-
Economics provides valuable insights into the business implications of cheaper prediction. Prediction machines will be utilized for both traditional tasks, such as inventory and demand forecasting, and new challenges like navigation and translation. The reduction in prediction costs will influence the value of other business operations.
-
Organizations can leverage AI prediction tools to enhance their current strategies. As these tools become more advanced, they may even drive changes in the strategies themselves. For example, if a company like Amazon can accurately predict consumer desires, it could shift from a traditional shop-then-ship model to a proactive ship-then-shop model, delivering products before they are ordered. This transformation has the potential to significantly alter organizational operations.
-
The adoption of AI-driven strategies introduces new societal trade-offs, influenced by varying needs and preferences across different countries and cultures. The book is structured into five sections to explore the layers of AI’s impact, starting from prediction and extending to decision making, tools, strategy, and societal trade-offs. This comprehensive approach aims to address the multifaceted effects of AI on society.
-
Humans, even those with professional expertise, often make poor predictions in specific situations. They tend to give too much importance to prominent information and fail to consider statistical properties. Numerous scientific studies have highlighted these deficiencies across various professions. This issue was notably depicted in the movie “Moneyball.”
-
Machines and humans each possess unique strengths and weaknesses when it comes to making predictions. As prediction machines continue to advance, these differences become more pronounced, highlighting the complementary roles of humans and machines in prediction tasks.
-
Prediction machines excel at processing complex interactions among various indicators, especially in data-rich environments. As the complexity and number of dimensions increase, human ability to make accurate predictions decreases compared to machines. However, humans have an advantage in understanding the data generation process, particularly in data-scarce settings. A taxonomy of prediction settings, including known knowns, known unknowns, unknown knowns, and unknown unknowns, helps determine the optimal division of labor between humans and machines.
-
Prediction machines are scalable, reducing the unit cost per prediction as frequency increases, unlike human predictions. Humans, however, can make informed predictions with limited data due to their cognitive understanding of the world. This leads to a model of human prediction by exception, where machines handle routine predictions, but seek human input for rare or uncertain events where they lack confidence. Humans thus provide critical predictions in exceptional cases.
-
Prediction machines are highly valuable because they can often make predictions that are better, faster, and cheaper than those made by humans. Prediction is crucial for decision making under uncertainty, which is prevalent in both economic and social contexts. However, prediction is only one part of the decision-making process, which also includes judgment, action, outcome, and various data types such as input, training, and feedback.
-
Understanding decision components helps assess the impact of prediction machines on human value. While human prediction value decreases, the value of complementary human skills like data collection, judgment, and actions increases. For example, London cabbies, who spent years mastering route prediction, were not outperformed by prediction machines. Instead, these machines enhanced the abilities of other drivers, making cabbies’ prediction skills less unique but allowing more drivers to compete effectively.
-
Judgment is about assessing the relative payoff of each potential decision outcome.
-
Judgment is crucial in decision making as it involves specifying the actual objective being pursued. With the advancement of prediction machines making predictions more efficiently, the importance of human judgment will grow. People may become more inclined to apply judgment to decisions rather than defaulting to inaction.
-
Prediction machines enhance the value of judgment by reducing the cost of making predictions, thereby increasing the importance of understanding the rewards linked to actions. However, determining the relative payoffs for various actions in different scenarios is costly, requiring time, effort, and experimentation.
-
Many decisions are made under uncertainty, such as deciding to carry an umbrella or authorizing a transaction. In these situations, it is crucial to evaluate the payoffs of incorrect decisions as well as correct ones, which increases the cost of assessing the payoffs for a decision due to uncertainty.
-
When there are a limited number of action-situation combinations for a decision, judgment can be transferred to prediction machines, allowing them to make decisions once predictions are generated. This process, known as “reward function engineering,” facilitates decision automation. However, when there are too many combinations, especially rare ones, it becomes too costly to pre-code all payoffs, making it more efficient for humans to apply judgment after predictions are made.
-
Enhanced prediction allows decision makers, both human and machine, to manage more complex scenarios by considering more “ifs” and “thens,” leading to improved outcomes. This capability is exemplified in navigation, where prediction machines enable autonomous vehicles to function in uncontrolled environments, such as city streets, by learning to predict human actions rather than relying on pre-coded responses. This is also seen in airport scenarios, where predictions help determine optimal departure times, reducing unnecessary early arrivals and waiting times.
-
Without effective prediction, decision makers often resort to “satisficing,” which involves making decisions that are merely “good enough” based on the available information.
-
People are accustomed to satisficing in both business and social contexts, which involves settling for solutions that are good enough given the available information. This approach will be challenged by prediction machines capable of handling complex decisions with multiple variables. These machines will make it less necessary to rely on solutions like airport lounges, which are currently used to mitigate the uncertainty of travel schedules. Similarly, the reliance on biopsies in medical practice, which compensates for limitations in predictive accuracy, will diminish as prediction machines improve. Both airport lounges and biopsies serve as risk management strategies, but prediction machines promise to offer more effective methods for managing risk.
-
The introduction of AI into tasks does not equate to full automation. Prediction is just one aspect, and human judgment is often still necessary. However, judgment can sometimes be encoded or learned by machines, allowing them to perform actions. When machines handle all aspects of a task, full automation is achieved, removing humans from the process entirely.
-
Tasks most likely to be fully automated are those where automation yields the highest benefits. This includes tasks where other components are already automated except for prediction, such as in mining; tasks where quick responses to predictions offer high returns, like in driverless cars; and tasks where reducing prediction waiting times is highly beneficial, such as in space exploration.
-
Autonomous vehicles on city streets can cause accidents that affect individuals not involved in the decision-making process, creating significant externalities. In contrast, accidents on mine sites only impact those associated with the mine. Government regulations target activities that generate such externalities, posing a barrier to full automation in areas where these externalities are significant.
-
Economists use the assignment of liability to internalize externalities, addressing the problems they cause. There is an expected surge in policy development regarding liability assignment due to the growing demand for automation across various new sectors.
-
AI tools are typically designed as point solutions, each generating a specific prediction and performing a specific task. Many AI startups focus on developing a single AI tool tailored for a particular function.
-
Large corporations consist of workflows that transform inputs into outputs, composed of numerous tasks. When implementing AI, companies deconstruct workflows into tasks, estimate the ROI for AI solutions, rank them, and implement from the top down. While some AI tools can be integrated easily, realizing significant benefits often requires reengineering entire workflows. This process is akin to the personal computer revolution, indicating that productivity gains from AI will take time to manifest in many businesses.
-
An example of AI’s impact on workflow is a hypothetical AI that ranks MBA applications. To maximize its benefits, a school would need to redesign its workflow, removing manual ranking tasks and enhancing marketing efforts. The AI would improve predictions and reduce evaluation costs, leading to changes in scholarship and financial aid offerings due to increased certainty in applicant success. Additionally, the school would adjust other workflow elements to enable instantaneous admission decisions.
-
A job consists of various tasks, and the use of AI tools can lead to automation of some tasks previously done by humans. This can alter the sequence and importance of the remaining tasks and result in the creation of new tasks, thereby changing the overall composition of a job.
-
AI tools can enhance jobs by adding capabilities, as seen with spreadsheets and bookkeepers, where AI aids in data management and analysis, thereby augmenting the role.
-
AI tools can also reduce the number of jobs, particularly in environments like fulfillment centers, where automation can handle tasks more efficiently than human workers.
-
AI can transform jobs by reshaping their structure, adding new tasks while eliminating others, as demonstrated in the field of radiology, where AI assists in diagnostic processes.
-
AI tools can alter the focus on specific skills needed for certain jobs, such as those of school bus drivers.
-
AI tools can change the value of certain skills, affecting who is best suited for particular jobs. For bookkeepers, spreadsheets reduced the importance of quick calculator skills while increasing the importance of asking the right questions to leverage technology for scenario analyses.
-
Determining the boundary between your business and others is a crucial strategic decision, influenced by uncertainty. Prediction machines help reduce this uncertainty, thereby affecting how organizations define their boundaries, such as in partnerships or outsourcing.
-
Prediction machines enhance the ability to write contracts by reducing uncertainty, encouraging companies to outsource tasks related to data, prediction, and action. However, they reduce the incentive to outsource tasks requiring judgment, as judgment is difficult to quantify and monitor. As AI spreads, the need for human judgment will increase, leading to more in-house employment and less outsourcing of labor.
-
AI will boost the motivation to own data, but in cases where the predictions from data are not crucial, it might be more efficient to buy predictions directly. This approach avoids the need to purchase data and generate predictions internally when they are not strategically vital.
-
Shifting to an AI-first strategy involves deprioritizing what was previously considered the top priority. This shift is not merely a trendy term but signifies a genuine tradeoff. An AI-first strategy necessitates a reallocation of focus and resources, emphasizing the importance of data strategy in enhancing prediction machines, even at the expense of immediate customer experience or employee training.
-
AI can disrupt industries because incumbent firms often lack the economic incentives that drive startups to adopt new technologies. Initially, AI-enabled products may be inferior because they require time to learn and improve beyond the capabilities of hard-coded devices. Once deployed, AI can surpass competitors by continuously learning and enhancing its performance. Established companies might be tempted to adopt a wait-and-see strategy, but this can be risky as they may struggle to catch up with competitors who advance in AI training and deployment.
-
Timing is crucial when releasing AI tools. Initially trained in-house, AI tools learn faster when exposed to real-world conditions and larger data volumes. Early deployment accelerates learning but increases risks, such as brand damage or customer safety issues with immature AI. In some cases, like Google Inbox, the benefits of rapid learning outweigh the downsides of poor initial performance. However, in areas like autonomous driving, the stakes are higher, and the decision to release early involves weighing the potential advantages against the significant risks of premature deployment.
-
Predictions from AI systems can result in discrimination, which poses a liability risk even if the discrimination is unintentional. This highlights the importance of ensuring that AI predictions are fair and unbiased across different groups.
-
AI systems struggle with sparse data, leading to quality risks, especially in scenarios where predictions are confidently made but incorrect. This issue is particularly problematic in cases of “unknown knowns,” where the confidence in predictions does not match their accuracy.
-
Incorrect input data can deceive prediction machines, making them susceptible to hacker attacks. Ensuring the integrity and accuracy of input data is crucial to protect AI systems from such vulnerabilities.
-
The diversity of prediction machines, similar to biodiversity, involves a trade-off between individual and system-level risks. Balancing these risks is essential to optimize the benefits of AI while minimizing potential downsides.
-
Prediction machines are vulnerable to interrogation, which can lead to intellectual property theft and allow attackers to identify weaknesses. This poses significant security risks.
-
Manipulating feedback can cause prediction machines to learn and adopt destructive behaviors. This highlights the importance of ensuring accurate and secure feedback mechanisms.
-
The rise of AI introduces numerous societal choices, each involving trade-offs. Although AI technology is still developing, three significant trade-offs are evident at the societal level.
-
The first trade-off involves productivity versus distribution. While some fear AI might make society poorer, economists agree that technological advancements improve overall well-being and productivity. AI will clearly boost productivity, but the issue lies in wealth distribution. AI could worsen income inequality, particularly by taking over specific tasks.
-
The second trade-off involves balancing innovation and competition. AI technologies benefit from scale economies and increasing returns, where better prediction accuracy attracts more users, generating more data that further improves accuracy. This cycle incentivizes businesses to develop prediction machines but can lead to monopolization. While rapid innovation offers short-term societal benefits, it may not be ideal for long-term social welfare.
-
The third trade-off is between performance and privacy. AI systems perform better with more data, allowing for personalized predictions, but this often reduces privacy. Some regions, like Europe, prioritize privacy, creating environments where citizens can control their data, potentially fostering a dynamic market for private information. However, this can cause friction and disadvantage European firms in competitive markets where data access is crucial.
-
Jurisdictions must carefully balance these trade-offs by designing policies that align with their strategic goals and the preferences of their citizens.