The Hollowed Mind – ChatGPT Version
The extended mind thesis, as proposed by Clark and Chalmers, is one of those philosophical ideas that feels obviously true once you see it. The claim, roughly, is that cognition does not stop at the skull. Tools, notebooks, maps, calculators, phones, calendars, and other bits of the world can become part of the thinking system. A notebook is not just a passive store of information. In the right circumstances, it functions as memory. The mind leaks out into the world.
This was a useful corrective to a cramped view of intelligence. Human beings have always thought with tools. We think with language, diagrams, books, spreadsheets, whiteboards, institutions, habits, and other people. The romantic image of the pure individual mind, sealed off from the world and doing everything internally, is largely nonsense. Intelligence has always been partly scaffolded.
But there is a darker version of this idea, and LLMs make it newly visible. Call it the hollowed mind hypothesis.
The hollowed mind hypothesis is that certain cognitive tools do not merely extend the mind. Used badly, they hollow it out. They create the appearance of increased capability while quietly degrading the underlying human capacity to initiate, attend, interpret, decide, and express. The result is not a superpowered cyborg mind. It is a dependent, thinned-out, increasingly passive mind, surrounded by fluent machinery.
The canonical case is email.
Someone starts using an LLM to summarise long emails and Slack threads. Fair enough. There is too much noise. The machine compresses it. Then they use the LLM to draft a reply. Also fair enough. Writing can be tedious, and sometimes you just need the thing done.
But then a pattern emerges. The person no longer really reads the original message. They read the summary. They do not sit with the ambiguity, tone, politics, implied asks, hidden risks, or emotional texture of the communication. They consume the machine’s abstraction of it. Then, instead of forming an actual response, they ask the same machine to produce one. Perhaps they lightly edit it. Perhaps not.
At this point, what exactly has the human done?
They have not really read. They have not really thought. They have not really written. They have supervised a transaction between someone else’s words and a probabilistic text machine. The human remains “in the loop” in the thinnest possible sense: clicking, approving, nudging, forwarding. But the actual muscles of cognition are not being exercised.
The grim irony is that this person may feel no less stressed. In fact, they may feel more stressed. They are still surrounded by obligations. They still have the inbox, the messages, the social risk, the decisions, the vague fear that something important has been missed. But now they are also increasingly unable to start unaided. Every act of communication begins to require a prosthesis. The tool that was supposed to reduce friction becomes the condition of action.
This is the difference between extension and atrophy.
A notebook extends memory if you actively use it to think. A spreadsheet extends reasoning if you understand the model and can interrogate it. A map extends navigation if you still know what journey you are making. But a system that repeatedly substitutes for attention, judgment, and expression may not be extending the mind in any meaningful sense. It may be replacing the occasions on which the mind stays fit.
This is not an argument against LLMs. That would be stupid. They are extraordinary tools. They can accelerate drafting, critique arguments, explore options, summarise noise, translate between levels of abstraction, and help people get unstuck. Used well, they are genuinely powerful. The point is not that using an LLM is bad. The point is that delegation has a shape, and some shapes are corrosive.
The danger zone is when the LLM is inserted at both ends of cognition: first to digest the world for you, then to emit your response to the world. Input is outsourced. Output is outsourced. The remaining human role is managerial, but not in the strong sense of judgment and direction. More like a tired middle manager rubber-stamping work they barely understand.
Over time, this produces a person who can process more but understand less. They can respond faster but mean less. They can maintain the appearance of competence while losing contact with the sources of competence: attention, memory, taste, agency, and the ability to tolerate the blank page.
The practical test is simple. After using the tool, are you stronger or weaker? Could you explain the matter yourself? Could you write the reply unaided if needed? Has the LLM helped you think, or has it allowed you to avoid thinking while preserving the external appearance of thought?
The extended mind was a hopeful idea: the human mind reaching outward into the world. The hollowed mind is its inverse: the world reaching inward and scooping out the human contribution while leaving the interface intact.
The terrifying thing is that from the outside, the two can look almost identical.