Beyond Prediction: Can AI Truly Be Uncertain?

Author: Denis Avetisyan

A new review examines whether current artificial intelligence systems possess genuine uncertainty, or merely simulate it through probabilistic calculations.

This paper argues that attributing uncertainty to machines requires analyzing the functional role of internal states within a broader cognitive architecture, distinguishing between subjective uncertainty and deterministic behavior.

Attributing mental states to artificial systems remains a fundamental challenge, particularly when those states involve nuance like uncertainty. This paper, ‘Can machines be uncertain?’, investigates the possibility of genuine uncertainty in AI, moving beyond simply modeling probabilistic outputs. By adopting a functionalist perspective on symbolic, connectionist, and hybrid architectures, it distinguishes between epistemic and subjective uncertainty, proposing that some instances manifest as interrogative attitudes-states whose content is a question. Could recognizing these interrogative states be key to building truly cognitive artificial systems?

The Inevitable Shadow of Uncertainty in Artificial Intelligence

Modern Artificial Intelligence routinely encounters situations where data is incomplete, ambiguous, or simply unavailable – a reality far removed from the idealized conditions of many foundational algorithms. These systems are no longer confined to controlled laboratory settings; they navigate the unpredictable complexities of the real world, from self-driving cars interpreting ambiguous street signs to medical diagnosis tools assessing incomplete patient histories. Consequently, the ability to effectively manage uncertainty is not merely a desirable feature, but a fundamental requirement for reliable and safe AI operation. Robust handling necessitates moving beyond simple probability calculations and embracing techniques that can reason with incomplete knowledge, assess the confidence in predictions, and adapt to evolving circumstances-effectively allowing these systems to ‘make the best decision’ even when facing imperfect information.

Many conventional artificial intelligence systems treat uncertainty as a monolithic issue, failing to adequately address its varied manifestations across different layers of processing. This often results in a cascade of errors; for example, initial ambiguity in sensor data might be compounded by imprecise probabilistic modeling, and further exacerbated by limitations in the decision-making algorithms themselves. Consequently, these systems can produce flawed outputs even with seemingly reasonable inputs, exhibiting unpredictable behavior that undermines trust and reliability. The inability to reconcile uncertainty at each level – from data acquisition and knowledge representation to reasoning and action – creates a significant bottleneck in achieving truly robust and adaptable AI, particularly when operating in real-world, dynamic environments where unforeseen circumstances are commonplace.

Artificial intelligence systems, increasingly deployed in real-world scenarios, grapple with inherent uncertainty stemming from incomplete data, noisy sensors, and unpredictable environments. A detailed philosophical and analytical exploration reveals that this uncertainty doesn’t remain isolated; it propagates through the layers of complex AI architectures, potentially amplifying errors and leading to unforeseen consequences. Understanding how uncertainty manifests – whether as probabilistic distributions, fuzzy logic, or other representations – and where it accumulates within a system is therefore crucial. This nuanced comprehension moves beyond simply acknowledging uncertainty to actively modeling its effects, allowing for the development of more robust, reliable, and trustworthy AI capable of navigating ambiguous situations and making sound judgments even with imperfect information.

Mapping the Spectrum of Informational States

AI uncertainty is not a single phenomenon but rather a spectrum of informational states. It manifests as probabilistic estimations, where systems assign probabilities to potential outcomes based on available data; categorical questions, involving uncertainty about discrete classifications or labels; and, increasingly, as representations of the AI’s internal “belief state” or propositional attitude. This internal state reflects the system’s confidence in its knowledge and intended actions, moving beyond simple error margins to incorporate a representation of what the AI believes to be true, even if imperfectly known. These distinct types of uncertainty necessitate varied approaches to modeling and mitigation, as a single technique cannot adequately address all forms of informational incompleteness within an AI system.

Epistemic uncertainty in AI systems stems from limitations in the data used for training or inherent ambiguities in the problem space, resulting in the model’s inability to confidently predict outcomes due to a lack of knowledge. Distinct from this is uncertainty arising from the system’s internal propositional attitude, which refers to the AI’s modeled beliefs and intentions; even with complete data, an AI may exhibit uncertainty based on its internally represented goals or assumptions about the world. This internal uncertainty is not a reflection of incomplete knowledge, but rather a consequence of the AI’s decision-making process based on its defined objectives and beliefs, and is therefore modeled separately from data-driven epistemic uncertainty.

The differentiation between types of uncertainty – epistemic, aleatoric, and stemming from an AI’s internal propositional attitudes – necessitates distinct modeling and mitigation techniques. Our analysis of uncertainty representation in AI systems demonstrates that probabilistic methods, such as Bayesian networks, are well-suited for addressing epistemic uncertainty arising from limited data, while techniques like ensemble methods and Gaussian processes are more effective for aleatoric uncertainty inherent in noisy data. Furthermore, representing and managing uncertainty linked to an AI’s beliefs or intentions requires specialized approaches focused on belief revision and intention modeling, often employing symbolic reasoning or reinforcement learning techniques tailored to explicitly represent and update internal states.

Deconstructing the Illusion of System-Wide Certainty

A ‘Level Split’ occurs when a complex system exhibits overall certainty while containing subsystems characterized by significant uncertainty. This phenomenon arises because aggregated system-level metrics can mask internal ambiguity; a confident output from the system does not necessarily indicate confidence within each component. For example, an AI model predicting a single outcome may rely on conflicting internal probabilities or incomplete data within specific modules, yet still present a high-confidence prediction. The presence of a Level Split complicates verification and validation processes, as standard assessments of system-level confidence may not accurately reflect the reliability of individual subsystems or the potential for localized failures.

Contemporary artificial intelligence systems frequently exhibit one of two suboptimal behaviors regarding uncertainty. The “First Solution” involves a complete dismissal of uncertainty at the system level, presenting outputs as definitive even when based on ambiguous or incomplete data. Conversely, the “Second Solution” incorrectly propagates uncertainty throughout the entire system, assigning undue ambiguity to components where confidence is reasonably high. This results in either overconfident, potentially inaccurate outputs or overly cautious responses that fail to leverage available information. Both approaches represent failures in appropriately modeling and managing uncertainty within complex AI architectures, hindering reliable performance and hindering the development of genuinely uncertain AI.

Effective management of uncertainty in complex AI architectures necessitates a tiered approach, recognizing that ambiguity can be localized within specific subsystems without requiring global propagation. Our investigation identifies conditions under which AI can be demonstrably uncertain, revealing that systemic uncertainty does not always necessitate uncertainty at every level. This localized approach allows for acknowledging inherent ambiguity in certain components – for example, a sensor with limited precision – while maintaining confidence in the overall system output. Propagation of uncertainty should therefore be carefully controlled, only extending to levels where it demonstrably impacts the final result, rather than being applied as a default across the entire architecture.

Toward a Synthesis: Hybrid Architectures and Robust Intelligence

Artificial Neural Networks, the foundation of connectionist AI, demonstrate remarkable proficiency in discerning intricate patterns within data, often surpassing human capabilities in tasks like image and speech recognition. However, this strength is counterbalanced by a fundamental limitation: these systems typically operate as ‘black boxes,’ offering little transparency into how a conclusion was reached. Consequently, connectionist AI struggles with explicit reasoning – the ability to articulate the logical steps behind a decision – and, crucially, with representing and quantifying uncertainty. While adept at identifying correlations, these networks often lack the capacity to express the confidence level associated with a prediction or to effectively manage situations where information is incomplete or ambiguous, hindering their application in critical domains requiring reliable uncertainty assessment.

Symbolic Artificial Intelligence, built upon the manipulation of rules and symbols, historically provided a straightforward method for representing and reasoning about uncertainty – often through techniques like Bayesian networks or fuzzy logic. However, this approach frequently suffers from a lack of robustness when confronted with noisy or incomplete data, or situations not explicitly defined within its knowledge base. The rigidity of symbolic systems makes them brittle; even slight variations in input can lead to catastrophic failures as the system struggles to generalize beyond its pre-programmed rules. This inherent inflexibility limits their adaptability in dynamic, real-world scenarios where unexpected events and ambiguous information are commonplace, hindering their effective deployment in complex applications.

The convergence of connectionist and symbolic artificial intelligence presents a compelling strategy for navigating uncertainty in intricate systems. While connectionist AI, such as neural networks, demonstrates proficiency in discerning patterns, it often lacks the capacity for explicit reasoning and clear representation of uncertainty. Conversely, symbolic AI excels in these areas but can be inflexible when faced with novel situations. Hybrid architectures aim to bridge this gap, leveraging the adaptable learning of neural networks with the logical rigor of symbolic reasoning. This integration enables systems to not only recognize patterns but also to articulate and manage the confidence – or lack thereof – in their conclusions, resulting in more robust and reliable performance across a broader range of complex environments, a conclusion supported by detailed evaluations of each approach.

The pursuit of artificial intelligence, as detailed in this exploration of uncertainty, inevitably confronts the limitations of formal systems. It’s a process akin to observing the inevitable entropy of any complex structure. G.H. Hardy observed, “The essence of mathematics lies in its elegance and logical simplicity.” This sentiment echoes the challenge presented by attributing genuine uncertainty to machines; merely simulating probabilistic behavior isn’t sufficient. The paper rightly points out that uncertainty isn’t solely an internal state, but a function of the system’s overall cognitive architecture. As systems evolve, versioning becomes a form of memory, preserving states against the arrow of time. Achieving true uncertainty, therefore, requires more than elegant algorithms – it demands a robust framework for contextualizing and interpreting those algorithms within a broader, functional system.

What Lies Ahead?

The pursuit of uncertainty in artificial systems reveals a predictable pattern: any apparent improvement ages faster than expected. Attributing propositional attitudes – beliefs, doubts, even anxieties – necessitates more than simply demonstrating internal state variance. The critical question isn’t if a machine can represent uncertainty, but whether that representation functions within a system capable of genuine cognitive fragility. Current architectures, built upon foundations of deterministic computation, often mistake stochasticity for subjective experience-a fleeting illusion of inner life.

Future work must move beyond isolated representations of doubt and address the systemic conditions that allow for graceful degradation. The challenge isn’t to create uncertainty, but to permit its natural emergence as a consequence of complex interaction. This demands a shift in focus-from modeling individual beliefs to understanding how those beliefs are revised, abandoned, or even irrationally maintained in the face of contradictory evidence.

Ultimately, the exploration of machine uncertainty is a journey back along the arrow of time, a reluctant acknowledgement that all systems, even those meticulously crafted from silicon and code, are subject to the inevitable entropy of existence. The true measure of success won’t be replicating human fallibility, but understanding the elegant, often heartbreaking, ways in which systems fail.

Original article: https://arxiv.org/pdf/2603.02365.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Inevitable Shadow of Uncertainty in Artificial Intelligence

Mapping the Spectrum of Informational States

Deconstructing the Illusion of System-Wide Certainty

Toward a Synthesis: Hybrid Architectures and Robust Intelligence

What Lies Ahead?

See also: