Why AI as a Separate Field?:
AI emerged around 1956 as a distinct field alongside other disciplines like statistics, economics, and operations research. These fields shared a common approach of maximizing expected utility and making rational decisions to solve problems. However, AI aimed to tackle different types of problems using these same approaches, encountering resistance from existing fields.

AI’s Unique Contributions:
AI sought to handle large-scale data inference, beyond the capabilities of statistics, which focused on smaller datasets. AI pioneered methods for symbolic reasoning, enabling computers to understand and manipulate abstract concepts, a task challenging for other fields. AI advanced game playing, a domain where traditional methods struggled, demonstrating the potential of AI’s unique approaches.

Challenges in Defining AI:
Defining AI precisely remains a challenge, as it encompasses a broad range of subfields and applications. AI’s goals vary depending on the context, making it difficult to pinpoint a single, overarching objective. The field continues to evolve rapidly, with new developments and applications emerging frequently.

Human-Centered AI:
Human-centered AI emphasizes the importance of designing AI systems that align with human values and goals. It seeks to create AI systems that augment human capabilities, rather than replace them, and that are transparent, accountable, and fair. Human-centered AI aims to address societal concerns about AI’s potential negative impacts and ensure that AI benefits humanity as a whole.

The Difference Between Software and AI:
AI focuses on addressing problems with uncertainty, where there is no definitive answer, such as explaining quantum physics with rap lyrics. Traditional software engineering deals with complexity and building systems that do the right thing.

Human-Centered AI:
Involves early engagement of multidisciplinary teams, including ethicists, lawyers, sociologists, and multicultural perspectives. Considers the impact, benefits, and harms to all people involved in the AI system. Gives appropriate control to humans and aims to augment rather than replace them. Emphasizes transparency, accountability, security, privacy preservation, and fairness.

Multiple Constituencies in Human-Centered AI:
Considers not only the user but also other stakeholders affected by the AI system, such as pedestrians, other drivers, and society as a whole. Examines the impact on urban sprawl, traffic patterns, and unemployment. Takes a broader approach to problem-solving, leading to better acceptance and fewer resentments.

Benefits of Human-Centered AI:
Addresses the real goals of helping people and avoiding narrow optimization. Enhances the accuracy and robustness of AI systems by incorporating diverse perspectives and considerations. Promotes the development of AI systems that align with human values and societal needs.

Dimensions of Automation and Control:
The Society of Automotive Engineers (SAE) defines six levels of automation for self-driving cars, ranging from level 0 (no automation) to level 5 (full automation). As automation increases, the user’s role in operating the vehicle decreases. Ben Schneiderman proposed a two-dimensional framework for automation and control, allowing for varying degrees of both in different situations.

Choice and Transparency:
Choice refers to the ability to provide advice to the automation rather than taking direct control. Transparency refers to the degree to which the machine makes the user aware of its context and decision-making process.

Certainty and Speed:
Certainty refers to the machine’s confidence in its decisions. Speed refers to the rate at which the machine makes decisions.

Conclusion:
These dimensions are important considerations in the design and implementation of machine learning systems.

Evolution of AI Textbooks:
AI textbooks have shifted from focusing solely on algorithms to emphasizing data and objectives. In 1995, the approach was to maximize a given objective using algorithms. In 2003, the focus shifted to data acquisition and improvement. By 2020, the emphasis was on defining and maximizing objectives, with algorithms and data readily available.

Importance of Defining Objectives:
Defining clear objectives is crucial for successful machine learning. Traditional machine learning approaches relied on external guidance for objectives. This approach is flawed as it overlooks the significance of understanding and defining what needs to be optimized.

Challenges with Data:
Data quality and reliability are significant concerns. Data issues often arise from configuration errors rather than code errors. Strong typing and debugging tools for data are lacking.

COMPAS Case Study:
COMPAS is a recidivism prediction system used in criminal justice. ProPublica’s analysis revealed bias against black defendants. North Point, the creator of COMPAS, argued for calibration accuracy across racial groups. John Kleinberg demonstrated the mathematical impossibility of optimizing both calibration and fairness simultaneously. The resolution lies in societal values and determining fair metrics.

Societal Values and Metrics:
Determining fair metrics involves societal values and trade-offs. William Blackstone’s quote highlights the importance of weighing false negatives and false positives. Defining societal values and metrics is essential for fair and ethical AI systems.

Multidisciplinary Teams:
The COMPAS case study highlights the importance of multidisciplinary teams. Outside criticism exposed flaws that could have been addressed early on.

Conclusion:
The field of AI has shifted its focus from algorithms to data and objectives. Defining clear objectives and addressing data quality issues are crucial. Societal values play a significant role in determining fair metrics for AI systems. Multidisciplinary teams can help identify potential biases and ensure ethical AI development.

Underlying Assumptions and Access to Ground Truth:
AI systems may fail when underlying assumptions are incorrect, emphasizing the need for diverse stakeholder involvement and considering alternative perspectives to mitigate biases. Proxy data, such as arrest records, is often used due to the lack of access to ground truth data on crime. Biases can arise in both the arrest and conviction processes, affecting the accuracy of AI systems.

Focus on Internal Metrics vs. Societal Relevance:
The emphasis on internal accuracy metrics like 60% accuracy may not be meaningful to stakeholders and society. ProPublica’s approach of considering the harm caused to individuals is more relevant.

Justice and Implementation:
Implementing justice through laws and regulations may not be effective if the implementation process is cumbersome and inaccessible. Automating and simplifying processes using AI can improve access to justice.

Importance of Multidisciplinary and Multicultural Teams:
Diverse teams identify more use cases, expose more threat surfaces, allow for better generalization, and discover data coverage gaps.

Synthesizing Different Values for AI Systems:
Anthropic’s idea of constitutional AI involves defining principles and values for AI systems, allowing the public to contribute to the synthesis of these values.

Adversarial Attacks and Human Susceptibility:
Adversarial attacks, where small changes to inputs can fool AI systems, raise concerns about the reliability of AI. However, humans are also susceptible to similar attacks, highlighting the need for a balanced perspective.

Adversarial Attacks and Human-Centered Defenses:
The advertising industry exploits the quirks of human minds through catchy phrases and visual appeals, making individuals vulnerable to adversarial attacks. A partnership between AI and humans can be beneficial in developing human-centered defenses against these attacks.

Measuring Metrics and Finding Sensible Data:
Metrics used for different stakeholders often involve a tradeoff, and measuring easily accessible metrics rather than the right ones is a common issue. In cases of conflicting requirements, finding metrics that are sensible to all stakeholders can be challenging. Analyzing each example, considering available data, and generating additional data can help in finding sensible metrics.

Over-Optimization and Goodhart’s Law:
Optimizing for a specific metric can lead to over-optimization, where the metric is prioritized over the intended goal. Goodhart’s law highlights this issue, emphasizing the vulnerability of optimization efforts.

Approaches to Gathering Data:
Constructing the desired data can be an effective approach when it is unavailable. Combining multiple metrics, rather than relying on a single one, can provide a more accurate representation of the desired outcome.

Addressing Fear of AI:
Unfamiliarity with AI often contributes to fear and apprehension. Encouraging individuals to experiment with AI tools can help reduce this fear. In cases like the writer’s strike, AI tools can enhance the capabilities of individuals rather than replacing them.

Music Industry and Power Dynamics:
The disruption caused by the transition from physical media (CDs) to digital downloads led to a shift in power dynamics in the music industry. Artists were negatively impacted by this shift, receiving less compensation for their work. Unionizing has been effective in helping artists gain more power and secure better deals.

Responsible AI:
Peter Norvig was asked to create a 10-point agenda for discussing responsible AI. He acknowledged the difficulty of coming up with 10 points off the top of his head and suggested referring to his slides for more information.

Power Dynamics and AI:
Peter Norvig emphasized the importance of considering power dynamics when discussing AI and technology shifts. He highlighted the need to ensure that technology advancements do not exacerbate existing power imbalances.