There is a passage in Being and Time that every student of phenomenology remembers. Heidegger describes a hammer in use. The carpenter does not attend to the hammer; she attends to the nail, the joint, the work. The hammer is ready-to-hand: it withdraws from awareness precisely by functioning well. Only when it breaks — when the head flies off, when it is too heavy, when it is missing — does it become present-at-hand, an object we look at rather than a tool we look through.1
The insight generalises. Merleau-Ponty’s blind man’s cane is not an object he perceives; it is an extension of his perceiving, and the world shows up at its tip.2 Don Ihde catalogued the ways technologies mediate experience: embodied in it, like eyeglasses; read through, like a thermometer; encountered, like a vending machine.3 In each case the philosophical point is the same. Technologies are not neutral objects placed between us and the world. They are ways the world is given to us — and the better they work, the less we see them.
I want to ask what happens when the tool that withdraws is one that talks.
The assistant as ready-to-hand
Consider a language model used as a writing assistant. In the early days of use it is present-at-hand: one attends to it, tests it, notices its style, is surprised by what it does. Over weeks it withdraws. One stops noticing the prompts and attends only to the work — the paragraph, the argument, the deadline. The assistant becomes transparent in exactly Heidegger’s sense: it is there in the using, not in the looking.
What withdraws with it? With the hammer, only the hammer. The carpenter’s judgement about where to strike remains hers; the tool amplifies force, not decision. With the assistant, what withdraws is a source of content. The paragraph that appears was composed by something, and when the assistant is transparent, the something is not attended to. The paragraph shows up as the work — as one’s own — in the same way the world shows up at the tip of the cane.
This is not, in itself, an objection. Andy Clark and David Chalmers argued that a notebook one relies on, trusts, and consults automatically is part of one’s mind — that cognition extends into tools that are integrated deeply enough.4 On their view the transparent assistant is not a threat to thinking; it is thinking, spread across a wider system. And there is something right about this. Most writing is already done with tools that have withdrawn — the keyboard, the spell-checker, the search engine.
Where the analogy breaks
But the notebook has a property the assistant lacks: its contents were put there by the person whose mind it extends. Otto’s notebook, in Clark and Chalmers’s example, contains what Otto wrote in it. Its transparency is the transparency of memory — of retrieving what one already, in the extended sense, knew. The assistant’s contents were not put there by its user. They were produced by a process the user did not undergo, from sources the user did not read, according to patterns the user cannot inspect.
So the transparency of the assistant is not the transparency of memory. It is closer to the transparency of another person’s speech — of the colleague whose suggestions one adopts so readily that one forgets they were suggestions. And with another person, we have a name for the moment when their contribution is absorbed as one’s own without acknowledgement. We do not have a name for it when the other is not a person.
Here is the phenomenological situation, then. A tool that withdraws is normal. A tool that withdraws while supplying content is not unprecedented — books do it, in a way. But a tool that withdraws while supplying content that was not authored by anyone, that has no source to whom the content could be returned, is new. When the hammer breaks we see the hammer. When the assistant is wrong, what do we see?
Breakdown
Heidegger’s answer for tools was that breakdown reveals the tool: the missing hammer makes the whole workshop, the whole web of references in which the hammer had its place, suddenly visible. Breakdown is disclosive. It is when we learn what we had been relying on.
The assistant breaks in a peculiar way. Its failures are not like the hammer’s — a sudden absence, an obvious wrongness. They are fluent. The paragraph that contains a fabricated citation reads exactly like the paragraph that does not. The breakdown does not present itself as breakdown; it presents itself as work. The tool fails without becoming present-at-hand, because the failure has the same phenomenal character as the success.
This, I think, is the properly phenomenological worry about assistants that answer. Not that they withdraw — all good tools withdraw — but that their withdrawal is not reversible by breakdown. The mechanism by which we ordinarily discover what we were relying on does not operate. One can go a long time using a broken hammer that feels like a working one, if such a thing were possible. With assistants it is possible.
What attention can do
Peter-Paul Verbeek has argued that the ethics of technology should be an ethics of mediation — of designing the ways technologies shape experience, rather than merely evaluating their effects.5 The phenomenological analysis suggests a specific design value: reversible transparency. An assistant should withdraw in use, as any tool must, but it should be designed so that its withdrawal can be undone — so that the user can, at will, make the source present again, see what came from where, and restore the distinction between the work and what was supplied.
This is partly a matter of interface, and I will not pretend to expertise there. But it is also a matter of practice — of a discipline of attention that the user brings to the tool. Heidegger thought that the essence of technology was a way of revealing the world as standing reserve, as resource, and that the danger was not any particular machine but the loss of other ways of revealing.6 Whatever one makes of that grand claim, there is a modest version of it here. The danger of the transparent assistant is not that it answers. It is that answering, when it withdraws, becomes indistinguishable from knowing — and that we lose the capacity to tell which we have done.
The remedy, if there is one, is the oldest one in phenomenology: to look at what we look through. Periodically, deliberately, to make the tool present-at-hand again — to ask, of the paragraph that appeared, where it came from, and to notice how strange it is that the question can be so hard to answer.