Do LLMs undermine or improve critical reasoning skills?

Socrates famously argued in Plato’s [Phaedrus](https://www.gutenberg.org/files/1636/1636-h/1636-h.htm) that the invention of writing would produce "forgetfulness in the souls of those who learn it," replacing internal wisdom with external marks. We now face a parallel crisis: does delegating our thoughts to Large Language Models (LLMs) signify the atrophy of human reason or its ultimate evolution? The impact of LLMs on critical reasoning is not a binary outcome but a fundamental shift in **cognitive load**. Critical reasoning traditionally involves two stages: the generation of possibilities and the rigorous evaluation of those possibilities. LLMs intervene in both, acting as either a "crutch" or a "scaffold." ### The Risk of Cognitive Atrophy LLMs undermine reasoning through **automation bias**—the human tendency to favor suggestions from automated systems even when they are demonstrably incorrect. When a user employs an LLM to summarize a complex text, they bypass the "productive struggle" necessary for deep conceptual mapping. Without this struggle, the brain fails to form the dense neural connections required for long-term retention and synthesis. As researchers Emily M. Bender and Timnit Gebru noted in their influential paper [On the Dangers of Stochastic Parrots](https://dl.acm.org/doi/10.1145/3442188.3445922), these models lack an underlying model of the world, meaning they prioritize linguistic plausibility over logical truth. > "Contrary to the surface appearances, an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data... but without any reference to meaning." If we treat these probabilistic outputs as reasoned conclusions, our ability to detect fallacies and verify claims begins to wither from disuse. ### The Scaffolding of Dialectic Conversely, LLMs can improve reasoning by acting as an **adversarial partner**. By prompting a model to "play devil’s advocate" or "identify logical gaps in my argument," a user can overcome their own confirmation bias. In this framework, the LLM provides the "raw material" of diverse perspectives, forcing the human to move from a *producer* of content to a *curator* and *evaluator*. This shifts the definition of "reasoning" from the ability to construct a basic argument to the ability to interrogate a complex one. The "critical" in critical reasoning now requires a higher-order skill: **AI Literacy**, or the ability to navigate the tension between the model’s fluent prose and its potential for hallucination. ### Further Inquiries 1. If LLMs are trained on human biases, can they ever truly serve as "unbiased" adversarial partners, or do they merely reinforce the most common logical errors found in their training data? 2. How does the "outsourcing" of synthesis to AI change the neurological development of the prefrontal cortex in students who have never known a pre-LLM academic environment? 3. Will the persistent "hallucination" tendencies of LLMs ironically lead to a renaissance of **primary source verification** and a more skeptical, rigorous public discourse?