Predictive models, irreducible modalities, and sovereignty

What an LLM actually computes

A large language model (LLM) approximates a probability function over a discrete vocabulary. Given a context of tokens (x_1, \ldots, x_t), the model produces a distribution (P(x_{t+1} \mid x_1, \ldots, x_t)). At inference we sample or pick the highest-probability token; in training we adjust parameters to maximize the likelihood of observed data.

The correct technical analogy is not “a mind that thinks,” but a predictive compressor: the system learns statistical regularities of sequences. The better the predictive compression over a corpus, the more fluent the generated text. That is real, measurable engineering — perplexity, cross-entropy loss, throughput. It is not, by itself, a theory of meaning.

Shannon already distinguished quantity of information (average surprise) from semantic content. The LLM operates comfortably in the first register. The second requires other categories.

Irreducible modalities: Dooyeweerd’s map

Herman Dooyeweerd describes created reality as structured in irreducible aspects (modalities): numerical, spatial, kinematic, physical, biotic, psychical, analytical (logical), historical, lingual, social, economic, aesthetic, juridical, ethical, pistic, among others in the classical systematization. No aspect explains the others without remainder. Absolutizing one aspect is the typical gesture of modern theoretical idolatry.

Two aspects matter closely for LLMs:

1. Analytical (logical) aspect — distinction, formal coherence, inference, predicate consistency. Conditional probability, the linear algebra of attention matrices, and loss optimization live here.
2. Lingual aspect — symbolic signification, reference, speech acts, interpretation within a community of sense. Here belong assertoric truth, metaphor, promise, lie, and responsibility for what is said.

The operational thesis: token prediction is a typically analytical operation on representations that participate in the lingual without exhausting it. Reducing language to prediction is an aspectual closure — useful as an engineering model, false as ontology.

Didactic analogy: the checker and the signer

Imagine an advanced spelling checker. It suggests the next word with high precision. No one attributes authorship of a contract to the checker. A contract requires a signer: someone who bears juridical and ethical consequences of the utterance.

The LLM resembles a checker at scale. It can draft fluent clauses. It cannot, in the proper sense, sign. Analytical-statistical fluency does not, by magical leap, generate linguistic responsibility.

Another analogy: a voice synthesizer reproduces phonemes with fidelity. That does not make it a speaker in a debate. The sonic (or textual) form may be impeccable; the speech act remains human — or at least attributable to a legal person who answers.

Attention, context, and the limit of closure

The transformer architecture (Vaswani et al.) computes weighted relations between sequence positions. “Attention” here is a technical term: similarity in vector space, not moral care. Confusing the two is a dangerous metonymy.

The token-window context is finite. The linguistic community, by contrast, is historical: dictionaries, institutions, precedents, mutual correction. Dooyeweerd would insist that the historical and social aspects co-constitute living language. The model sees a statistical slice; language as practice exceeds that slice.

Basden, applying Dooyeweerd to information systems, recommends multi-aspectual analysis: ask which aspects the artifact activates, which it constrains, and which it ignores. An LLM in production strongly activates the analytical and the economic (inference cost); it can distort the juridical (privacy, attribution) and the ethical (deception, manipulation) if the design treats output as neutral.

Divine sovereignty and the refusal of two extremes

In the Reformed horizon — and convergently with the cosmonomic idea — God’s sovereignty affirms that no created sphere is autonomous in an absolute sense. Technique does not save; technique is also not a demon. It is cultural creation under norm.

Two extremes fail:

• Techno-utopianism: if the model predicts well, it “understands”; if it understands, it may replace human judgment. Absolutizes the analytical aspect.
• Dualist techno-phobia: if it is not a mind, it is irrelevant or impure. Neglects the legitimate craft of engineering within creation.

Divine sovereignty authorizes another path: technical excellence without worship; limits without panic; personal and institutional responsibility that cannot be outsourced. The content steward who publishes model-assisted text remains the moral author of what they sign.

Practical criteria of discernment

For editorial and engineering use, four simple tests:

1. Reduction test — does the metric (BLEU, human preference, loss) measure the analytical/productive, or does it claim full linguistic understanding?
2. Signer test — who answers if the text causes harm? If the answer is “the model,” governance is broken.
3. Multi-aspectual test — beyond accuracy, were juridical, social, and ethical aspects of deployment considered?
4. Attribution test — can the reader distinguish statistical assistance from responsible voice?

input (prompt)
  → analytical operation: P(token | context)
  → fluent output (participates in the lingual)
  → responsible act: human review + signature

The final arrow is not optional. Without it, there is simulated discourse without an economy of truth.

Conclusion

LLMs are powerful instruments of sequential modeling. Dooyeweerd supplies the grammar to say what they are and are not: excellence in the analytical aspect does not liquidate the lingual aspect, nor God’s sovereignty over pluriform creation. Predicting the next token is one task; meaning before the neighbor and before God is another. Mixing the two is the theoretical — and pastoral — error this periodical refuses with clarity.