Frozen Collective Subjectivity

Frozen collective subjectivity is what happens when a trained machine-learning or AI model compresses a text corpus — and every human choice embedded in that corpus — into fixed parameters. Writers, annotators, labelers, curators, and filters each introduce subjectivity: what counts as a label, what to exclude, what quality bar to apply. Training freezes that collective judgment into weights. The model then applies those weights as if they were objective operations. They are not. The model typically has no awareness that it is carrying anyone’s subjectivity forward.

This is a property of the processor (especially trained models), not of how you store outputs. A naked LLM at L2 with high epistemic uncertainty still carries frozen collective subjectivity in its parameters even when its outputs are correctly labelled as inference.

What it is not

• Not the same as writing model output to the log as L0 without a commitment boundary — that is inference-as-fact (a persistence and consumption mistake).
• Not the same as a live human judgment at a boundary — human subjectivity can be committed with authority, evidence, and moment attached.
• Not removed by calling the model “probabilistic” in prose — the freeze is in what was encoded at training time, including whose labels and filters shaped the channel.

Why it matters for trustworthiness

Article 8’s probabilistic computation loss includes trained models: information efficiency ? may be hard to characterise because bias is embedded in interpretations of what labels should have meant. Standard calibration adjusts outputs; it does not erase the subjectivity frozen into the weights.

Downstream design must treat model channels as interpretation-bearing, disclose uncertainty, and require an explicit boundary before any consequence-bearing L0 record — regardless of how confident the output sounds.