When Machines Reason About Uncertainty, Which Framework Are They Using?
Artificial intelligence systems are increasingly asked to reason under uncertainty: to weigh incomplete evidence, assess risk, and produce judgments where no fact set is complete. Almost without exception, the uncertainty frameworks embedded in these systems are probabilistic. Neural networks output probability distributions. Large language models are trained to predict likely continuations. Risk-scoring systems assign probability estimates to outcomes. Probability has become the default vocabulary in which machine judgment is expressed.
This default is well earned. Probability theory is mathematically mature, computationally tractable, and closely matched to the statistical, data-driven inference that underlies most contemporary machine learning. But defaults are worth examining, particularly as intelligent systems take on a larger role in governance, security, and decision-making. Is probability actually the right framework for every kind of uncertainty these systems encounter, or has it simply become the only tool most of them are built to use?
The Case for an Alternative
Possibility theory, developed from Lotfi Zadeh’s foundational 1978 work on fuzzy sets and substantially extended by Didier Dubois and Henri Prade, was built to formalize a different kind of uncertainty than the one probability theory targets. Where probability handles randomness, variation across repeated trials or populations, possibility theory targets imprecision: the vagueness inherent in natural language, expert judgment, and single, non-repeatable situations.
This distinction is not merely academic when applied to intelligent systems. Consider a system tasked with reasoning over expert-authored rules, threat intelligence reports, or policy documents, sources that are frequently qualitative, hedged, and linguistically imprecise rather than statistically derived. Forcing that kind of input into a probabilistic frame, treating “significant concern” as if it were an empirically calibrated percentage, can manufacture a false sense of precision that the underlying evidence never supported.
Possibility theory’s structure offers a way to represent this honestly. Its dual measures, possibility and necessity, allow a system to distinguish between an outcome that is merely not ruled out and one that is genuinely well supported, a distinction that a single probability number tends to collapse. Researchers working at the boundary of symbolic and neural reasoning have explored possibilistic and hybrid approaches for exactly this reason: some forms of knowledge resist being cleanly reduced to a probability distribution without losing something important in the translation.
Why This Matters for Governance and Trust
As AI systems inform decisions in domains like security triage, regulatory compliance, or intelligence assessment, the uncertainty framework embedded in a system is not a purely technical detail. It shapes what the system can express, and by extension, what a human relying on its output is able to understand about the actual state of the evidence.
A system that reports “72 percent probability” when its actual epistemic state is closer to “not ruled out, but weakly supported” is not lying, but it is translating a more honest, less precise judgment into a falsely precise one. That translation can shape trust, resource allocation, and downstream decisions in ways that are difficult to trace back to their source, precisely because the output looks authoritative.
This does not argue for abandoning probabilistic AI in favor of possibilistic AI. The two frameworks suit different kinds of uncertainty, and meaningful research already explores how possibilistic and probabilistic representations can be related and transformed into one another, rather than treated as rival paradigms. The more useful question for governance and system design may simply be: does this system’s uncertainty framework actually match the kind of uncertainty in its input data, or was that framework chosen by default, inherited from the dominant toolchain rather than selected for fit?
A Question Worth Carrying Forward
As intelligent systems take on more of the work of judgment, understanding the assumptions built into how they represent uncertainty becomes part of understanding the systems themselves, and part of holding them accountable. Possibility theory is a useful reminder that probability, however dominant, is one particular way of formalizing not-knowing, not the only coherent one, and that the choice of framework is itself a design decision with consequences.
References
Zadeh, L. A. (1978). Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems, 1(1), 3-28.
Dubois, D., & Prade, H. (2012). Possibility theory: An approach to computerized processing of uncertainty. Springer Science & Business Media.