AI VS HUMAN WRITING:
WHAT ACTUALLY CHANGES UNDER THE HOOD?

1. The Mathematics of Next-Token Prediction

Autoregressive language models generate text by calculating a probability distribution over the entire vocabulary for the next token, given all preceding tokens. Even when sampling with temperature, the generated sentence tends to follow the path of maximum average likelihood.

In practical terms, this means an AI almost never selects a truly strange analogy or an ungainly sentence fragment unless explicitly instructed. It optimizes for smooth, plausible continuity. The consequence is text that feels frictionless—and strangely devoid of friction, edge, or personal vulnerability.

2. Human Asymmetry: The Friction of Thought

Human composition does not operate on probabilistic token sampling. When a human sits down to write, they are grappling with messy cognitive memory: a vivid recollection of an argument, a lingering frustration with a software bug, or an idiosyncratic metaphor derived from childhood.

This produces what computational linguists call asymmetric entropy:

• Sudden shifts from formal technical terminology to colloquial slang.
• Intentional sentence fragments for dramatic emphasis.
• Unorthodox punctuation patterns (dashes, parentheses, ellipses).
• Specific, non-round empirical details ("spent $142.18 on cloud egress" vs "scaled our cloud pipeline efficiently").

3. The Perplexity Paradox

If a text is too predictable, it reads as robotic. If it is too unpredictable, it reads as gibberish. Human writers operate in an oscillating middle zone—moving from high perplexity (novel ideas) to low perplexity (clear explanations). Modern LLMs, without strict prompting, gravitate toward uniform low perplexity across entire paragraphs.