AI writing often feels generic and devoid of personality, but the underlying technical cause goes deeper than poor prompting. Writing in The Register, commentator Claudio Nastruzzi highlights a fundamental algorithmic phenomenon termed "semantic ablation"—the systematic erosion of high-entropy information during model generation.
How RLHF and Greedy Decoding Drive Semantic Ablation
Unlike hallucinations, which represent additive errors where a model introduces false facts, semantic ablation is a subtractive bias built into model design. It occurs as a byproduct of greedy decoding and reinforcement learning from human feedback (RLHF).
To maximize statistical probability, LLMs gravitate toward the center of a Gaussian distribution during text generation. In doing so, the model discards rare, precise, and complex "tail data" tokens in favor of low-perplexity outputs. Aggressive safety and helpfulness tuning further exacerbate this problem by deliberately penalizing unconventional phrasing and linguistic friction.
If hallucination is an additive error where AI sees what isn't there, semantic ablation is a subtractive error that destroys what is actually present.
The Three Stages of AI Text Degradation
Nastruzzi explains that running human text through successive AI refinement loops causes a collapse in vocabulary diversity, measured by a drop in the type-token ratio. This degradation unfolds across three distinct stages:
- Metaphoric cleansing: The AI flags unconventional metaphors or sensory imagery as noise because they deviate from training set averages, swapping them for dead, predictable clichés.
- Lexical flattening: Domain-specific jargon and high-precision vocabulary are sacrificed for accessibility, replacing rare 1-of-10,000 tokens with generic 1-of-100 synonyms.
- Structural collapse: Complex, non-linear logical arguments are forced into standardized, low-perplexity templates to satisfy artificial readability metrics.
The result is what Nastruzzi terms a "JPEG of thought"—an output that appears visually coherent and polished on the surface but has been stripped of its underlying data density and original intent.
Affected Users and Technical Trade-offs
This algorithmic smoothing directly impacts researchers, technical writers, and creators who rely on AI tools to polish draft copy. While aggressive filtering reduces erratic output, the trade-off is a severe loss of semantic density and unique argument structure.
An honest caveat of this system is that ablated text remains syntactically flawless and easy to read. Because the output looks clean and frictionless, human reviewers frequently fail to notice that critical nuance and structural integrity have been removed.