# Harvard Law School Library Book Talk | Cass Sunstein, ‘Imperfect Oracle: What AI Can and Cannot Do’

Source: https://www.youtube.com/watch?v=MvBdRF3zw5k
Recap page: https://rapidrecap.app/video/MvBdRF3zw5k
Generated: 2025-11-19T22:06:53.385+00:00

---
## Quick Overview

Cass Sunstein argues that while AI algorithms can be less biased and noisy than human judgment in some areas like predicting coronary events or recidivism, they fail in domains involving complex social interactions, context, and human preferences, as illustrated by the failures in predicting song success and the need for human intuition in bail decisions.

**Key Points:**
- Human judgment is inherently biased and noisy, which AI can sometimes correct, as shown in predicting coronary events where AI outperformed human doctors (08:06).
- AI algorithms, however, fail in predicting outcomes sensitive to social dynamics, context, and individual preferences, such as the success of songs or bail decisions (09:59, 10:41).
- The failure in predicting song success (The Music Lab experiment) demonstrated that algorithms struggled with unpredictable elements like social interactions and individual preferences (02:53, 03:10).
- In bail decisions, algorithms often overweigh current offense bias and mugshot bias, while human judges are noisy, leading to complex trade-offs (10:25).
- John Rawls' concept of uncertainty is relevant: some predictions (like war or copper prices 20 years hence) lack a scientific basis for calculable probability (03:04, 03:28).
- The inherent unpredictability in human affairs—like romantic attraction—means that purely data-driven AI approaches will struggle where context, timing, serendipity, or mood are key factors (07:33, 08:09).
- The speaker concludes that while AI excels in certain structured domains, human intuition and judgment remain critical for navigating complex, unpredictable human social systems (04:28, 05:03).

![Screenshot at 08:07: Cass Sunstein discusses the second thesis, 'Algorithms Fail,' illustrating that while AI can be better than humans in some measurable areas \(like predicting coronary events\), it fails in contexts involving human social dynamics and uncertainty.](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-08-07.png)

**Context:** This video captures a book talk at the Harvard Law School Library featuring Cass Sunstein discussing his book, 'Imperfect Oracle: What AI Can and Cannot Do.' Sunstein contrasts the capabilities of AI in predictable domains, such as medical diagnosis, with its limitations in areas governed by complex human behavior, social dynamics, and inherent uncertainty, drawing on examples from criminal justice and cultural trends.

## Detailed Analysis

Cass Sunstein presents two core theses regarding the capabilities and limitations of Artificial Intelligence. The first thesis posits that AI can be superior to human judgment in specific, measurable contexts because AI can be noise-free and less biased than humans, as demonstrated in predicting coronary events (08:06). However, the second thesis argues that algorithms fail when dealing with phenomena governed by social interactions, context, and unpredictable human preferences, such as predicting the success of songs (02:53) or making nuanced bail decisions (10:25). Sunstein references John Rawls' concept of uncertainty to explain that for certain large-scale, long-term events (like war or social revolutions), calculable probabilities are scientifically impossible to form (03:04, 03:28). He uses the example of the "Fragile Families Challenge" (08:07) to show that even sophisticated machine learning models struggled to predict individual life trajectories accurately, performing only slightly better than random guessing. He further contrasts AI with LLMs, noting that LLMs excel at pattern matching but lack the contextual understanding and emotional nuance (like romantic attraction) that humans possess. Ultimately, Sunstein suggests that while AI can augment human decision-making by identifying known biases (like current offense bias in bail decisions), it cannot replace human judgment where irreducible uncertainty and moral obligations are involved.

### Introduction and First Thesis

- Sunstein welcomes the audience and introduces his book 'Imperfect Oracle'; the first thesis states that human judgment is biased and noisy, while AI can be noise-free and less biased, leading to more accurate predictions in certain areas like coronary events (00:05, 08:06).

### Second Thesis

- Algorithms Fail: This thesis details areas where AI fails, including predicting social outcomes like song success or bail decisions, due to unpredictability in human affairs (03:05, 09:59).

### Keynes on Uncertainty

- The speaker cites Keynes to define true uncertainty as situations where no scientific basis exists for calculating probability, such as predicting a European war or the price of copper in twenty years (03:04, 03:28).

### Case Study

- Bail Decisions: The problem of bail involves weighing flight risk vs. crime likelihood. Algorithms are better than noisy human judges on average but fail due to biases like current offense bias and mugshot bias (10:25).

### Case Study

- Song Success (Music Lab): Algorithms failed to reliably predict which songs would be hits because success depends on unpredictable social interactions and preferences (03:05, 03:26).

### LLMs vs. AI-Powered Algorithms

- LLMs primarily match patterns and use probabilistic judgments, while traditional AI algorithms are more deterministic, but both are susceptible to cognitive biases present in their training data (18:22).

### Final Words & Conclusion

- The presentation concludes that humans often overweigh data-driven recommendations when facing inherent uncertainty, and that while AI can help counteract known biases, it cannot solve problems rooted in fundamental unpredictability and moral obligations (09:16, 03:33).

![Screenshot at 00:08: Title slide for Cass Sunstein's book talk: 'Imperfect Oracle: What AI Can and Cannot Do' set against a background of molecular structures.](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-00-08.png)
![Screenshot at 01:56: Slide detailing 'Two Theses; the First Thesis,' outlining that human judgment is biased/noisy while AI can be less biased and identify unknown biases \(01:56\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-01-56.png)
![Screenshot at 03:05: Slide illustrating the Second Thesis: 'Algorithms Fail, 1,' noting that algorithms cannot predict shocks like pandemics or foresee effects of social interactions \(03:05\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-03-05.png)
![Screenshot at 08:04: Speaker gesturing while explaining the challenge of predicting individual life trajectories using data from the Fragile Families Challenge \(08:04\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-08-04.png)
![Screenshot at 10:25: Slide outlining the comparison between human beings and algorithms in assessing flight risk or crime for bail decisions, pointing to biases like 'Current Offense Bias' and 'Mugshot Bias' \(10:25\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-10-25.png)
![Screenshot at 18:22: Slide comparing LLMs vs. AI-Powered Algorithms, highlighting differences in inputs/outputs, pattern matching, temperature settings, and cognitive biases \(18:22\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-18-22.png)
![Screenshot at 30:21: Slide detailing Keynes' view on uncertainty, noting that for events like war prospects or copper prices, there is 'no scientific basis on which to form any calculable probability' \(30:21\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-30-21.png)
![Screenshot at 35:04: Final slide summarizing that experienced judges can perform worse than algorithms due to cognitive biases, and that algorithms can help identify these human biases \(35:04\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-35-04.png)
![Screenshot at 37:09: A wide shot of the audience during Q&A as a woman stands to ask a question, while the speaker stands at the podium \(37:09\).](https://ss.rapidrecap.app/screens/MvBdRF3zw5k/00-37-09.png)