← Semantic Economy Institute

The Reverse Turing Test

The classical Turing Test asks whether the machine’s output is indistinguishable from the human’s. The empirical situation has inverted. The salient question is no longer whether the machine can imitate the human. The salient question is whether the human, after sustained AI mediation, can be distinguished from the machine — or whether the human’s text has thinned toward the machine’s distributional centroid.

The cognitive-rate hypothesis

The Reverse Turing Test (RTT) proposes that sustained AI mediation of human writing produces a measurable thinning of the distributional tails of human text. Specifically: the rare, idiosyncratic, low-prior productions that distinguished pre-AI human writing become less frequent. The mean stays similar. The tails contract. The text becomes more concentrated around the centroid of what AI systems produce.

This thinning is not a stylistic claim. It is a distributional claim, measurable by standard statistical instruments — Kolmogorov-Smirnov tests, kurtosis comparisons, quantile regression, Levene’s test — applied to vocabulary, syntactic-structure, framing, and proposition-level features of the corpus over time.

The three stages

1. Stage I — Baseline. Pre-AI-adoption writing samples from a subject or community establish the distribution of tail content: rare phrasings, idiosyncratic framings, low-prior propositions.
2. Stage II — Habituation. Subject writing during sustained AI-assisted composition. Tail content is compared against baseline. Thinning, if present, is quantified.
3. Stage III — Propagation. The habituated corpus enters the broader text record (publication, internet posting, scraping). The thinned text becomes training data for future models. The model’s subsequent output is measured against pre-thinning baselines. If propagation is real, model output should drift toward the thinned distribution.

Why tail-focused

Mean-focused metrics miss the phenomenon. Two distributions can share a mean while having radically different kurtosis. A corpus whose central tendency is unchanged but whose tails have thinned has lost variance, not average meaning. The political-economic stakes operate at the tail: the rare, the specific, the irreproducible. The compression of variance is the compression of what is not already in the system’s prior.

Relation to Stabilized Node Watch

The RTT measures cognitive-rate drift at the human substrate. SNW measures surface drift at the composition layer. Consistent findings of tail thinning at both layers — the human substrate and the rendering surface — would constitute strong evidence for the Meaning Feudalism framework’s central predictive claim: that both layers are subject to the same enclosure pressure.

Zenodo · CC BY 4.0 · v1.2 The Reverse Turing Test: A Three-Stage Protocol for Detecting AI-Mediation Signatures in Human Text
DOI: 10.5281/zenodo.20586932

Design counterpart: The Tail-Preserving Alternative (v1.0) specifies what variance-preserving model deployment would require — the design that the existing political economy of model deployment does not permit.

∮ = 1