Mind, Matter, and Meaning

Derived and Original Intentionality

Some things mean what they do because a user or practice lends them meaning. Other states have accuracy conditions that do not depend on anyone interpreting them that way. John Searle calls the first derived intentionality and the second original or intrinsic intentionality. The distinction concerns dependence on interpretation, not possession by an owner.

A stop sign really means stop, but it means that through a public practice. Remove the practice and the metal retains its shape and causal powers but not its instruction. By contrast, if a frog’s producer-consumer circuitry was stabilized because one state guided prey-directed behavior under one condition and another state under another, that mapping may help explain what the states can get right or wrong without waiting for an observer to assign them meanings.

Three Questions

Artificial systems make three questions especially easy to confuse:

  1. Content fixing: Does the proposed mapping explain the mechanism’s stabilization and survive interventions on its producers, consumers, and success conditions, or is it merely a convenient gloss?
  2. Level of attribution: Is the content properly attributed to a component, a coupled mechanism, or the whole system?
  3. Scope of capacity: Does content in one mechanism support perception, belief, understanding, planning, or consciousness at the level of the larger system?

An affirmative answer to the first question does not settle the other two. A subpersonal state can carry content without becoming a tiny subject. A system can contain content-bearing mechanisms without thereby understanding their content.

To consider a robot as possibly having original content with a subject, both the content-fixing and attribution questions need independent support. The relevant history and producer-consumer use must fix a world-involving correctness standard not exhausted by an interpreting practice. An inherited task does not by itself decide whether they do. Interventions must also locate one relatively maximal, persisting control organization in which those states update a shared economy of perception, expectation, memory, correction, planning, and action. The first finding makes original content a live attribution; the second identifies its candidate subject. Neither finding establishes consciousness, and the second does not create the first.

Original Does Not Mean Uncaused

Original content need not arise from nowhere. Natural selection, individual learning, and perhaps artificial training can establish producer-consumer mappings whose semantic interpretation does explanatory work. Original concerns whether history and use fix accuracy conditions not exhausted by an interpreting practice. It does not require freedom from causal or semantic inheritance. A training process can make an inherited interpretation causally indispensable without making it independent; inheritance alone cannot show that independence is absent either.

The positive test matters. A representational description earns realism when the mapping helps explain why the mechanism has the organization it does, predicts its success and failure, and remains privileged across relevant interventions. If rival mappings perform equally well throughout, the content may remain indeterminate at that grain. If the mapping merely redescribes what an engineer intended the device to do, the content remains derived.

Original Does Not Mean Novel

A second confusion travels under the same word, and it costs more arguments than the first. In ordinary usage, original content names something no one has produced before — a new sentence, a new image, a new tune. Language models produce that in quantity, and some of it repays attention. Nothing in this distinction denies it.

The philosopher’s sense asks a different question. Not whether the output has appeared before, but what fixes the standard that makes it right or wrong. Novelty belongs to the product; originality in this sense concerns the standard’s dependence on interpretation. A system can produce unprecedented output indefinitely while every standard it answers to remains dependent on an interpreting practice, and a system can produce dull, repetitive output whose accuracy conditions its own history and use fix. The two properties vary independently.

Keeping them apart matters because the popular sense makes the philosophical claim sound like a denial of the obvious. It denies nothing about what these systems can write.

The LLM Case

An emitted word carries public, derived meaning as an English token. Inside a language model, training produces elaborate geometry, learned sensitivities to textual distributions, and states that make indispensable causal differences to later computation. Fintan Mallory applies consumer-based teleosemantics to word2vec and argues that training can give its embeddings original content. His claim concerns actual word-types and the contexts in which they occur, not the numerical vectors themselves or the objects the words name.

On Mallory’s account, an embedding combines a description with a directive: a word-type occurs in a particular context, and the downstream mechanism should produce a distribution over context words. Training links the two. Its updates respond to differences between predicted distributions and those in the data. The proposed content helps explain how producers and consumers acquired their organization; it does not rest on the geometry alone.

The external dictionary raises a question about which relations fix content, not a quick refutation. A coordinated change to word indices and the corresponding matrices can preserve the relation between actual words and their contexts. Changing the external pairing instead changes a candidate content-fixing relation. Neither operation alone establishes dependence on an interpreting practice. Human children inherit public standards too; both human and artificial learners require assessment through the history and use that make their states answerable to what they concern.

Linguistic distributions belong to the world too. A model can track facts about the word maple without tracking the maple outside the window. Changing the tree while holding the corpus fixed therefore tests a proposed reference to that tree, not the proposed content about word use. The distinction limits the claim without dismissing it. Where content does remain dependent on an interpreting practice, derived still does not mean inert, superficial, or optional: such content can operate inside a network and help explain what it does.

Nor does mechanism content establish whole-model understanding. That requires a further account of integration across memory, learning, inference, conflict resolution, planning, and action. Appropriate perceptual organization and poising would be needed for the book’s strong representationalist account of consciousness. These are questions about what the larger system can do with content, not extra ingredients that turn content on.

Self-Maintenance and Autonomy

Reciprocal self-maintenance may help explain autonomy, agency, persistence, welfare, and the boundaries of an individual. It does not create accuracy conditions. Adding a self-repair loop to a content-bearing mechanism need not change what its states represent; removing the loop need not make their content derived. Ownership can still matter to whose belief, plan, or experience a state helps constitute. It does not mint the state’s content.

My View

I retain the original/derived distinction as a diagnostic of semantic dependence, which can vary by content and grain. Selection, learning, and artificial training may establish original content where history and producer-consumer use fix accuracy conditions not exhausted by an interpreting practice. An objective causal role alone does not settle that question; neither does the ancestry of the vocabulary, corpus, task, or loss. Mallory supplies a serious account of mechanism-level content about actual word-types and linguistic distributions. Whether that content is original at this limited grain remains unsettled. Either kind of content leaves open whether the whole system believes, understands, perceives, or experiences anything.

This gives internal states their fair due. They can carry derived content without being passive inscriptions, miniature subjects, or proof of understanding by the machine that contains them. Mechanism-level names where the content operates; it does not name a third semantic grade between derived and original.

  • Searle: the original/derived/as-if distinction
  • Dennett: the interpretationist alternative