The Limits to Forecasting

Can AI predict everything?

In today’s world of abundant data, advanced algorithms, and immense computing power, it's tempting to believe that nearly everything can be predicted. We constantly hear about AI’s remarkable predictive capabilities. Studies suggest that AI can forecast which social media post will go viral, which Hollywood movie will be a blockbuster, or which song will top the charts. Some even claim AI could predict rare events like civil wars. The enthusiasm for prediction is everywhere, even during games, where screens flash probabilities of which team might win after every ball.

But how much of this is truly accurate? Is there a fundamental limit to what we can predict? When should we trust these predictions, and when should we be skeptical?

For builders of AI systems, ability to clearly state what AI can predict and what it cannot, is the cornerstone of responsible AI. Not overhyping its capabilities, and being transparent about its limitations can drive responsible use of these predictions.

Physical Domain vs. Social Domain

When discussing the limits of prediction, it's essential to differentiate between the physical and social domains, as they operate under distinct rules. Physical systems often adhere to deterministic laws, while social systems are influenced by human behavior, which is far more unpredictable. Additionally, the frequency of events plays a crucial role—it's much harder to identify underlying patterns behind rare events.

Rare but Deterministic: Earthquakes Earthquakes are rare, but they are governed by physical laws. Although extremely difficult to predict precisely, their occurrence follows geological patterns that science continues to study. For example, new AI models have found success in predicting when a cyclone will make landfall, helping administrators issue advance warnings so that people can get out of harms way sooner.

Rare and Complex: Black swan events such as Civil War or Epidemics : Civil wars, epidemics and financial crises are rare and arise from highly complex and unique social, political, and economic dynamics. Predicting them involves understanding deep factors, making accuracy very elusive. In simple words, there is not enough "training" data for models to learn.

Frequent and Deterministic: DeepMind’s Predictions on Material Properties, Protein folding In fields like material science, predictions occur frequently because physical properties follow well-defined laws. Machine learning models can make highly accurate predictions based on structured data. This field has gained a lot through the predictive power of AI, we can see recognitions in the Nobel prize for chemistry for 2024.

Frequent but Complex : Viral Posts, Hit Songs, Blockbuster Movies, Stock Market Winners, or Fashion Trends While trends in social media, entertainment, and markets emerge regularly, they’re difficult to predict due to the complexity of human behavior and changing tastes. Success in these domains often feels random or driven by sudden shifts.

Across all these examples, one thing is clear: predictions in social domains are far more challenging than in physical domains. This difficulty arises from several factors:

• Feedback Loops: For instance, in finance, a prediction that a stock might rise could lead to a surge in buying, which artificially inflates the stock, creating a self-fulfilling prophecy. Feedback loops are always present when humans are involved.
• Hidden Data: People's motivations can be deeply buried, sometimes even unknown to them. While AI can excel in rule based games like Chess and even Go, try using it for a game like Poker !
• Sudden Events: A single unexpected event—like a game-changing throw (hail Mary throw!) or an unforeseen business decision—can drastically alter outcomes.
• Timing: Success can depend heavily on timing. What worked for a viral post or hit movie a few years ago may not work today.

Moving Forward: Lessons from the Limits of Prediction

Understanding the boundaries of prediction offers valuable lessons on how we should approach data and technology. Here are a few takeaways:

1. Understanding the Underlying Theory: In many cases, understanding why things happen is as important as predicting what will happen. In medicine, for instance, predictive models may forecast the likelihood of heart disease, but understanding the role of lifestyle factors like diet and exercise allows for better intervention. Similarly, predicting material properties for alloys has become possible due to atomic-level understanding, not just by relying on empirical observations.
2. Ethical Considerations: As predictive models become more widespread, it’s crucial to recognize their limitations and potential biases—especially when they impact people's lives. This is the foundation of responsible AI.
3. Humility vs Hubris in Forecasting: We should be cautious when making bold predictions, especially in complex social systems where human behavior plays a crucial role. Acknowledging the presence of known-unknowns and unknown-unknowns needs to be factored in. Several predictions have gone spectacularly wrong, and with the benefit of hindsight one could see what got missed.

Conclusion: Prediction Has Limits, Especially in Human-Centric Domains

Two years before the great depression, famed economist John Maynard Keynes confidently declared that “ we will not have any more crashes in our time”

Fed had predicted in 2006 “ The worst is behind us”

Several economists had predicted an immediate global recession after Donald Trump’s victory.

The funny thing is that if you get a prediction right, everyone goes crazy and treats you like a divine oracle. When you get it wrong, people simply forget about it.

As we push the boundaries of prediction, let’s also cultivate the wisdom to acknowledge what we cannot foresee. That is how we can help drive responsible use of AI.

Thank you for reading! The views expressed here are my own.

Additional Reading

1. https://www.cs.princeton.edu/~arvindn/teaching/limits-to-prediction-pre-read.pdf
2. Podcast (created by NotebookLM)