Mind, Matter, and Meaning

Information and Meaning

“Information” and “meaning” name different achievements. A system can carry and process information without every informative state having semantic content. The first question is whether anything fixes an accuracy condition at all. A further question then asks whether that content is derived from an interpreting practice or original because its correctness standard is fixed independently of one. Ownership settles neither question.

Two Senses of Information

Shannon information is a measurable statistical relation: one state carries information about another when it reduces uncertainty about it. Smoke carries information about fire; tree rings about age; a thermometer’s strip about temperature. This notion by itself contains no truth, reference, or error.

Semantic content concerns what a state represents and how it can misrepresent. A representation can be causally effective while false. That possibility requires more than covariance.

Dretske’s Bridge

Fred Dretske tried to naturalize content by beginning with lawful indication and adding a history in which an organism learns to use the indicator. The acquired function distinguishes what the state was recruited to indicate from whatever happens to cause it now. Teleosemantics generalizes the strategy through selected producer-consumer organization.

The bridge succeeds only if the proposed mapping does genuine explanatory work. It should help explain why the mechanism was stabilized, predict consumer success and failure, and remain privileged under interventions that separate rival correlations. When several mappings remain explanatorily equivalent, content may be indeterminate at that grain. When the mapping merely repeats an engineer’s intention, the content remains derived.

Observer Dependence

Searle is right that unrestricted “information processing” talk can be observer-relative. Almost any physical system permits some mapping from its states to symbols. Calling one mapping semantic explains nothing unless the mechanism itself supplies constraints that privilege it.

But it does not follow that physical mechanisms contain only causal traffic until an owner appears. Selection, learning, and artificial training can organize producers and consumers so that one correspondence rule explains performance better than its rivals. Such a state may have non-derived mechanism-level content. The state need not be a subject, and its containing system need not understand it.

The AI Case

A text-trained language model contains causally potent states sensitive to linguistic distributions. If a linguistic-context mapping survives interventions on training, producer states, downstream consumers, and loss-relevant success, those states may carry causally internalized derived content at that narrow level. The vocabulary, corpus, task, and correctness standard still descend from public linguistic practice. This grants more than the claim that an engineer’s interpretation is non-arbitrary without pretending that training originated the semantic standard.

The grant remains limited. Change the world while holding the corpus fixed and the trained mechanism does not follow the world; change the corpus while holding the world fixed and it follows the corpus. Textual content need not amount to reference to particular distal objects. Nor does content in a submechanism establish whole-model belief, understanding, perception, or consciousness.

Those larger attributions depend on how content-bearing states are integrated and poised across memory, learning, inference, conflict resolution, planning, and action. Self-maintenance may bear on whether the system counts as an autonomous agent. It does not turn information into meaning.

My View

Information remains necessary background and nowhere near sufficient. Semantic content begins where an accuracy condition is fixed, although that content may still be derived from an interpreting practice. Originality begins only where a producer-consumer organization and its history fix the correctness standard independently of such a practice; it does not require ownership by a proprietor.

This preserves both cautions the machine case requires. Causal sophistication alone does not guarantee semantics. But lack of self-maintenance does not disqualify a mechanism whose history and organization genuinely fix correctness conditions. Mechanism content, distal reference, system-level understanding, and phenomenal character remain separate achievements.