What Happened
Developer lr0 published an essay titled "I'm not consulting an LLM" that lays out a specific argument against using language models as learning tools. The core claim: LLMs optimize for arrival at an answer, not for the intellectual process of getting there.
The essay uses a thought experiment. Imagine if every Google search returned a single perfect answer instantly. You would get information, but you would lose the experience of encountering diverse perspectives, contradictory sources, bad arguments, and the friction that forces you to build genuine understanding.
The author argues that LLM outputs share this problem, but worse. "You get the answer but nothing else. You don't learn how ideas fight, mutate, or die." The concern is not primarily about accuracy (though the essay notes LLMs only guarantee plausible answers, not correct ones). The concern is about what the author calls "epistemic judgment" - the intuitive sense that something is wrong before you can formally prove it.
The essay also highlights what it calls the hidden danger: "A tool can be efficient and still be intellectually corrosive, because it lies well enough. Its smoothness hides uncertainty."
Why It Matters
This argument is relevant to anyone who has noticed a shift in how they approach problems after relying heavily on AI assistants. If you use ChatGPT or Claude dozens of times per day, you have probably experienced the efficiency gains. But have you also noticed that you are less likely to dig into documentation, read source code, or sit with a problem until you understand it from multiple angles?
The essay puts a name on something many practitioners sense but rarely articulate. The risk is not that AI gives wrong answers. The risk is that correct answers delivered without friction slowly erode the skill of finding answers yourself.
This is particularly relevant for developers and researchers who are still building foundational knowledge in their fields. Senior practitioners who already have strong mental models can use LLMs as accelerators. Juniors who skip the struggle may never build those models in the first place.
Our Take
The argument has merit, but it is incomplete. The essay presents a binary: either you struggle through learning or you get spoon-fed by an LLM. In practice, most people use AI tools somewhere in between. You can ask Claude to explain a concept, then go read the primary sources it references. You can use ChatGPT to get unstuck, then spend time understanding why you were stuck.
The real takeaway is not "stop using LLMs" but "be intentional about when you use them." There is a difference between using an AI assistant to skip a learning opportunity and using one to accelerate past a roadblock you have already wrestled with. The danger the essay identifies is real - smooth, confident answers do mask uncertainty - but the solution is better habits, not abstinence.