Demographic Bake-In

trigger

""Human baseline: 19 pages" — all male tech essayists, all English, all 2000–2023."

what it is

The training data defines what "normal" looks like, and the analysis inherits that definition without declaring it. The calibration's Category A — the "human baseline" — is 19 male tech essayists writing in English between 2000 and 2023. Every feature that discriminates "human" from "AI" is actually discriminating "this demographic writing in this genre in this era" from "AI." A woman writing in French, an academic writing a paper, a non-native English speaker writing a blog post — all would score differently against a baseline they were never part of. The demographic is invisible because it is the default. The same mechanism as RLHF: whoever picked the training set picked the answer.

what it signals

In prose slop: the unstated premise the reader is expected to accept without noticing. In analytical slop: the unstated demographic the analysis treats as universal. The hiding mechanism is identical. The medium is different. The medium makes it harder to see because "19 pages across 6 categories" sounds like science. It is not science. It is a convenience sample dressed as a representative one.

instead

Declare the demographic. "Baseline: 19 English-language tech essays by 11 male authors, published 2000–2023." Then let the reader decide how far to generalise. The honest version names the sample for what it is. The sloppy version calls it "human baseline" and lets the reader assume universality.

refs

• AnotherPair calibration v3 session 2026-03-01
• Captain: 'How do I control for slop inside the analysis?'
• Layer 7 in the bias stack
• Same mechanism as RLHF reward function bias