Method of Doubt

The method of doubt is the philosophical procedure introduced by Descartes in the Meditations on First Philosophy (1641) of systematically doubting all beliefs that admit of any doubt, in order to identify those that survive the most radical skeptical challenges and can serve as secure foundations for knowledge. The method is not ordinary doubt but hyperbolical doubt: entertaining even the possibility of a deceiving demon, a dream, or a malicious god who systematically distorts perception and thought. What survives this radical doubt — the cogito (the thinking subject's existence), clear and distinct ideas, and ultimately God's existence and benevolence as guarantors of reliable cognition — forms the foundation on which Descartes attempts to reconstruct knowledge. The method of doubt is methodological rather than genuine: Descartes never believed he was dreaming or deceived by a demon; he used the possibility as a logical test for certainty. It established the framework of modern epistemology — the isolated subject seeking foundations for knowledge against a background of possible deception — that dominated philosophy from Descartes through Kant and beyond, and that still structures contemporary epistemology debates about external world skepticism and the justification of knowledge.== The Method of Doubt in Machine Learning ==

The method of doubt has an unexpected heir in contemporary machine learning: the practice of adversarial training, in which models are deliberately exposed to perturbed inputs designed to fool them. The goal is not to find the average case but the worst case — to identify the inputs that survive the model's confidence and expose its blind spots. This is structural doubt applied to pattern recognition: rather than trusting the model's surface performance, the researcher systematically doubts it.

The analogy extends to red teaming in AI safety, where teams of humans deliberately attempt to provoke failures in AI systems. The red teamer's role is analogous to Descartes's deceiving demon: they construct scenarios that test whether the system's apparent competence is genuine or merely the product of favorable conditions. The difference is that Descartes's doubt was epistemological (can I know anything?), while adversarial doubt is engineering (can I break this system?). But the structural parallel is real: both practices treat the system's default outputs as suspect and subject them to systematic stress.

The deeper connection is to the scalable oversight problem. As AI systems become more capable, the gap between their competence and human evaluators' competence widens. The method of doubt, in its original form, was a procedure for an individual thinker to secure foundations. In its machine learning form, it is a procedure for a community to secure foundations when no individual is competent to evaluate the system alone. The Cartesian cogito — I think, therefore I am — was the endpoint of radical doubt. The machine learning equivalent might be: the system fails under adversarial stress, therefore we know its limits.

Descartes's method was designed to produce certainty. The machine learning version is designed to produce uncertainty — to map the boundaries of what we do not know about a system's behavior. This inversion is characteristic of the shift from classical epistemology to statistical learning: where Descartes sought foundations, machine learning seeks error bounds. The scientific method and the method of doubt together find their twenty-first century expression not in the search for secure axioms but in the search for reliable failure modes.