Does AI Understand? The Evidence Is Hard to Ignore

The debate over AI’s capacity for understanding hinges on a single, ambiguous word. On October 15th, Gregory Kurtzer, co-founder of CentOS and Rocky Linux, stated on Twitter that LLMs are “statistical model[s] that spew out tokens…not an understanding of the subject matter.” David Shapiro countered that LLMs’ demonstrable capabilities do constitute understanding.

Kurtzer's impressive credentials, which include years developing high-performance computing software at Lawrence Berkeley National Laboratory, underscore the substantive nature of this dispute. The core ambiguity lies in "understanding" itself: it can refer to a private, conscious experience or to a demonstrable ability to interpret, explain, and use information. These definitions answer fundamentally different questions.

This article adopts a testable approach, evaluating functional competence through observable behavior. We will examine what AI models can do, rather than speculating on their internal states. Consciousness and subjective experience remain separate, unresolved questions, distinct from the objective assessment of understanding.

How do we objectively measure “understanding”? Educators have refined this question for decades. Bloom’s revised Taxonomy offers a six-rung ladder, defining understanding through observable actions: interpreting, summarizing, inferring, comparing, and explaining—not a vague internal state. Its rungs ascend from remember, understand, and apply, to analyze, evaluate, and create.

AI models now demonstrate capabilities across these rungs. For “remember,” frontier models tackling Humanity’s Last Exam (a 2,500-question benchmark) have improved from under 10% to approximately 61% in independent testing. For “understand,” Harvard physics students tutored by AI learned more than twice as much as those in active-learning classes, while even Gregory Kurtzer concedes LLMs can “explain difficult concepts.”

Moving up the ladder, for “apply,” a Google DeepMind AI system achieved a gold medal score of 35 out of 42 on the International Mathematical Olympiad, with its solutions graded by Olympiad coordinators. For “analyze,” AI systems in DARPA’s AI Cyber Challenge found 54 of 63 planted software vulnerabilities, plus 18 unplanted ones.

Higher rungs also show progress. For “evaluate,” GPT-4’s grading aligned with expert human evaluators over 80% of the time. For “create,” AI research systems have generated novel mathematical algorithms, demonstrating capabilities that meet Bloom’s highest-level definitions. These benchmarks demonstrate specific abilities, not universal or human-equivalent understanding.

Describing large language models (LLMs) as merely “predicting tokens” explains how they operate but doesn't define their capabilities. A mechanism doesn't inherently limit function; the brain, after all, is just firing neurons, yet it achieves profound understanding. This distinction is crucial for evaluating AI’s true capacities.

Consider Ashby’s Law of Requisite Variety, which states that an effective controller must possess at least as much variety as the system it regulates. The Conant–Ashby theorem further posits that "Every good regulator of a system must be a model of that system." This means successful regulation implies the development of a sufficiently similar internal model, whether explicit or emergent. AI agents that adapt to novel situations and goals often demonstrate this internal modeling.

Evidence for such learned models is compelling. An Othello model, trained solely on game moves, developed an internal representation of the board; manipulating this internal picture directly altered the model's gameplay. Similarly, an AI model successfully folded laundry with unfamiliar industrial robot arms, performing as well as expert human operators using the same equipment for the first time. These examples support the existence of functional modeling, indicating that AI builds internal representations of its environment, even if it doesn't "think" like a person. For a deeper dive into educational frameworks, explore Bloom’s Taxonomy - Vanderbilt University Center for Teaching.

Inferring inner experience from outward behavior presents a classic philosophical challenge, exemplified by the Chinese Room thought experiment. Functional tests, however, establish useful competence without resolving the metaphysical question of consciousness. We can assess what a system does without knowing what it feels.

Subjective confidence also proves an unreliable benchmark. People frequently overestimate their grasp of familiar mechanisms, leading to errors. Similarly, AI systems can produce confidently incorrect answers, demonstrating that internal certainty does not equate to objective accuracy.

AI understanding, while impressive, remains uneven and fragile in certain contexts. This functional capability is not evidence of consciousness. Regardless of AI’s evolving capabilities, humans and organizations retain full responsibility for the systems they deploy, the decisions they inform, and any resultant harm.

Does AI actually understand anything?

It depends on the definition. By observable measures such as explaining, applying, and analyzing, AI can demonstrate functional understanding, though its performance is uneven.

Does functional understanding mean AI is conscious?

No. Demonstrating useful capabilities does not establish that a system has subjective experience or consciousness.

How can Bloom’s Taxonomy help evaluate AI?

It offers observable cognitive tasks—from remembering and explaining to evaluating and creating—that can be used to assess performance.

Why does Ashby’s Law matter to the AI debate?

It suggests that an effective regulator needs a model of the system it controls, connecting successful adaptation to predictive modeling.