How Aleatoric and Epistemic Uncertainty Separate Inherent Randomness from Lack of Knowledge

On October 6, 2025, researchers presenting at the 42nd International Conference on Machine Learning published a paper that broke a foundational assumption in artificial intelligence: the idea that uncertainty neatly divides into two distinct categories. For years, engineers and data scientists have relied on a strict dichotomy to measure risk. They separated "aleatoric" uncertainty—the inherent, irreducible randomness of a coin flip—from "epistemic" uncertainty, the reducible lack of knowledge that disappears when you gather more data. But as large language models began to scale, that clean boundary collapsed. The 2025 findings demonstrated that what the industry marketed as a solved mathematical framework was actually insufficiently expressive to capture how modern neural networks process probability. The distinction between what a model cannot know and what it simply does not know yet had become a fluid threshold.[5]

The origins of this dichotomy lie far outside computer science. In 2008, the journal Structural Safety published a baseline framework defining aleatoric uncertainty as strictly irreducible physical randomness. If an engineer is building a bridge, the exact wind speed on a Tuesday in 10 years is aleatoric—it is a game of chance, derived from the Latin "alea" for dice. No amount of historical data will give you the exact number. Epistemic uncertainty, derived from the Greek "episteme" for knowledge, represents the things the engineer could know but does not. If the bridge's steel has a microscopic flaw that the inspector missed, that is epistemic. The uncertainty exists only because the observer's knowledge is incomplete.[2]

Machine learning adopted this framework wholesale. A 2019 paper in the Machine Learning journal formalized the concepts for neural networks, treating the classic model as a probabilistic engine. In this setting, epistemic uncertainty was interpreted as the uncertainty of the network's parameters. If a model was unsure whether an image showed a dog or a cat because it had never seen that specific breed, feeding it 10,000 more images of that breed would reduce the epistemic uncertainty to near zero. The model simply needed more data.[1]

Aleatoric uncertainty, on the other hand, appeared in the probabilistic prediction itself. If a self-driving car's camera is blinded by a sudden glare of sunlight, the exact position of the lane lines is lost. Even if the model has seen a billion hours of driving footage, it cannot definitively know where the lines are in that specific, corrupted frame. The uncertainty is baked into the data-generating process. For years, the goal of "uncertainty quantification" in AI was to measure the total uncertainty, isolate the aleatoric baseline, and treat the remainder as the epistemic gap that engineers needed to close.[1]

But the reality of deploying massive models exposed the cracks in this theory. In April 2022, researchers published a deeper look into disentangling the two uncertainties, revealing that they often interact in ways the strict dichotomy cannot handle. If a surrogate model learns from computer experiments that contain aleatoric noise, the resulting model exhibits epistemic uncertainty that is fundamentally dependent on that initial randomness. The two types of uncertainty bleed into one another, making it mathematically impossible to cleanly separate them in complex, real-world deployments.[4]

The breaking point arrived with the massive scale of large language models. In February 2024, researchers demonstrated that language models could be used to distinguish the knowable from the unknowable at the token level. When an LLM generates text, it assigns a probability distribution to the next word. If the prompt is "The capital of France is," the model has near 100% certainty that the next word is "Paris." That is a solved epistemic gap. But if the prompt is "The first card drawn from the shuffled deck was," the model faces pure aleatoric uncertainty. The next word could be any of the 52 cards.[3]

However, the 2024 LLM experiments revealed something deeper: the boundary between the two uncertainties is entirely dependent on the observer's computational budget. If a model is small and trained on limited data, a vast swath of the world appears aleatoric to it. It lacks the capacity to recognize the underlying patterns, so it treats the variance as irreducible noise. But when a massive model with a trillion parameters processes the exact same data, it finds the hidden correlations. What was aleatoric to the small model becomes epistemic to the large one, and is subsequently reduced to zero.[3]

This shift fundamentally changes how we evaluate AI safety. Philosophers and statisticians have long debated whether it is practical to distinguish the two types of uncertainty. In classical frequentist statistics, the likelihood function describes the aleatoric uncertainty while the parameters represent the epistemic. But in the age of generative AI, marketing departments often sell "uncertainty quantification" as a guarantee of safety. They claim their models can perfectly isolate inherent randomness from a lack of knowledge, promising that the system knows exactly when to ask a human for help.[6]

The 2025 MLR Press findings dismantle that marketing narrative. Freddie Bickford Smith, lead author of the study, stated the problem plainly: "We identify incoherence in existing discussions of these ideas and suggest this stems from the aleatoric-epistemic view being insufficiently expressive to capture all the distinct quantities that researchers are interested in." The dichotomy is too simplistic to capture the statistical dispersion in modern data. When an AI agent is deployed in the real world—whether it is diagnosing a patient or routing a power grid—its uncertainty is not a fixed property of the universe. It is a spectrum, dictated by the limits of its own architecture.[5]

The separation of inherent randomness from a lack of knowledge is not a mathematical absolute, but an engineering threshold. As models continue to scale, they will push that boundary further, converting what we once thought was pure chance into predictable patterns. The challenge for the next generation of AI is not to perfectly categorize what it does not know, but to build systems robust enough to operate safely in the gray area between the two. The 2025 ICML paper sets the new baseline: any safety framework that relies on a clean split between aleatoric and epistemic risk is measuring a world that no longer exists.[5]