PDF Summary:Artificial Intelligence, by Melanie Mitchell

Over the past decades, artificial intelligence (AI) has experienced cycles of heightened excitement and remarkable progress, followed by periods of disappointment and reduced enthusiasm, often termed "AI winters." In Artificial Intelligence, Melanie Mitchell provides an overview of the current state and developmental progress of machine intelligence, exploring its future potential while navigating its limitations and societal implications.

While acknowledging the remarkable achievements of AI in fields like computer vision and gameplay, Mitchell highlights the still-considerable challenges that separate modern AI from human-level reasoning and comprehension. She examines concerns around AI model robustness and bias, discusses the need for transparency and explainability in AI decision-making, and addresses the ethical quandaries AI raises regarding job displacement and privacy infringement—all while underscoring the fundamental question of whether AI will ever attain true understanding comparable to the human mind.

(continued)...

Practical Tips

• Optimize your personal finances by employing an AI budgeting app that analyzes your spending patterns and suggests optimizations. Unlike traditional budgeting methods, an AI-driven app can provide real-time insights and forecasts based on your financial behavior, helping you to identify areas where you can save money. It might alert you to recurring subscriptions you no longer use or suggest changing your utility provider when it detects a pattern of overpayment.
• You can enhance your intuitive thinking by keeping a daily 'Intuition Journal' where you record decisions made based on gut feelings and the outcomes. This practice will help you become more aware of your intuitive abilities and learn from the patterns in your decision-making. For example, if you have a hunch about taking a new route to work, jot it down and later note whether it saved time or led to an unexpected positive encounter.
• You can enhance your daily productivity by using AI-powered task management apps that specialize in organizing specific types of tasks. For example, if you have a lot of repetitive tasks, find an app that uses AI to automate scheduling and reminders. This way, you're leveraging the AI's strength in handling narrow tasks to free up your time for more complex projects.
• Experiment with cross-disciplinary learning to broaden your cognitive flexibility. Pick two hobbies or subjects that are unrelated, like gardening and coding, and spend time each week learning about both. Then, challenge yourself to draw parallels and transfer insights from one to the other. This could mean using the systematic approach of coding to plan your garden layout or taking the trial and error aspect of gardening to debug a program.
• Explore the potential of AI in your hobby by using available AI tools to analyze or enhance your performance, whether it's in games, art, or any other interest. For instance, if you enjoy chess, use an AI chess engine to review your games and learn from the suggested improvements. This can give you insights into strategic thinking and decision-making that you can apply to other areas of your life.
• Engage with a chatbot on a customer service website and then switch to a human service representative to tackle the same issue. Take note of the differences in how the chatbot handles the query versus the human. This will illustrate the limitations of AI in handling tasks outside of its programmed domain, similar to how AlphaGo is limited to Go.
• Experiment with a smartphone app that uses object recognition to assist visually impaired individuals. By using the app in different environments, you can witness the practical applications of neural networks and understand their impact on accessibility and daily life.
• Try teaching a child or a friend about a new object without using the object's name, and then observe how an AI assistant struggles with the same task. This will help you appreciate the depth of human understanding and the nuances AI currently misses. For instance, explain to a child what a "frisbee" is by describing its shape, use, and the feeling of playing with it, then ask a voice-activated AI to do the same and note the differences.
• Develop creative thinking by sketching everyday objects with unusual features or in abstract forms. Use a sketchpad to draw items like a chair with exaggerated proportions or a car with non-standard shapes. This activity encourages your mind to maintain the essence of the object while altering its conventional markings, reinforcing your cognitive flexibility in understanding the core concept of the object.
• Develop quick decision-making skills by setting up a 'first impression' challenge with a friend. Share pictures of various doorways or objects with each other and guess what's behind them or what the object's function is based solely on the initial view. This activity will help you trust your instincts and make swift judgments, akin to recognizing a slightly open door without extensive examples.

Artificial intelligence systems lack the profound understanding, the ability to generalize, or the versatility that are essential characteristics of human intellect.

Mitchell emphasizes that even the most sophisticated AI technologies lack the depth of understanding that humans, even young children, routinely bring to everyday tasks. Deep learning systems are trained on labeled examples (for example, photos with labels identifying the objects they contain) provided by humans, but these systems do not yet exhibit the human ability to learn new concepts on their own or to form abstractions on the basis of only a small number of examples. Present-day artificial intelligence models do not possess the capability to transfer insights acquired in one area to improve understanding and solve problems in different fields, an ability referred to as knowledge adaptation within AI research.

Mitchell underscores this concept by discussing Bongard problems, which are visual puzzles created to challenge abilities in abstraction and forming analogies akin to human cognition. Bongard problems are effortlessly solved by humans, yet the most advanced deep-learning systems find it challenging to surpass the accuracy of random guesses, despite training on thousands of examples.

Practical Tips

• Enhance your critical thinking by regularly engaging in debates or discussions on topics outside your comfort zone, which will force you to understand and generalize concepts in a versatile manner. By doing this, you'll practice the very human skills of profound understanding and generalization that AI currently lacks. For example, if you're interested in technology, join a book club that focuses on philosophy or sociology to challenge your thinking patterns.
• Engage in a creative project like writing a story or painting a picture, focusing on the emotional and subjective aspects of the creation. As you work on your project, reflect on the unique human elements you bring to it, such as your personal experiences, emotions, and the sensory details you perceive. This exercise will help you understand the depth of human creativity and contextual understanding that AI has yet to achieve.
• Try teaching someone else a new skill without using any labeled examples to appreciate the human ability to learn independently. Choose a simple task, like making a paper airplane, and guide them through the process using only verbal instructions and answering their questions. This exercise will give you a firsthand experience of the human capacity for independent learning and the interactive nature of our learning processes.
• Develop a habit of reflective journaling to connect insights from different experiences. After each significant activity or event in your day, take a few minutes to jot down what you learned and how it might apply to other areas of your life. If you attended a music concert, for example, reflect on how the harmony in music could inform the way you approach teamwork in your professional life.
• Use analogies to solve problems by relating them to different domains with which you're familiar. When faced with a new challenge, try to think of a similar situation you've encountered before and how you resolved it. For instance, if you're trying to improve a recipe, consider how you've adjusted variables in a different context, like balancing elements in a work project, and apply that thinking to your cooking.
• You can enhance your adaptability by engaging in improvisational theater classes to practice responding to unpredictable scenarios. Improv theater requires you to quickly adapt to new information and situations, mirroring the adaptability that AI currently lacks. By participating, you'll train your brain to think on its feet and transfer knowledge to various contexts, a skill that's valuable in both personal and professional settings.
• Create a simple game with friends where you each take turns designing visual patterns for others to interpret. Use everyday objects or draw shapes on paper, and come up with a rule that distinguishes two sets of patterns. This social activity not only taps into the human ability to discern patterns but also fosters communication skills as you explain your reasoning and understand others' perspectives.
• Engage with interactive online platforms that allow you to 'train' a simple AI model, such as sorting images or classifying text. Pay attention to how the model's performance varies with the type and amount of data you provide. This will give you a practical sense of how data quality and quantity impact AI accuracy without needing to delve into the complexities of AI development.

Adversarial attacks expose the fragility of AI systems, highlighting that their grasp of concepts falls short when compared to the cognitive processes of humans.

A major concern regarding AI is its vulnerability to adversarial attacks, stemming from its inadequate understanding. Researchers have shown through their work that even sophisticated deep learning models can be tricked with methods that are unexpectedly straightforward. Subtle changes in images that might escape human detection can cause a convolutional neural network to misclassify an item, confusing it with another category, or cause a voice recognition system to misinterpret a command, perceiving harmless melodies as an instruction to download harmful software.

The fragility of deep learning systems, illustrated by their susceptibility to adversarial attacks, has spurred new interest in AI research on understanding and addressing the potential risks of applying these systems in the real world. As AI systems become more integrated into practical applications such as internet search and autonomous vehicles, the need for them to perform reliably and withstand various attacks is increasingly critical.

Context

• In image recognition, adding noise to an image can cause a model to misidentify objects. In audio processing, slight alterations to sound waves can lead to incorrect voice command interpretations.
• Ongoing research aims to develop more robust models that can resist adversarial attacks. This includes techniques like adversarial training, where models are trained on adversarial examples to improve their resilience.
• Humans recognize objects based on high-level features and context, while CNNs focus on low-level patterns. This difference makes CNNs more susceptible to being fooled by minor alterations that do not affect human perception.
• Misinterpretation of voice commands can lead to unauthorized actions, such as executing commands that could compromise security, like unlocking devices or accessing sensitive information.
• In image recognition, adding noise to an image can cause a model to misidentify objects. In audio processing, slight alterations to sound waves can lead to incorrect voice command interpretations.
• Vulnerabilities in AI can be exploited for malicious purposes, such as causing autonomous vehicles to misinterpret road signs or bypassing security systems.

Assessing how artificial intelligence measures up to human standards in fundamental cognitive tasks.

Artificial intelligence has matched or surpassed human abilities in certain tasks involving perception and gameplay, although the benchmarks used to assess human-like performance can occasionally lack depth.

Mitchell argues that claims about AI matching human proficiency in tasks such as object recognition or surpassing humans in games often fail to consider the different approaches that humans and machines use to perform these tasks, as well as the different criteria used to evaluate their performance. Learning systems that have been educated using an extensive collection of categorized data often exceed human performance when assessed on new examples from these datasets. The comprehension that these systems have typically lacks the deep and subtle understanding inherent to human intelligence, underscored by how effortlessly adversarial examples can deceive them.

For example, while a ConvNet that has been educated using the ImageNet dataset might outperform humans when evaluated by the "top-5" accuracy standard, this criterion might not be entirely indicative of its precision, as it only shows that the correct image classification is among the five guesses the system is most sure about, not necessarily that it is the system's first choice.

Context

• The deployment of AI in perception and gameplay raises questions about its application in real-world scenarios, where ethical considerations and the potential for unintended consequences must be addressed.
• Human-like performance involves understanding context, nuance, and ambiguity, which many benchmarks fail to measure, as they often focus on isolated tasks without considering broader situational awareness.
• Humans can adapt to new and unforeseen situations using creativity and problem-solving skills. Machines may require retraining or reprogramming to handle scenarios outside their initial training scope.
• Humans naturally incorporate context and experience into their decision-making processes, which is not easily replicated or evaluated in AI systems that rely on pattern recognition without deeper contextual awareness.
• Human intelligence encompasses emotional and social understanding, enabling empathy and complex social interactions, areas where AI is significantly limited.
• AI models, especially those based on deep learning, are highly sensitive to these small perturbations because they rely on specific patterns in the data that can be easily disrupted.
• The "top-5" accuracy metric does not necessarily reflect a model's ability to make the most confident and correct prediction first, which can be crucial in applications where precision is critical.

On more complex tasks requiring contextual understanding and commonsense reasoning, current AI still falls far short of human capabilities.

The author emphasizes that despite considerable advancements in AI's ability to discern objects and understand spoken language, the most advanced systems still fall short of human capabilities in understanding context, applying common sense reasoning, or planning for future occurrences. As one of many examples, automatic caption-generating systems take as input an image and generate a caption describing the image, with results both amusing and sometimes surprisingly accurate. When faced with visually complex scenarios that require a level of understanding, these systems fail to deliver satisfactory results.

For example, when shown a picture of a soldier embracing her pet at the airport, adorned with a celebratory "Welcome Home" balloon, the most sophisticated programs would still simply describe the scene as "A person hugging an animal." As Mitchell notes, humans immediately perceive the nuances, emotions, and implied meanings inherent in the illustrated scenario, which the accompanying text does not communicate.

Practical Tips

• Improve your personal security by using AI-driven apps that recognize faces and voices to create custom alerts. Download an app that allows you to input voices and faces of friends and family, and then set it to alert you when an unrecognized person or voice is detected in your home. This way, you're immediately informed of potential intruders or unknown visitors.
• Use images as prompts for short stories or poems to explore the depth beyond their visual appearance. Choose an image and write a narrative or poem inspired by it, aiming to capture the essence and nuances that a straightforward caption might not. This exercise can help you delve deeper into the storytelling potential of images and enhance your creative writing skills.
• Create a personal blog or online gallery where AI-generated captions are paired with abstract art or ambiguous images. This allows you to explore how AI interprets less straightforward visuals and provides a platform for others to see the amusing or surprisingly accurate results. It can become a space for discussion and reflection on the intersection of AI, art, and language.
• Try interpreting the subtext in various forms of media as a way to practice nuance recognition. While watching a film or reading a novel, pause to consider what the characters are not saying directly but are communicating through their actions, tone, and expressions. Discuss your interpretations with others to compare perspectives and deepen your understanding of subtle communication cues.

The inability of AI to clarify its thought processes and to apply knowledge in novel situations also distinguishes it from human intelligence.

Deep learning systems, unlike human cognitive processes, typically lack the ability to articulate the reasoning behind their decisions, a feature that is commonly present in humans when they provide justification for their specific choices or conclusions. As deep neural networks grow more complex, guaranteeing their dependability is increasingly difficult because their understanding is encapsulated within numerous numerical values that change during the training phase. Although a technique can be devised to display the "features" a network identifies, these visual representations do not provide an intelligible explanation for human understanding. Deep-learning systems are unable to transfer knowledge from one domain to enhance learning in another, necessitating a fresh start with each new task they are trained for.

Practical Tips

• Engage in regular debates with a friend on various topics. Choose subjects you both have an interest in and take turns presenting your viewpoints and justifications. This will sharpen your ability to articulate your reasoning and expose you to different perspectives. For instance, you might debate the merits of electric cars versus gasoline cars, providing justifications related to environmental impact, cost, and technology.
• Explore the basics of neural networks through interactive web simulations to visualize how numerical values change during learning. Websites like TensorFlow Playground allow you to tweak neural network parameters and observe the effects in real-time, providing a hands-on understanding of the concept without needing any coding skills.
• Try simplifying network concepts by using analogies from everyday life. For instance, compare a network's nodes to a group of friends and the links to the conversations they have. This method can help you grasp the abstract concepts of network features by relating them to familiar situations.

The emergence of artificial intelligence-based technologies has societal and ethical ramifications.

Developing AI systems that are dependable and worthy of trust.

Concerns about the dependability and safety of AI systems arise due to their opaque nature, as well as their susceptibility to biases and intentional interference.

Mitchell emphasizes the difficulties in developing AI that is dependable and ethically sound, highlighting that these challenges are deeply embedded in both technological and socio-political spheres. Our understanding of the limitations of artificial intelligence, along with the lack of transparency in how these systems make decisions, makes it challenging for us to understand their choices and predict their errors, even as these technologies are swiftly adopted for real-world use. Artificial intelligence systems, which develop their understanding from data that mirrors societal biases, can amplify these biases, potentially causing harm to marginalized communities.

Context

• Explainability refers to the ability to describe how AI systems reach their conclusions. Many AI models, while highly accurate, do not provide clear explanations for their decisions, which can be problematic in critical applications like healthcare or criminal justice.
• The algorithms themselves can introduce bias if they are not carefully designed and tested. This can occur if the algorithm favors certain outcomes over others due to the way it processes data.
• Building public trust in AI systems is crucial, yet challenging, as people may be wary of technologies they do not understand or that have previously caused harm.
• AI systems often use complex algorithms that involve numerous parameters and layers of computation, which can obscure the decision-making process even from the developers who created them.
• AI systems may lack cultural sensitivity, leading to misinterpretations or inappropriate responses that can negatively impact marginalized groups.

In the European Union, efforts are being made to ensure that individuals have the ability to comprehend the logic behind decisions made by AI, emphasizing the improvement of transparency and the assurance of equity.

A diverse group of specialists, organizations, and authoritative bodies are working diligently to overcome the challenges linked to dependable artificial intelligence by establishing standards and protocols that guide the development and deployment of AI technologies, with a focus on improving their transparency, strengthening their robustness, and minimizing bias. In 2018, a legislative mandate was established requiring that automated systems explain their decision-making processes affecting EU citizens in a clear, comprehensible, and accessible manner, using language that is easy to understand. The vague language used in the regulation results in uncertainty regarding its interpretation and implementation.

Practical Tips

• Encourage transparency by asking service providers about their AI ethics. When interacting with companies that use AI, such as chat support or recommendation systems, inquire about their AI ethics policy. This could be as simple as sending an email to customer service or asking a sales representative. Your interest can signal to businesses that consumers care about ethical AI, potentially influencing them to adopt standard practices.
• You can become more informed about your rights by researching the General Data Protection Regulation (GDPR) and its requirements for automated decision-making transparency. Understanding these regulations will empower you to question and challenge automated decisions made about you by companies operating within the EU. For example, if you receive an automated credit score decision, you can request an explanation of how the score was calculated.
• Engage in role-playing exercises where you act out scenarios involving the interpretation of vague regulations or instructions. Swap roles with a partner and discuss the different interpretations you each had. This can sharpen your critical thinking and negotiation skills when dealing with ambiguity.

Debates persist regarding the importance of AI systems developing deep understanding rather than just surface-level skills to ensure their reliable use in important sectors.

For Mitchell, a central issue is the compromised dependability of these systems due to their lack of adequate understanding. Machines that rely on data, particularly those that employ deep learning techniques, have an intrinsic vulnerability and are prone to unexpected errors, as shown by their vulnerability to intentionally deceptive inputs designed to mislead them. Many experts, including Mitchell, are of the opinion that artificial intelligence must include systems that understand as deeply as humans do in order to achieve genuine reliability. The effectiveness of the machine would be demonstrated by its ability to utilize learned information in new situations, quickly adapt to unexpected changes, and explain its decision-making process in a way that is understandable to humans.

Practical Tips

• Start a habit of questioning the reasoning behind AI recommendations in everyday technology. When you receive suggestions from AI, such as movie recommendations on streaming services or product suggestions while shopping online, take a moment to consider why the AI might have made those suggestions. If the reasons aren't clear or seem shallow, use the platform's feedback mechanisms to indicate the recommendation wasn't relevant. This practice encourages the development of AI that can better understand and match human preferences.
• Develop your ability to explain complex ideas simply by starting a blog or vlog where you teach concepts you're familiar with to an audience with no background knowledge. Focus on breaking down the concepts into fundamental parts and use analogies or stories for clarification. This exercise will train you to articulate your thought process clearly, akin to how AI can be designed to explain its decisions in a way that humans can understand.

Considering the societal changes that have been instigated as a result of advancements in artificial intelligence.

The possibility that AI could render numerous occupations obsolete has sparked worries about substantial job displacement and the necessity for novel economic frameworks.

The discussion about AI's role in society is driven by worries that automation could lead to the displacement of human employment, resulting in considerable joblessness, economic upheaval, and profound changes to the societal structure. Mitchell notes that despite the shift where machines have assumed roles once held by people, including those of elevator operators and human calculators, this change has not resulted in enduring, widespread unemployment. The conversation has gained new momentum due to the vast potential of artificial intelligence and robotics to assume positions typically occupied by humans, such as driving vehicles and diagnosing health conditions. The influence of AI on various job markets is challenging to predict because of the unknowns regarding when and how widely it will be adopted, yet some contend that it could balance out job displacement by generating an equivalent number of new positions, as has been the case with past technological progress.

Other Perspectives

• Novel economic frameworks may not be necessary if existing systems can be adapted to accommodate the changes brought by AI, through reforms and policy adjustments.
• Historical evidence suggests that technological advancements, including automation, have often created more jobs than they have destroyed by generating new industries and economic growth.
• The concern about joblessness often does not consider the potential for policy interventions, such as universal basic income or job guarantees, which could provide a safety net for those displaced by automation.
• Economic upheaval is not a foregone conclusion; with appropriate policies and investments in education and training, the transition can be managed to minimize negative impacts.
• In some cases, the introduction of machines to perform certain tasks has improved the quality of the work environment for humans, by eliminating the need for people to perform dangerous, repetitive, or physically demanding tasks.
• The assertion that there has not been enduring, widespread unemployment overlooks the regional and sectoral impacts of technological change, which can be severe and long-lasting for certain areas or industries.
• The effectiveness of AI and robotics in assuming human positions is often industry-specific, and there may be sectors where human labor remains more cost-effective or preferred by consumers.
• Some argue that the core functions and capabilities of AI are well understood by experts in the field, which can allow for more accurate predictions about its impact on job markets than the statement suggests.
• Historical patterns of technological adoption show that once a technology reaches a certain level of maturity and cost-effectiveness, its adoption becomes more predictable and widespread.
• The geographic distribution of new jobs created by AI might be uneven, benefiting certain regions over others, which could exacerbate economic inequalities.
• The quality and pay of the new jobs are not guaranteed to be equivalent to the ones they replace, which can lead to a deterioration in working conditions and living standards for some workers.

The incorporation of artificial intelligence in sectors like surveillance, criminal justice, and healthcare highlights ethical issues related to privacy, fairness, and the appropriate level of technological participation in decision-making processes.

Mitchell notes that the ethical issues associated with AI extend well beyond the possibility of job displacement. It is crucial to stay vigilant so that the advancement of AI technologies does not exacerbate existing societal biases. When systems are trained with data that includes biases, those biases will unavoidably be reflected in the systems' behavior, which can result in detrimental consequences, even if such results are unintended. Systems that recognize faces are quickly being integrated into various applications such as airport security checks, financial transaction authentication, and surveillance operations, despite their propensity for errors and bias against individuals who are not white males, and are even employed in sensitive areas like electoral processes and law enforcement.

Other Perspectives

• AI in healthcare can lead to more personalized and effective treatments by analyzing patient data and identifying patterns that may not be apparent to human practitioners.
• The statement assumes that all biases are negative, but some biases may be reflections of reality or beneficial for certain applications, and thus the goal should be to understand and manage biases rather than to prevent them entirely.
• The statement does not acknowledge the efforts and progress made in the field of AI ethics, where researchers are actively working on developing more equitable and transparent systems, suggesting that the problem is being addressed, albeit gradually.
• Errors and biases in facial recognition systems are not exclusive to non-white males; they can also misidentify women and individuals of other demographics, which is a problem that developers are actively working to address.
• Facial recognition technology can enhance the security and integrity of electoral processes by ensuring that individuals casting votes are who they claim to be, potentially reducing voter fraud.

The development of AI systems with advanced language and cognitive skills has initiated discussions about their possible impact on the interactions between humans and AI, as well as our understanding of our own identity and purpose.

Debates intensify alongside the progression of AI systems in understanding and utilizing language, regarding the possibility that they might achieve or exceed human intelligence levels. What criteria can we use to equate a machine's demonstration of intelligence with that of a human's? Is it essential for the evolution of this machine to include the development of self-awareness, emotional capabilities, and a survival instinct? The progress in artificial intelligence necessitates a reassessment of our understanding of human intelligence, our essential characteristics, and our distinct place in the cosmos.

Practical Tips

• Volunteer for initiatives that require human sensitivity, such as mentoring or counseling. These activities demand a level of emotional intelligence and understanding that AI currently lacks. By participating, you not only contribute to society but also reinforce the value of human connection and empathy. For instance, mentoring a young person can help you realize the impact of personal guidance and support that AI cannot provide.
• Engage with AI in a controlled environment by using AI-powered apps that simulate conversation or companionship, like Replika or Woebot. As you interact with these apps, take note of moments when you wish the AI were more self-aware or emotionally intelligent, and other moments when you're glad it's not. This hands-on experience can give you a personal sense of the potential benefits and drawbacks of emotionally capable AI.