Rich Hickey on Why AI is Automation, Not a Tool

The Distinction Between Tools and Automation

Rich Hickey, creator of Clojure, argues that the common classification of Large Language Models (LLMs) as "tools" is a category error that obscures their true nature. He posits that while tools are implements that one wields to build expertise and exert direct control, LLMs function as automation—systems that perform tasks independently, often at the cost of the user's agency and learning.

Defining the 'Tool'

Hickey defines a tool as something that is "wielded." This implies three core characteristics:

• Control and Intention: The user maintains continuous direction over the process. The tool does not act on its own; it transmits the user's energy and intent.
• Sensory Feedback: Using a tool provides immediate, physical or logical feedback that allows the user to adjust their approach, fostering the development of expertise.
• Skill Acquisition: The act of using a tool is the act of learning. By engaging with the physics, logic, or craft of a task, the user gains deep understanding of the domain.

In this framework, a tool is something you use to do something, not something that does the work for you. When you use a tool, you are actively involved in the process, and the outcome is a direct result of your effort and decision-making.

LLMs as Automation

Hickey contrasts this with the current paradigm of LLM usage, which he characterizes as automation. Automation, by definition, operates with little or no direct human control. When a user prompts an LLM, they are not "wielding" the model; they are making a request for the model to perform a task on their behalf.

This shift in modality has significant consequences:

• Abdication of Process: By offloading tasks like coding or writing to an LLM, the user misses the opportunity to learn the underlying logic, formulate arguments, or notice edge cases. The user is no longer the practitioner but a consumer of the model's output.
• Lack of Predictability: Unlike a physical tool or a deterministic software utility, LLMs are probabilistic. Users often find themselves "gaming" the system—using specific prompts or "voodoo" to elicit the desired output—rather than mastering the craft itself.
• Fragility: Because the user does not understand the underlying process, they are vulnerable to changes in the model's behavior. A software update can render a user's learned "prompts" obsolete, leaving them unable to perform the task without the automaton.

The Fast Food Analogy

To illustrate the difference, Hickey uses the example of a fast-food drive-thru. While a drive-thru is useful for obtaining food, it is not a "tool for cooking." Using it does not teach you how to cook; it is a mechanism for avoiding the act of cooking. Similarly, using an LLM to summarize research or write code is a mechanism for avoiding the cognitive work of reading or programming. The convenience of the result masks the loss of the developmental process that occurs when one performs the task manually.

Industry Perspectives

The categorization of AI remains a subject of intense debate within the technical community. Supporters of the "AI as a tool" perspective argue that the definition of a tool is a continuum, and that LLMs simply represent a more advanced, albeit complex, instrument that can be integrated into a professional workflow. Critics of Hickey's view suggest that the focus on "manual effort" is a romanticized view of labor, and that the value of AI lies in its ability to abstract away boilerplate, allowing humans to focus on higher-level domain logic.

However, proponents of Hickey's argument emphasize that the distinction is not merely semantic but political. If AI is viewed as a tool, the user retains accountability. If AI is viewed as an autonomous agent, the responsibility for the output becomes diffused. As one observer noted, the danger lies in the industry's tendency to position these systems as "just tools" to avoid the moral and economic implications of replacing human expertise with automated, non-deterministic systems.