The Induction Machine: How AI Predictions Trap Us in Recursive History
AI predictions don't foresee the future—they mathematically reinforce past patterns, creating self-fulfilling feedback loops that Hume warned against.
AI predictions are not true forecasting but formalized induction—David Hume's logical fallacy that the future will resemble the past. Machine learning systems like COMPAS mathematically weight historical data, producing a fundamental logical gap when environments shift or data reflects prior observer bias. In predictive policing, this creates a dangerous closed-loop: more policing produces more arrests, which the algorithm uses to justify deploying even more resources to the same neighborhoods.
When we use AI to predict crime, we are often merely repeating the 18th-century logical error that David Hume warned would lead to false certainty. In his Enquiry Concerning Human Understanding, Hume argued that induction—the assumption that the future will resemble the past—is a habit of mind rather than a logical necessity. Modern machine learning is the technical peak of this habit. It does not "see" the future; it mathematically weights the past, creating a fundamental logical gap whenever the environment shifts or the data is tainted by the observer's own hand. The COMPAS recidivism algorithm illustrates the danger of treating this inductive leap as an objective truth. By relying on historical arrest data to calculate risk scores for future sentencing, the system assumes that past policing patterns are a neutral mirror of reality. However, if historical arrests were concentrated in specific ZIP codes due to previous policy decisions, the algorithm "predicts" high risk in those same areas. This creates a closed-loop induction: the model’s output justifies deploying more resources to those neighborhoods, which leads to more arrests, which then feeds back into the model as "proof" of its accuracy. Data-driven policing is not a neutral window into tomorrow; it is a formalization of the logical fallacy that the past is a perfect, static template for the future. While a model might achieve high statistical probability within its training set, it remains trapped by its own inductive bias. In social contexts, this version of AI doesn't just fail to account for change—it actively traps us in a recursive loop of our own history.