When People Start Using LLMs to 'Think': A Real Snapshot of AI Dependency

A Post That Sparks Reflection

On Hacker News, a short post with only 8 points quietly appeared: "So I guess everyone uses LLMs to think now." No body, no link—just the title itself. That seemingly casual exclamation precisely hit an unspoken reality in today's AI circles: more and more people are no longer "using their brains" themselves, but outsourcing the thinking process to large language models.

The post itself is an interesting sample—it's short enough that it needs no extra explanation or data support, yet it resonates because of the "seen-it-all" tone in the title. It doesn't criticize anyone or draw conclusions; it simply describes an ongoing phenomenon and leaves readers to see themselves in it.

From 'Auxiliary Tool' to 'Cognitive Infrastructure'

Using LLMs as thinking tools is nothing new. From initially polishing text, explaining concepts, and generating code snippets, to now asking the model first when deciding what to eat for dinner, evaluating a relationship, or diagnosing work confusion—the boundary of human-model interaction is quietly shifting.

The key to this shift is the replacement of the "default option." In the past, when faced with a problem, people's first reaction was to search, flip through books, or talk to friends; now, opening a chat box and typing a natural language prompt has become the path of least resistance. When a path is smooth enough, it becomes a habit; when a habit is widespread enough, it becomes cognitive infrastructure. The words "everyone" and "now" in the post title hint at exactly this transition from exception to norm.

The term "cognitive infrastructure" borrows from the perspective of infrastructure studies: a typical feature of infrastructure is that it is almost invisible when functioning normally, and people only become aware of its existence when it fails. Electricity, tap water, and search engines have all undergone this role transition. Calling LLMs cognitive infrastructure means they have moved from "tools that need to be deliberately invoked" to "background conditions that are just there by default, used without asking." The value of this perspective lies in reminding us to pay attention to dependency and single-point-of-failure risks: if a large amount of thinking is done through the same type of model, then the model's biases, hallucinations, or interface outages will affect the entire cognitive activity like a power outage. Another layer of meaning of infrastructure is publicness and governance—who trains it, who prices it, and who decides its default behavior will all indirectly shape the thinking habits of millions.

What Does 'Thinking with LLMs' Actually Mean

Saying "thinking with LLMs" does not mean humans have stopped thinking. A more accurate description is: the "scaffolding" of thinking has changed. People still have to ask good questions, judge whether answers are reliable, and choose among multiple options, but the intermediate labor of diverging, retrieving, and organizing language has been taken over by the model.

There are two very different interpretations here. Optimists believe this is another amplification of cognitive leverage—just as calculators didn't make people lose arithmetic ability but allowed them to focus on more complex problems; search engines didn't make people stop seeking knowledge, they just changed the entry point to knowledge. What the model takes over is the tedious "mental heavy lifting."

Cautious people worry that what is being outsourced is precisely the most valuable struggle in thinking. Writing a stumbling first draft, repeatedly trying and erring between concepts, forcing yourself to reorganize logic after getting stuck—these "inefficient" steps are often where true understanding is formed. If every time you get stuck you immediately hand it to the model, you may retain the answer but lose the path to understanding where the answer comes from.

The "cognitive leverage" mentioned in the text can be understood more concretely: the essence of leverage is to use a small active input to pry a larger output, provided the fulcrum is stable enough. Calculators, search engines, and spell checkers have all been seen as cognitive leverage; their common feature is taking over mechanical steps that are deterministic and verifiable. What makes large language models different is that the boundary of the steps they take over is blurry—they can do mechanical labor like retrieval and wording, but also output judgments, frameworks, and tendentious suggestions. Therefore, whether leverage holds depends on whether the user can stably distinguish "the mechanical part that can be outsourced" from "the judgment part that must be borne by oneself." This provides a common analytical coordinate for the optimistic and cautious interpretations in the text: the real debate is not whether leverage exists, but where the fulcrum is placed.

An Overlooked Cost: The 'Traceability' of Thinking

When a large number of judgments and decisions are based on model outputs, a more hidden problem emerges: can we still trace how we arrived at our conclusions?

Views derived through one's own reasoning usually come with a complete chain of reasoning—even if the chain has flaws, at least you know where the flaws are. Answers obtained from a model are often a packaged finished product; accepting it takes only seconds, but unpacking it takes much longer. Over time, personal judgments become a collection of conclusions of unknown origin, losing the clue of "why I think this way." This is especially worth vigilance for professional decisions that require responsibility—writing code, making diagnoses, giving advice.

Traceability corresponds to the "justification" problem in epistemology: a belief must not only be true, but the subject must also be able to explain why it is true. Conclusions derived through one's own reasoning, even if wrong, carry the reasoning path and can therefore be checked, corrected, and transferred to new situations; while model outputs accepted directly often provide only conclusions and seemingly reasonable explanations, and the explanations themselves may not reflect the model's actual generation process. This constitutes a substantial risk in professional scenarios: code review, clinical diagnosis, and compliance judgments all require decision-makers to be accountable for conclusions, and accountability presupposes reproducible evidence. Therefore, "asking for evidence" is not just a good habit, but an operational means of transforming untraceable finished products back into traceable judgments.

Perhaps the Question Is Not 'Whether to Use' but 'How to Use'

That short post did not take a stand, but the question it raises is valuable enough in itself. Rather than taking sides on "should we use LLMs to think," it's better to focus on more practical dividing lines:

• Use LLMs as sparring partners, not substitutes: First give your own first version of an idea, then use the model to compare, find errors, and fill blind spots—rather than turning a blank input directly into a finished output.
• Preserve time to 'get stuck': When encountering difficulties, set a buffer period during which you don't ask the model for help, letting the struggle itself play a role.
• Ask for sources: For key conclusions given by the model, cultivate the habit of asking "what is the evidence?" and reopen packaged answers.

The slightly teasing "I guess" in the title actually leaves room. It both describes the current situation and seems to remind us: if everyone defaults to thinking with models, then "not doing so" itself may become a scarce ability again.

Frequently Asked Questions

Will thinking with LLMs degrade people's thinking ability? There is no definitive answer yet. The key lies in how it is used: treating the model as a first reaction can easily weaken independent thinking, while treating it as a second opinion may amplify cognitive ability.

Does this Hacker News post provide specific data or cases? No. The post itself only has a title, no body or data. It is an observational expression of a phenomenon, not empirical research.

How can I tell if I am overly dependent on LLMs? A simple signal is: whether your first reaction to a problem is to open a chat box, and whether it is noticeably difficult to organize your thoughts independently without the model.