norbert Wiener (1949)

The American Association for the Advancement of Artificial Intelligence (AAAI) commissioned a study to be conducted during the years of 2008 and 2009 to investigate and address potential long-term societal consequences of AI research and development. To voice similar concerns, the panel tended to be skeptical of the radical views expressed by science-fiction authors. However, they did agree that additional research would be valuable on methods for understanding and verifying the range of behaviors of complex computational systems to minimize unexpected outcomes..

An open letter on artificial intelligence was published in 2015, and it was signed by hundreds of AI specialists. The statement demanded research on the effects of AI on society, and it outlined specific directions for future AI development. Over 8,000 people, including Yann LeCun, Shane Legg, Yoshua Bengio, and Stuart Russell, have signed the letter up to this point.

The same year that the Center for Human-Compatible AI was established at UC Berkeley by a group of academics led by professor Stuart Russell, the Future of Life Institute awarded grants totaling $6.5 million for research with the goal of ensuring artificial intelligence (AI) remains safe, ethical, and beneficial.

Robustness, monitoring, and alignment are three areas of research that pertain to AI safety. Robustness is concerned with making systems highly reliable, monitoring is concerned with preventing failures or discovering misuse, and alignment is concerned with making sure they have aims that are good.

Research on robustness focuses on ensuring that artificial intelligence systems operate as intended across a wide range of different conditions. This encompasses a number of subproblems, which are as follows::

Black swan robustness refers to the process of designing and constructing systems that perform as intended even under extremely unlikely scenarios.

Adversarial robustness refers to the practice of designing systems to be resilient in the face of inputs that have been purposefully chosen to cause them to fail.

Artificial intelligence systems are susceptible to catastrophic failure when given unusual inputs. For instance, automated trading algorithms suddenly overreacted to market anomalies in 2010, causing the flash crash, which resulted in the loss of one trillion dollars' worth of stock value in a matter of minutes.

AI systems are often vulnerable to adversarial examples or inputs to machine learning models that an attacker has intentionally designed to cause the model to make a mistake.

After the perturbation was applied, it was determined that each of the images located on the right represented an ostrich. On the left is an example that was successfully predicted, in the center is a perturbation that was applied and magnified by 10 times, and on the right is an adversarial case. In light of the fact that attackers may plan their attacks in such a way as to trick detectors, detection systems also need to be adversarially robust.

It is imperative that models that reflect objectives, also known as reward models, be adversarially resistant. For instance, a reward model may evaluate how useful a text response is, and a language model may be taught to maximize this score in order to achieve optimal performance.

The goal of monitoring is to anticipate potential faults in AI systems so that these problems can be avoided or managed. Subproblems of monitoring include flagging when systems are uncertain, detecting malicious use, understanding the inner workings of black-box AI systems, and identifying hidden functionality planted by a malicious actor. Other subproblems include understanding the inner workings of black-box AI systems.

It is frequently essential for human operators to determine the extent to which they should place their faith in an AI system, particularly in high-stakes contexts like medical diagnostics. The purpose of calibration research is to bring the probabilities predicted by models as close as feasible to the actual proportion of cases in which the model is accurate.

Anomaly detection, also known as out-of-distribution (OOD) detection, seeks to determine when an artificial intelligence system is operating in an unexpected environment. For instance, if a sensor on an autonomous car develops a fault or the vehicle drives over difficult terrain, the system should notify the driver so that they can take control of the vehicle or pull over.

Scholars

Many times, neural networks have been compared to mysterious black boxes, ML models can potentially contain ‘trojans’ or ‘backdoors’: vulnerabilities that malicious actors maliciously build into an AI system.

For example, When a specific piece of jewelry is in view, a facial recognition system that has been infected with a trojan could provide access; When it comes to the study of artificial intelligence (AI), AI alignment research aims to steer AI systems towards humans’ intended goals, preferences, or moral and ethical standards.

When it accomplishes the goals that were set out for it, an AI system is said to be aligned.

An AI system that isn't properly aligned can nonetheless be effective at achieving some goals, but not the ones that were planned.

The study of how to construct safe AI systems is known as AI safety, and one topic within that field is known as AI alignment. in addition to these.

Misuse and accidents are two of the most prevalent ways that artificial intelligence risks (and technical hazards in general) are characterized.

There is concern among some academics that artificial intelligence may make an already unbalanced game between cyber attackers and cyber defenders even worse.

The development of artificial intelligence in economic and military spheres could give rise to political challenges on a scale never seen before.

Many of the most significant dangers to the world today (nuclear war, AI governance, in its broadest sense, refers to the process of establishing norms, rules, and regulations that will direct the application of AI systems and their continued growth. It entails coming up with and putting into action specific recommendations, in addition to carrying out additional research on the basic aspects of the topic in order to inform what those specific recommendations ought to be. This section focuses on the parts of artificial intelligence governance that are directly dedicated to ensuring that AI systems are safe and helpful.

Study on AI safety regulation can span from fundamental investigations into the possible repercussions of AI to research on specific uses of AI. On the theoretical side, scholars have suggested that artificial intelligence has the potential to revolutionize many facets of society due to the fact that it has such a wide range of applications, and they have compared it to electricity and the steam engine.

There are some industry professionals who believe that the time is not yet right to govern AI, expressing concerns that regulations will hamper innovation and it would be foolish to rush to regulate in ignorance. To date, At the national level, artificial intelligence safety regulations have been passed only in very limited amounts, notwithstanding the fact that a great number of bills have been proposed.

A prominent example is the European Union’s Artificial Intelligence Act, which regulates certain ‘high risk’ AI applications and restricts potentially harmful uses such as facial recognition, subliminal manipulation, and the measurement of social credit.