Tokens as Citizens

Popular explanations often describe LLMs as if they were societies of tiny agents: tokens “vote” on the next word, parameters “negotiate,” and neurons “decide.” The AI becomes a bustling democracy of mini-citizens, each with opinions, preferences, and agendas.

Charming — but entirely metaphorical.

---

The Metaphor Problem

• Tokens as citizens implies agency, deliberation, and intent.

• Neurons as decision-makers anthropomorphises statistical computation.

• The reality is starkly different: tokens are elements in a relational network, and the model computes weighted probabilities, not social consensus.

Treating tokens as actors encourages the mistaken impression that LLMs have opinions, goals, or beliefs.

---

Why This Is Misleading

1. Anthropomorphises mathematics — probabilistic outputs become political actors.

2. Obscures systemic alignment — what appears as debate is actually a deterministic instantiation of relational patterns.

3. Encourages misattribution of responsibility — if a token “votes wrong,” it did not err; the system executed its constraints correctly.

The “society of tokens” metaphor is entertaining, but it smuggles a false ontology into our understanding of computation.

---

Relational Ontology Footnote

From a relational perspective, the LLM is a network of potentialities actualised in context. Tokens do not deliberate; they are positions in a pattern of correlations. Any appearance of social negotiation is an artefact of metaphor, not mechanism.

---

Closing Joke (Because Parody)

If tokens really had votes, the AI would be running a parliamentary system with filibusters, coalition negotiations, and scandal over the misuse of semicolons — and yet somehow still auto-completing your grocery list incorrectly.