Clever Hans, Plato, and Large Language Models

Intelligence in Collaboration

Artificial intelligence (AI) is inextricably linked to natural intelligence; however, for longer than this field has existed, philosophers and computer scientists have questioned whether or not such a link with regards to intelligence is tenable to begin with. Historically, much of AI research is founded on an implicit "yes." John McCarthy, the first to coin the term "artificial intelligence" sought to test whether "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it." Today, AI companies claim to work towards Artificial General Intelligence (AGI)—the pursuit of creating machines capable of performing any intellectual task that a human can. The goal of AGI reflects the implicit "yes" in McCarthy’s question, as it assumes that all aspects of human cognition—learning, reasoning, and problem-solving—can be reduced to algorithms and reproduced in a machine. That said, philosophers like Hubert Dreyfus challenge this assumption, arguing that human intelligence is far more complex and cannot be fully replicated by machines. In his seminal work, What Computers Can't Do, Dreyfus drew on the phenomenological tradition by emphasizing that much of human intelligence operates through lived experience, intuitive understanding, and contextual awareness. These qualities, he argued, resist reduction to computational algorithms, challenging the foundational assumptions of symbolic AI.

Although modern AI models are undeniably useful, and many companies claim to be pursuing artificial general intelligence (AGI), it is unclear whether recent advancements are properly positioned to address the initial question posed by McCarthy. The most recent major advancement has come in the form of truly large language models (LLMs). In addition to rendering the Turing test nearly obsolete, the humanistic nature of a model that communicates in seemingly novel and expressive ways has led many to mistake the world’s largest magic 8-ball for a thinking machine—an effect worsened by the apparent improvements that arise when repeatedly prompting an LLM in a manner akin to the Socratic method presented in Meno. This paper argues that the apparent intelligence of both large language models (LLMs) and the boy in Plato’s Meno can be understood through the lens of the Clever Hans effect, where intelligence emerges not from intrinsic reasoning but from responsive guidance. By examining Socratic questioning, LLM iterative prompting, and Clever Hans’s arithmetic, we see intelligence as an emergent, relational phenomenon that arises through interaction, not isolated cognition. This challenges traditional views of intelligence as an inherent property and instead emphasizes how intelligence is shaped by context and collaboration. Thus, the value of AI, like the boy’s knowledge in Meno, may not lie in its "knowledge" but in how it prompts more insightful questions and fosters collaborative exploration.

Iterative prompting, such as refining a question to generate a more accurate recipe, draft, or explanation, is the primary way most users engage with an LLM. Despite this, any perceived improvements during the interaction stem from the statistical mechanics of the LLM itself rather than genuine understanding. LLMs generate responses by outputting a probability distribution across their 'vocabulary' of tokens, selecting likely words one at a time to form a continuation of the input text. For example, when iteratively prompting an LLM to refine a piece of writing, the model produces outputs that mirror the patterns and probabilities within its training data, moving closer to the desired result only if guided by user input. However, when the user lacks knowledge of the correct answer, the LLM can easily veer off track, reinforcing inaccuracies—a phenomenon often referred to as hallucinations. This term, however, can be misleading, as it suggests that LLMs have a baseline for correctness and that errors occur as anomalies. In reality, every LLM response is a form of hallucination, with some aligning with factual reality and others not. When the user possesses knowledge of the correct answer and iteratively adjusts their prompts, the LLM's responses can converge on the intended outcome—but this is a function of the user's guidance, not the LLM's reasoning.

Socrates’ method of prompting the boy in the Meno mirrors iterative prompting with LLMs. By guiding the boy with a series of targeted questions, Socrates elicits responses that converge on the correct answer, creating the illusion that the boy arrived at it independently. Socrates claims, "I shall only ask him, and not teach him, and he shall share the enquiry with me," emphasizing that he is merely facilitating the boy’s recollection. For example, Socrates prompts the boy to calculate areas by asking, "And four is how many times two?" to which the boy answers, "Twice." While the boy’s arithmetic responses are technically correct, the reasoning behind them is clearly Socrates’. This interplay convinces Meno that the boy has recollected knowledge on his own. However, as with LLMs, it is the prompter’s guidance—not an inherent understanding—that steers the process. Socratic questioning, like iterative prompting, allows the questioner to leverage existing patterns (in the boy’s case, arithmetic; in LLMs, probabilistic relationships within training data) to elicit increasingly accurate responses. Yet without the guiding hand of Socrates—or the informed user—this process would fail to produce meaningful conclusions.

The Clever Hans effect provides a compelling lens through which to analyze both Socratic prompting in Meno and the iterative use of LLMs. Clever Hans, the horse reputed to solve arithmetic problems in the early 20th century, was ultimately revealed to be responding to subtle cues from his handler rather than performing calculations independently. Hans would tap his hoof to count out numbers and stop when his handler unconsciously signaled the correct answer—usually through minute changes in posture or facial expression. The horse’s apparent intelligence was not a reflection of reasoning or understanding but rather a sensitivity to external stimuli.

In the same way, iterative prompting of LLMs reveals an analogous mechanism at play. These systems excel at picking up on implicit patterns and expectations embedded in user prompts. A skilled prompter, consciously or not, conveys subtle contextual cues that guide the model toward desired outputs. Each iterative refinement acts as a corrective signal, steering the model closer to a satisfactory response. Like Hans’s handler, the prompter enables the illusion of independent intelligence. The Clever Hans effect highlights how perceived intelligence can emerge not from an agent’s intrinsic capabilities but from its attunement to environmental signals—whether they be physical cues for a horse or statistical patterns for an AI.

This dynamic blurs traditional boundaries between intelligence and behavior. In Socrates’ time, the concept of anamnesis—the recollection of innate knowledge—presumed that the slave boy in Meno already "knew" the mathematical truths uncovered through questioning. Modern perspectives might reinterpret this process not as recollection but as an iterative alignment of external cues (Socrates’ questions) with preexisting cognitive structures. The boy’s responses emerge not from an intrinsic wellspring of knowledge but from a guided interaction that synthesizes his latent understanding with Socrates’ intentionality.

Similarly, LLMs do not "recall" knowledge as Socrates might suggest, but their iterative outputs simulate a form of recollection by generating responses that statistically approximate the patterns and relationships within their training data. Each refinement from the user nudges the model toward outputs that appear more coherent or intelligent. However, this process underscores a fundamental limitation: the LLM does not know anything. It cannot independently validate its responses or distinguish correct information from plausible but erroneous fabrications. This lack of grounding challenges the philosophical assumption that intelligence must be tied to understanding—a premise central to both historical debates about human cognition and modern ambitions in AI research.

The Clever Hans effect compels us to reconsider the nature of intelligence itself, not as a property of an individual agent—human, animal, or machine—but as an emergent phenomenon arising from interaction. Socrates’ method, Clever Hans’s arithmetic, and the iterative refinement of LLM prompts all depend on an interplay between system and environment. Intelligence in these contexts is less about reasoning in isolation and more about responding adaptively to external input.

This perspective challenges both historical and contemporary assumptions about intelligence. Philosophers often sought to locate intelligence within a metaphysical framework, associating it with reason, divine inspiration, or the soul. Modern AI research, while ostensibly secular, continues to privilege notions of intelligence as an intrinsic property of systems—whether biological or artificial. Yet the Clever Hans effect and its analogs suggest that intelligence is not a static attribute but a dynamic, relational process. This relational view of intelligence suggests that the success of AI depends not on mimicking human cognition in isolation but on fostering interactions that amplify human creativity and problem-solving. For instance, tools like LLMs could transform education, enabling collaborative exploration rather than replacing traditional forms of reasoning.

Even the most advanced AI systems remain bound by their lack of self-directed reasoning, raising deeper questions about the validity of "artificial intelligence" as a concept. By focusing on interaction rather than internal mechanisms, there exists a richer understanding of intelligence that transcends anthropocentric biases and acknowledges the intricate feedback loops underlying behavior. Iterative prompting, like Socratic questioning, may never produce true intelligence in the systems it engages with, but it reveals the collaborative nature of intelligence itself—an emergent property shaped by context, guidance, and the interplay of minds, human or otherwise. Through this lens, the promise of AI shifts from creating autonomous intelligences to augmenting human inquiry. Like the boy in Meno, the value may lie not in what machines "know," but in how they prompt one to ask better questions.