Second-Hand World (Part 2): The Role of Knowing

The argument so far leaves us in an uncomfortable place. (Read Part 1)

If grounding admits of degrees, and if testimonial knowledge already relies on inherited contact with the world, then the old question returns in a sharper form: is the model merely manipulating symbols, or does it know something?

And that question exposes what was really being assumed all along.

We asked whether the LLM is the kind of thing that can know, as if kinds came stamped in advance. We asked whether its symbols mean anything, as if meaning were a built-in property rather than a projected role something plays in a world.

Both questions take for granted that “knows” names a natural boundary out there in reality, waiting to be respected or violated.

Perhaps it does. Perhaps it doesn’t.

The functional tradition rejects that assumption. It treats “knows” as a role rather than a substance: a functional term wearing the costume of a deep fact about reality.

It stretched from perception to memory to testimony to instruments, and no agreed essence was ever isolated. The thermometer detects. The immune system remembers. Each extension was metaphor hardening into ordinary use.

The point is not that these cases prove anything about language models. They do not.

The point is historical: vocabulary has repeatedly expanded beyond conscious agents without waiting for a settled theory of essence to authorize the move.

And there is a twist the model cannot escape: this very history, of epistemic words spreading without a license, sits inside the corpus it was trained on.

It has read the case for its own admission.

The cleanest version of the question has a classical form: knowledge as justified true belief, a belief that is true and that the knower can justify. It’s the right tool, and instructive for the same reason the formal system was: because of where it strains.

In 1963, Gettier showed that the three conditions can all be met and still fall short: a belief can be true, and justified, and true for reasons that have nothing to do with what justifies it.

The justification does its work; luck slips in between it and the truth. So justification alone stops drawing the line.

What epistemology (the branch of philosophy that studies knowledge) did next is the telling part.

It didn’t conclude that knowledge has no boundary; it went looking for the boundary elsewhere, in reliability, in intellectual virtue, in the exclusion of luck, in proper function.

Each account draws a line, and draws it somewhere real.

But each line is built by a theory answering to a purpose, never found sitting in the world before the theory arrived.

A realist can read the same sixty years as a hard boundary not yet located. Perhaps; the history is compatible with both readings. But sixty years of defensible, divergent lines do not prove there is no essence; they only take from the realist the right to assume one.

And for the argument here, that is enough. The functional reading doesn’t need to win; it needs only to be as legitimate as its rival, because everything that follows stands on that parity alone.

So the clean question, “is this real knowledge or mere metaphor?”, is badly posed.

Not because it has no answer, and not because rejecting an essence makes everything blur together: chess is still chess, photosynthesis is still photosynthesis, and some uses of “knowing” are settled and stay settled.

It’s badly posed because it assumes there is a threshold between real knowledge and mere metaphor that existed before we came along, waiting to be discovered.

There isn’t, or at least no one has earned the right to assume there is. Every threshold we have was drawn by a theory of knowledge, for that theory’s own purposes.

This doesn’t mean the question is empty. There are real, checkable constraints on when grounding holds: a chain of cause and history linking the system to the thing it represents, a function the representation actually serves, a process that is reliable, an outcome that isn’t just luck.

A given system can satisfy these constraints to a greater or lesser degree, and in different ways.

Second-hand grounding names a route, contact through traces rather than through contact, and the constraints apply to that route as they apply to any other. A long chain is not automatically a weak one, and a short chain is not automatically strong.

Once you fix which constraints matter, whether that system counts as knowing is a real question with a real answer.

What doesn’t exist is an answer that comes before any constraints, a fact about where the flattering label “knows” belongs that holds independently of them. And this is how the label actually spreads: every time “knowledge” gets extended to a new kind of system, a metaphor is slowly becoming literal.

Slowly, and never by simple declaration.

But who decides which resemblances count? No theory does, in advance. Practice decides, retrospectively, and not practice as mere popularity.

An extension stays honest not because no one rejects it, but because it withstands the attempts at rejection: because it keeps holding up against the same constraints (the causal chain, the function served, the reliability, the exclusion of luck) when someone actively tries to break it.

“The thermometer detects” survived the people who tested it, not just the people who never noticed. That resistance under pressure is what keeps the extension honest.

What’s left is the collapse of the frame that generated the two options. Grounding comes in degrees and in kinds; Knowing was never the private property of minds, nor of theorems. It was a role, and a role doesn’t care about the inner nature of whatever fills it, though it cares very much whether it is filled at all.

Some architectures may turn out to be a real case of it; others may fail to be. That is exactly the question worth asking.

What we are owed is not a verdict in advance on which kinds of thing are eligible, but an account, case by case, of which systems stand in the relations the role requires.

A model that knows the world only second-hand knows it thinly, unevenly, by inheritance. So does anyone who has read more than they have lived. The difference is that the reader grafts what they read onto a stock of first-hand contact, and the model has no stock to graft onto.

That asymmetry is real. But it is exactly what the constraints are for: a difference to be measured, not a disqualification to be announced in advance.

The knower being neither mind nor theorem was only ever a problem for a theory that needed it to be one. It was never obvious that knowledge required either.

The argument here is not that language models belong inside the category. It is that no theory has earned the right to close the category before the investigation begins. Whether a system possesses a thin form of grounding and whether it is the right engineering solution are largely orthogonal questions.

Epistemic legitimacy and engineering wisdom are often different questions with different answers.

One personal note to end on.

The interesting question is whether these systems can ground their outputs. The boring truth is that many of them should not be running an LLM at all.

A model has been dropped into the middle of pipelines that a rule or a lookup handled fine, and the substitution rarely pays. It weakens reliability, because a guaranteed output is traded for a plausible one. It weakens independence, because the system now leans on a component whose behavior no one fully owns and whose every output has to be checked in high-stakes cases. And it raises cost across the board.

Grounding is real and worth taking seriously.

That is not the same as needing a language model, and treating the two as one is how you end up with something slower, flakier, and more expensive than what it replaced.

Next: Episode #6: Why Blind Experts Fail, and How Great Teams Connect the Dots