Michael Kearns: Algorithmic Fairness, Bias, Privacy, and Ethics in Machine Learning

Certainly. Here is the English translation of the provided Chinese investment research notes, following all specified rules.

At a Glance

This episode's guest is Michael Kearns, a professor at the University of Pennsylvania and co-author of the new book The Ethical Algorithm. He is a scholar with expertise spanning machine learning theory, game theory, and algorithmic trading. The podcast's main thread explores the challenges and solutions related to algorithms in fairness, privacy, and ethics. The most significant judgment of the entire episode is: Kearns believes there is a fundamental, irreconcilable contradiction in the field of algorithmic fairness—you cannot simultaneously satisfy three seemingly reasonable and desirable definitions of fairness, making this problem far more intractable than algorithmic privacy (which already has the widely accepted solution of differential privacy).

Algorithmic Fairness: A More "Chaotic" Battlefield Than Privacy

Michael Kearns argues that the field of algorithmic fairness is far less mature than algorithmic privacy because it faces fundamental definitional conflicts and philosophical dilemmas.

• Core Contradiction: Kearns points out that theorems already exist in the algorithmic fairness literature proving that you cannot simultaneously possess three "perfectly reasonable, desirable notions of fairness." This means that no matter which definition of fairness you choose, you inevitably sacrifice other equally important aspects. This stands in stark contrast to the privacy field, which has largely reached a consensus that differential privacy is the "correct definition."
• Group vs. Individual: Kearns delves into the tension between group fairness and individual fairness. Current fairness definitions are mostly group-based (e.g., ensuring similar false rejection rates across different races). However, Kearns notes that this group-level guarantee can mask discrimination against specific individuals (e.g., a disabled, Hispanic, low-income woman over 55). He calls this phenomenon "fairness gerrymandering" —just like political gerrymandering, you provide guarantees at a macro level, but at a more granular level, you are effectively achieving this by favoring some groups and discriminating against others.
• Trade-offs Are Inevitable: Kearns emphasizes that pursuing fairness inevitably involves trade-offs. He introduces the concept of a "Pareto curve" : one axis represents the algorithm's prediction error, and the other represents some measure of unfairness. You cannot simultaneously achieve "zero error" and "zero unfairness." You can only choose a point on this curve, a choice that is fundamentally a value judgment and should not be made by computer scientists, but by policymakers and stakeholders. Kearns believes that explaining this curve to non-technical people and letting them make the choice is a key interface for future work.

Algorithmic Privacy: Differential Privacy is a "Savior," But Not a Panacea

Kearns believes that differential privacy provides a powerful and practical solution to data privacy problems, though its application scenarios still have limitations.

• Core Idea of Differential Privacy: Kearns explains differential privacy with a vivid analogy: it compares two worlds—one containing your data, and one without it. Differential privacy guarantees that any "harm" that could potentially occur because your data is included is almost exactly the same as the "harm" that could occur without your data. This means your data itself cannot be the "culprit" for any harm you might suffer.
• Implementation Mechanism: Adding Noise: Kearns notes that the algorithmic mechanism for achieving differential privacy involves "cleverly and decorously adding noise" during the computation process. For example, when calculating the average of a set of numbers, instead of outputting the exact value, you first compute it and then add Gaussian noise with a mean of zero. This way, the output is close to the true average, but it's impossible to reverse-engineer any single input value.
• Limitations and the Future: Kearns acknowledges that differential privacy is not suitable for all scenarios. For instance, when you search Google for "mountain bike," you want Google to know your query precisely, not a noise-distorted result. In such cases, other types of privacy protection ideas are needed, such as protecting privacy through obfuscation—for example, a browser plugin that continuously sends a large number of irrelevant, fake queries to Google while you search, thereby "drowning out" your true intent.

Game Theory and Machine Learning: From "Selfishness" to "Stability"

Kearns argues that the combination of game theory and machine learning is key to understanding how many of today's platforms (e.g., navigation apps, social media) operate, and reveals how "selfish" behavior can lead to collectively suboptimal outcomes.

• The "Selfish" Nature of Machine Learning: Using navigation apps (like Google Maps, Waze) as an example, Kearns points out that these apps essentially compute a "selfish best response" for each user—planning the fastest route for you given the current behavior of all other users. This effectively drives all users towards a Nash equilibrium.
• Equilibrium is Not Optimal: Kearns emphasizes one of the most important lessons from game theory: reaching an equilibrium does not mean everyone gets the best possible outcome. In the navigation example, all users following the app's "selfish" advice could lead to a collective commute time that is significantly higher than other possible solutions. Similarly, on social media, algorithms optimize content for you, showing you only what you like, which can ultimately lead to the polarization of political discourse in society.
• The Time Scale of Algorithmic Trading: Kearns attributes the success of algorithmic trading to its time scale. He believes algorithms excel in high-frequency trading and statistical arbitrage (predicting price movements over seconds to days) because these are pure optimization and control problems. However, for Buffett-style long-term investing (holding for 10-20 years), algorithms need to understand economic cycles, political situations, and human nature. Kearns believes "we are still far from a robot Warren Buffett."

Mentioned Positions

This section is an interview and does not involve specific investable positions.

Judgments Worth Remembering

1. Fairness Has an "Impossible Trinity" (Michael Kearns): Theorems in algorithmic fairness prove that you cannot simultaneously satisfy three seemingly reasonable and desirable definitions of fairness, making this problem far more intractable than algorithmic privacy (which has the widely accepted solution of differential privacy).

2. "Fairness Gerrymandering" (Michael Kearns): Protecting fairness only for macro-level groups like race and gender can mask discrimination against specific intersectional groups (e.g., "a disabled, Hispanic, low-income woman over 55"). This phenomenon is called "fairness gerrymandering."

3. The Trade-off in Fairness Should Be a Societal Decision (Michael Kearns): There is a "Pareto curve" between an algorithm's error and unfairness. Choosing a point on this curve is a value judgment that should not be made by computer scientists, but by policymakers and stakeholders who understand the curve.

4. The Core of Differential Privacy is "Your Data is Not the Culprit" (Michael Kearns): Differential privacy guarantees that any "harm" that could potentially occur because your data is included is almost exactly the same as the "harm" that could occur without your data. This means your data itself cannot be the "culprit" for any harm you might suffer.

5. "Selfish" Algorithms Drive Us Towards Suboptimal Equilibria (Michael Kearns): Navigation apps compute a "selfish best response" for each user, driving all users towards a Nash equilibrium. But game theory teaches us that equilibrium is not optimal, which can lead to a collective commute time significantly higher than other solutions.

6. Long-Term Investing is a "Blind Spot" for Algorithms (Michael Kearns): Algorithms excel in high-frequency trading and short-term statistical arbitrage, but for long-term investing that requires understanding economic cycles, political situations, and human nature (like the Buffett style), algorithms are far from replacing humans.

7. Computer Scientists Should Proactively "Engage" with the World (Michael Kearns): Given the growing impact of computer science on society, computer scientists can no longer work in isolation. They must proactively go out, understand, and participate in broader social, ethical, and political discussions.

8. Algorithmic Fairness Research Needs "Subjective" Input (Michael Kearns): Current fairness definitions are mostly "conjured up" by scholars, lacking surveys of ordinary people's subjective sense of fairness. More behavioral experiments are needed in the future to understand what people truly consider fair.