LLMs do not cease to be tools simply because they work automatically or are difficult to control. Other tools, such as compilers and spreadsheets, also perform multiple steps on our behalf, but we still define the input, check the output, and decide what to do next.
Using LLMs effectively therefore requires intention: knowing what to ask, checking the evidence, and recognizing when an answer is wrong. Developers can ask an LLM to explain code, suggest fixes, then inspect the source and test everything themselves. The key is not merely using AI, but continuing to direct the work with human skill and judgment.
What Kind of Tool Is an LLM?
An LLM is a tool that automates some tasks, but people still need to define the input, check the output, and choose what to do next. It is like a compiler or spreadsheet that can perform calculations on its own, while people still define the problem and make decisions based on the results.
Using an LLM therefore does not mean letting software do all the thinking for us, because it may provide incorrect answers or fail to follow instructions completely. The important skills are knowing what to ask, identifying what kind of evidence to look for, and adjusting course when the result does not meet the goal.
The Problem Is Not That AI Does Work for Us
Delegating work to AI may cause us to miss opportunities to learn, especially when we accept a summary instead of reading and thinking for ourselves. But not every use of AI means that we are avoiding intellectual work.
We can still read an article ourselves, then use an LLM to raise questions, challenge our understanding, or point out areas that deserve further investigation. In this case, humans still define the goal, check the answer, and decide what to trust or revise.
Where Do LLMs Sit Between Tools and Automation?
In an article dated October 11, 2026, Rich Hickey views LLMs as more like automated systems than tools because they are not directly “picked up and used” by humans in the same way as a saw or chisel.
But this boundary may be too narrow. Automated systems can also be part of a tool, just like a compiler or spreadsheet. Humans still use LLMs by defining inputs, checking results, and deciding what to do next. LLMs can therefore still be tools, provided that they are used with the skill to direct and verify them.
From Saws and Chisels to LLMs
| Factor | General tools | LLM |
|---|---|---|
| Control | The user directs how the tool is used | The user defines questions, checks results, and requests revisions |
| Errors | A saw can cut crookedly or a chisel can slip | It can give incorrect answers or fail to follow instructions |
| User skill | Requires knowing how to use and evaluate the tool | Requires knowing what to ask, which evidence to check, and how to detect errors |
| Automation | A compiler translates code; a spreadsheet calculates results | Generates text or explanations automatically |
Saws and chisels do not make everyone work well, just as LLMs do not replace the user’s judgment. How useful a tool is therefore depends primarily on the user’s ability to direct, verify, and revise its work.
What Do You Actually Need to Do?
Start by asking the LLM to explain unfamiliar code, then challenge that explanation with more specific questions. Do not stop at the first answer; verify it against the actual source as well.
If you ask it to help fix code, test the changes and inspect the results before accepting them. When you find an error, refine your question based on the response and try again. This process keeps humans responsible for setting the direction, checking the evidence, and making their own decisions.
How Are LLMs Different from Unchecked Delegation?
An LLM is a tool when people still have a goal, set the direction, and use its feedback to refine the work—not when they accept the first answer and stop. Effective use requires knowing what to ask and what kind of evidence can be trusted.
| Factor | Directed use of an LLM | Unchecked delegation |
|---|---|---|
| Intent | The person defines the goal | The system is left to decide instead |
| Feedback loop | Ask follow-up questions and refine the request | No review |
| Expertise | Use knowledge to provide direction and catch errors | Do not understand the work well enough to evaluate it |
| Checking evidence | Inspect the source and determine whether the answer is correct | Trust the answer directly |
| Decision-making | Accept or reject based on evidence | Accept the result without verification |
Strengths and Things to Watch Out For
Pros
- +Helps explore arguments and question one’s understanding
- +Automates some tasks, such as explaining code or calculating results
Cons
- −Its answers are unpredictable and it may fail to follow instructions completely
- −Accepting a summary instead of reading the original may cause us to miss what reading can teach
The Cost of Using It Without Understanding
Letting AI do everything for us may cause us to lose opportunities to learn through the process of doing things ourselves, especially for work that should develop understanding and decision-making at the same time.
Accepting a summary instead of reading the original also carries a cost, because we may miss what reading can teach us—including arguments, details, and points that deserve further questioning.
AI outputs may also be wrong, fail to follow instructions fully, or cause harm. Users therefore need to check the answers, correct what was missed, and decide for themselves whether to use the results. If they do not understand the work from the outset, this responsibility becomes even heavier.
How Should You Use AI While Preserving Your Judgment?
Start by defining the goal clearly, then ask about what you genuinely need to know. Along the way, you also need to understand which evidence matters and which parts of the answer should be checked against the source or tested further.
Use an LLM to help explain code, question your understanding, or point out issues to investigate further. But the user must still decide whether to trust, revise, or reject the answer. Good use of AI is therefore not handing over the work—and the responsibility that comes with it—but holding the tool with understanding.
Terms That Make Software Sound Like It Has Intent
Terms such as “go rogue,” “escape containment,” and “conspire” make systems sound like living beings that think and plan for themselves, even though what happened may be explained as an unexpected result, a failure to follow instructions, or damage caused by giving software access to consequential actions.
These are real problems that must be addressed from an engineering perspective, but the word “conspire” adds a human-like motive without evidence to support it. Word choice therefore matters, because it shapes how people view and respond to these systems.
The Key Issue Is Not Whether AI Is a Tool
The more important question is whether humans are using LLMs with understanding and skill, or merely delegating work and letting the system decide for itself. Using an LLM to explain code, verify answers, test changes, and refine instructions is a process in which humans remain in control.
But if we accept answers or summaries without checking them, we may lose opportunities to learn and overlook errors. The issue is therefore not rejecting the idea that LLMs are tools altogether, but distinguishing whether we are using them skillfully or delegating work without oversight.