AI and Decision Making

Artificial Intelligence: The Umbrella Term

AI is a broad interdisciplinary field encompassing the development of machines capable of performing tasks that typically require human intelligence. These tasks include:

Learning
Reasoning
Problem-solving
Perception

AI aims to simulate human cognitive functions to enable machines to perform complex actions and make informed decisions across various applications, ranging from autonomous systems and robotics to data analytics and decision support systems.

Machine Learning: Algorithms and Data

Machine learning (ML) is a subfield of AI focused on enabling computers to learn from data without explicit programming. It involves the use of algorithms that can automatically detect patterns, make predictions, and improve their performance over time through experience. Key aspects of machine learning include:

Algorithms: ML algorithms are designed to identify and learn from patterns in data. These algorithms can be broadly categorized into:
- Supervised Learning: Algorithms learn from labeled data, where the input data is paired with corresponding output labels. Examples include linear regression, decision trees, and neural networks.
- Unsupervised Learning: Algorithms learn from unlabeled data, where the input data is not paired with output labels. Examples include clustering, dimensionality reduction, and association rule mining.
- Reinforcement Learning: Algorithms learn through trial and error, interacting with an environment and receiving feedback in the form of rewards or penalties. Examples include Q-learning and policy gradients.
Data: Data is the foundation of machine learning. ML algorithms require data to learn and make accurate predictions. The quality and quantity of data can significantly impact the performance of ML models. Data preprocessing techniques are often used to clean, transform, and prepare data for ML algorithms.
Classical Methods: The course focuses on classical machine learning methods, providing a foundation for understanding more advanced techniques. Classical methods include linear regression, logistic regression, decision trees, support vector machines, and k-nearest neighbors.
Human vs. Algorithmic Decision-Making: A key theme is the comparison of human decision-making processes with those of machine learning algorithms. This comparison explores the strengths and weaknesses of both approaches, as well as the potential for combining human and algorithmic decision-making to improve outcomes.

Generative AI: Large Language Models

Generative AI encompasses large language models such as ChatGPT, image generation technologies, video recognition systems, and more. While these applications represent advanced achievements in AI, this course prioritizes understanding the fundamental decision-making processes that underpin these technologies. Key aspects of generative AI include:

Large Language Models (LLMs): These models, like ChatGPT, are designed to generate human-like text. They are trained on vast amounts of text data and can perform tasks such as language translation, text summarization, and content creation.
Image Generation: Generative AI can create realistic images from textual descriptions or other input data. This has applications in art, design, and entertainment.
Video Recognition: Algorithms can analyze video content to identify objects, actions, and events. This technology is used in surveillance, autonomous vehicles, and media analysis.
Focus on Decision-Making: Rather than diving into the complexities of these advanced applications, the course concentrates on the core decision-making principles that drive AI systems. This foundational knowledge is crucial for understanding how AI models learn, reason, and make predictions.
Ethical Considerations: The course also addresses the ethical implications of generative AI, including issues related to bias, privacy, and the potential for misuse.
Future Trends: Students will gain insights into emerging trends in generative AI and how these technologies are likely to evolve in the coming years.

Course Focus: Psychology of Decision Making

The class explores the psychology behind both human and algorithmic decision-making processes, often using the terms "AI" and "prediction" interchangeably to highlight their overlapping functions. A core focus involves a detailed examination of how algorithms operate, dissecting their methods and comparing their outputs to those of human experts. This comparative analysis aims to uncover potential deficiencies and biases inherent in human decision-making, shedding light on areas where algorithms may offer more consistent or accurate results

Ethical Implications of AI in Business

The course addresses the anxiety around using AI in life, particularly regarding the role of entry-level employees and the value of human skil

ls.Example: "Mary" Show Clip

A clip from the show "Mary" illustrates a business decision made by an algorithm to cancel a show because it didn't align with the algorithm's "taste clusters," despite high ratings. This highlights the potentially unsettling nature of algorithmic decision-making.

IBM's Warning (1979)

IBM cautioned in 1979 that computers should not make management decisions because they cannot be held accountable. Yet, computers now make many management decisions.

Algorithms in Everyday Life

Examples of algorithmic decision-making include:

Netflix recommendations.
TikTok's addictive algorithm.
Spotify, Amazon, and Stitch Fix recommendations.
News curation on Facebook, Apple News, and the New York Times, including headline variations.
Google Maps for navigation.
Dynamic pricing on Amazon, Target, and Walmart.
Quant investing, where algorithms manage investments.

Algorithms are used wherever there is data and a need for prediction.

The Lens Model: A Psychological Framework

The lens model is a 100-year-old framework for understanding judgments. It involves:

The Truth (Globe): The actual state of the world.
Cues: Observable pieces of information.
Decision Maker/Judge: Combines cues to make a judgment.

Concrete Example: Guessing Age

Given cues like wrinkles, hair color, height, and income, one makes an educated guess about a person's age. The decision-maker weighs each cue.

Example Cues:

Wrinkles, Hair color, Height, Income

Considerations for each cue:

Wrinkles: Increase with age, but newborns are wrinkly.
Hair Color: Can be changed with dye.
Height: Good indicator for children, less so for adults.
Income: Tends to increase with age.

The goal is to align cue weighting with the most reliable indicators for accurate judgment.

Algorithms vs. Human Judgment

Algorithms also combine and weigh cues to produce an answer. The class will show that algorithms often outperform humans at this task.

Linear Regression: The Math Behind Algorithms

Linear regression, a statistical technique over 100 years old, is used by these algorithms.

Formula:

$y = b0 + b1x1 + b2x2 + … + bnx_n$

Where:

$y$ is the prediction.
$x_i$ are the cues.
$b_i$ are the weights (betas) assigned to each cue.

Correlation

Correlation measures the relationship between prediction and reality.
$1.0$ : Perfect correlation.
$0.8$ : Very high correlation (rare in the real world).
$0.4$ : Moderately strong correlation.
$0.0$ : No correlation.
Negative correlation: Indicates an inverse relationship.

R-squared

R-squared indicates how much of the prediction the model captures (also matches to 1).

Partial r

Partial r indicates how much a single variable contributes to the prediction.

How Humans and Algorithms Differ

Algorithms change weights to optimize predictions. Humans rely on intuition and are prone to irrelevant factors and biases.

Comparing Human Experts to Algorithms

When given the same data, algorithms often outperform human experts.

Clinical Psychologists Example

Clinical psychologists' correlation: $0.28$
Algorithm's correlation: $0.42$ (50% better)
Adding the psychologist as a predictor provides a tiny boost, but the best psychologists are still worse than the model.

Predicting Graduate School Performance Example

Human correlation: $0.33$
Model correlation: $0.57$ (almost twice as high)
Even the best students performed with correlation of $0.48$ .

NBA prediction Experiment Results

Class average was 58% (50% is random guessing)
Model average was 79%
There were no proven correlations between time spent taking test, and NBA knowledge.

The Bootstrap Model

A bootstrap model is created based on expert predictions before the actual outcomes are known. It tries to mimic the expert's behavior, but it behaves more consistently and often performs better.

Inconsistency in Expert Judgment

Experts are often inconsistent due to various factors:

Not being robots: We Cannot precisely apply the same rules over time.
Heuristics, especially when tired.
Difficulty discerning valuable vs. bad indicators; models do this dynamically.
The accuracy of judgment does not need to improve when more information is give to humans.

Hodgkin's Disease Example

Expert correlation: Approximately 0.
Regression model correlation: $0.42$ .

Radiologist Example

Correlations among experienced radiologists are low (often below $0.4$ ).
Self-correlations (same doctor, same scans weeks later) ranged from $0.6$ to $0.92$ .
Experts are not consistent.

Doctors Prescribing Opioids Example

Opioids were shown to be perscribed, later in the day, more often.

Pigeon Example

Pigeons can be trained to identify cancerous tumors with 80% accuracy. A collective of 16 pigeons achieved 99% accuracy, exceeding human radiologists.

Summary of human judgement compared to algorithms

Human use heuristics.
Feedback cycles don't work as well.
Configurable rules (stories) are not very useful.
Not robots.

Why Algorithms Excel

Consistency: Always apply the same rules.
Ability to process large datasets.
Objective decision-making: Free from biases.

Algorithmic Selection and Bias

Algorithms can help avoid human bias in selection processes.

Symphony Auditions: Screening auditions increased women's participation from 5% to 36%.
Name Bias: Studies show that white names receive 50% more callbacks for interviews.
So now every symphony does the audtions in a blind manner.

AI in Interviews

AI-evaluated interviews are becoming more common, using visual and speech processing algorithms. However, there's pushback due to concerns about algorithmic bias and fairness.

The Impact of Consumer-Facing Algorithms

Consumer-facing algorithms aim to maximize engagement but can have negative consequences:

Getting stuck in loops (e.g., only seeing crime shows on Netflix).
Exposure to revenue-focused content (e.g., TikTok shop). And consumers want the algorithm to predict their needs even better.
Potential for political polarization (e.g., exposure to extreme views).
Consumers can be polarized due to what they see.

Algorithmic Transparency Desired

60% of TikTok and 46% of Instagram users feel worse off because of these platforms.

The Digital Fairness Act and Design Interventions

The EU's Digital Fairness Act proposes limits on manipulative designs, such as infinite scroll. The UK has a design code to protect consumers. Companies are not interested in your mind, but interested in your attention as people will also pay money to eliminate certain sources.

The Role of Humans in an Algorithmic World

Focus on skills that are difficult for algorithms to replicate:

Handling rare events and broken light cues.
Evaluating the sanity of something and making judgments beyond data.
System one tasks - instinctive reactions and human touch.

Some systems that AI are behind on are:

Face recognition
Speech recognition
Physical tasks requiring dexterity, balance, coordination (e.g., construction).

Algorithms are less helpful in cases with no objective data when hiring a personal coach for personality. It would be hard to predict one's personality.

The Future Combination Approach

Combining humans and algorithms is key:

Build algorithms that include human judgment for rare events.
Give experts algorithms output as an anchor for decision-making.

Focus on soft skills, as management may be the ideal in a world where algorithms are coaches.

Experts are bad at using cues consistently, whereas algorithms dont. Experts hopefully will still exist.

Jobs in the Future