Philosophical Method
philosophy's working tools, conceptual analysis, thought experiments, and consistency arguments, settle what a claim means and whether it holds together, while observation settles which of the coherent options is actually true, and confusion about philosophy's authority comes from letting either side do the other's job.
Essence
an experiment cannot tell you what "knowledge" or "fairness" means, and an armchair cannot tell you the boiling point of water at a given pressure. Philosophy's jurisdiction is the prior question, what a claim amounts to and what it is consistent with; once that is fixed, which world we are actually in passes to observation.
Question
A brain scan can log every signal moving through a machine while it answers questions fluently. Suppose the scan is perfect: every unit, every weight, every pass recorded. Does that settle whether the machine thinks?
It does not, and the reason is worth sitting with. "Thinks" is a word, and before any scan can bear on the question, something has to fix what the word requires: whether thinking demands a subjective feel to the processing, or an internal model of the world, or the right causal history, or nothing more than producing the right outputs under the right conditions. Different answers to that prior question make the very same scan count as evidence of thought or as evidence of nothing of the kind. No amount of additional recording resolves it, because the dispute is not about what the machine is doing. It is about what the word is asking for.
That is two questions tangled into one sentence: a question about meaning, and a question about fact. This entry is about the tools philosophy uses to work the first kind of question, what they can establish, what they cannot, and how to tell which kind of question is in front of you before reaching for either an argument or an instrument.
Definition
Call the first tool conceptual analysis: the search for the conditions a concept requires, stated precisely enough that a case can be checked against them rather than merely felt to fit or not. A necessary condition for a concept is one nothing falling under it can lack; a sufficient condition is one whose presence guarantees the concept applies. Conceptual analysis is the discipline of finding conditions that are both, jointly, for a target concept, then testing candidate conditions against cases built to strain them. This is not guessing at a dictionary entry. It is treating a concept the way an engineer treats a specification, stating exactly what must hold and then hunting for the case that breaks the statement.
Call the second tool the thought experiment: a scenario built to hold everything fixed except one variable, run in imagination because the real-world case is unavailable, too costly, or does not naturally occur in the form needed. A thought experiment is controlled variation of a concept, not idle storytelling. The trolley problem is the familiar instance: a runaway trolley will kill five people unless diverted onto a track where it kills one, and the variants (divert it yourself, or push a bystander onto the track to stop it, same numbers each time) hold the arithmetic fixed and vary only the means, physical distance, or intention involved. The scenario is not evidence about trolleys. It is a device for testing whether a moral concept like permissible harm tracks outcomes alone or also tracks how those outcomes are brought about. Edmund Gettier's 1963 cases work the same way for the concept of knowledge: each holds justification and truth fixed and asks whether a case with both, arrived at by lucky accident, still counts as knowing. Three pages of carefully constructed cases were enough to unsettle a definition of knowledge that had stood for centuries, because the cases did real testing work rather than illustrating a conclusion already reached.
Call the third tool the consistency or regress argument: checking whether a position, taken together with its own commitments, can be held without contradiction, or whether it generates a demand it can never satisfy. A regress argument shows that answering a question in a certain way requires answering the same question again one step back, so that the answer never actually terminates. These arguments establish nothing about how the world happens to be. They establish what a position costs, given what it already claims.
What these three tools can settle, between them, is meaning, coherence, what follows from what, and the price of holding a given view. What none of them can settle is a contingent fact about how the world happens to have turned out. No amount of analyzing the concept of water will tell you its boiling point at a given pressure; that requires putting water in a container and measuring it.
The boundary runs both directions, and this is the detail most often missed. Philosophy does not only wait for science to fill in the facts once the concepts are fixed. When a philosophical account rests on an empirical presupposition, a claim about how minds, brains, or societies actually work, an empirical result can retire that account outright, not merely add detail to it. An analysis of memory that presupposes memories are stored and replayed like recordings is a philosophical account built on a psychological claim, and evidence that memory instead reconstructs each time it is used can retire that analysis, not just supplement it. The border is patrolled from both sides.
None of this makes the border perfectly sharp. Timothy Williamson has pressed the point that the armchair and the laboratory are less cleanly separated than the tidy division suggests: philosophers routinely lean on assumptions about ordinary usage, cognition, or possibility that are themselves open to empirical challenge, and some scientific work leans on conceptual commitments its practitioners rarely examine. That is a live complication to keep in view, not a reason to give up on locating, case by case, which question is being asked.
Common mistakes
The first mistake asks philosophy for what only evidence can give: treating a careful definition as if it could, by itself, generate a fact about how the world turned out. No amount of analyzing "cause" tells you which of two events caused a particular fire; that is a question for investigators, not for conceptual analysis.
The second mistake asks evidence for what only analysis can give, and the machine-thinking case is the clean instance: expecting a more detailed scan, a bigger dataset, or a longer test to settle a dispute that is actually about what the target word requires. More data answers a factual question. It cannot answer a question about what counts.
The third mistake treats thought experiments as decorative fiction rather than as working instruments, dismissing the trolley variants or Gettier's cases as unrealistic scenarios that prove nothing because they could not literally happen. The point of the construction is control, not realism; an experiment engineered to isolate one variable is doing its job precisely by being unlike ordinary life in every respect except the one under test.
The fourth mistake declares philosophy obsolete because science now answers questions philosophy once asked, without noticing that the declaration is itself a philosophical claim, a position about which questions belong to which method and why, defended or refuted by the same conceptual tools it is dismissing.
Test yourself
Take one live, contested question that plausibly mixes both jurisdictions: does an AI system understand language, is addiction best treated as a disease, is a fetus a person at a given stage. Split the question into its conceptual half and its empirical half.
State the conceptual half as the specific condition in dispute: what exactly "understands," "disease," or "person" is being required to mean for the question to have an answer at all, stated precisely enough that two people could check a case against it.
State the empirical half as a specific observation or measurement that would bear on the question once the conceptual half is fixed, something that could in principle be looked for.
Success is a clean split, stated in one sentence each, plus a third sentence naming where the two halves meet, the exact point where an answer to the conceptual question is needed before the empirical answer means anything at all. If either half turns out to secretly depend on the other with no such meeting point stated, the split was not finished.
Primary sources and further reading
- Timothy Williamson, The Philosophy of Philosophy (2007)Argues the armchair and the laboratory are less cleanly separated than the standard picture claims, a caution this entry states as a live complication.
- Frank Jackson, From Metaphysics to Ethics: A Defence of Conceptual Analysis (1998)The working defense of conceptual analysis as a method with its own standards of success and failure.
- Edmund Gettier, Is Justified True Belief Knowledge? (1963)Three pages that moved a field, the model instance of a thought experiment doing methodological work.