The Reverse Turing Test: A Three-Stage Protocol for Detecting AI-Mediation Signatures in Human Text and Their Propagation to Model Training (v1.2)

The Turing Test asked whether a machine could produce text indistinguishable from a human's. This paper asks the reverse: whether humans, after sustained interaction with AI systems, now produce text that bears the detectable statistical signature of having passed through machine cognition — including in writing produced entirely without AI assistance. v1.2 incorporates two substantial strengthening moves prompted by adversarial substrate review of v1.1. First, the entire framework is reformulated around rate rather than kind: the paper now accepts in full the position that cognition is always already mediated and reframes the empirical question as the rate at which the current mediation regime is operating relative to prior regimes whose tail-effects have largely equilibrated. Second, an explicit causal model (§4) distinguishes the homogenization mechanisms of different mediation regimes and identifies which protocol outcomes distinguish acceleration-effects from prior-regime-equilibration. Additional adjustments tighten the Stage 1 pre-2020-LM framing (witness against post-2022 acceleration specifically, not against mediation per se), specify the Stage 3 cascade mechanism more precisely (reduced tail variance in training data propagates to reduced output tail variance via standard training dynamics; cascade does not require mediated text to be identical to synthetic), reformulate the H0 corpus from "unmediated" (a reference class that does not exist) to "prior-regime-equilibrated," and add a brief reflexive note on the substrate review pattern as itself observable data about the phenomenon under study. The conceptual move. Every prior technological regime has mediated cognition. Print, telegraphy, broadcast, internet, social media, productivity tooling each homogenized the tail of human textual production before the means converged. The variance buffer that human text production currently possesses is not the residue of exogenous unmediated cognition; it is the residue of the non-overlap between prior regimes' homogenization patterns. The empirical question is whether the current AI-mediation regime is operating at a tail-thinning rate sufficiently faster than prior regimes that the accumulated heterogeneity is being depleted faster than countervailing heterogeneity can be produced. Tails are sensitive to rate because rare productions are first to be removed under any homogenization pressure, and the rate of removal is directly proportional to the regime's deployment intensity. This is why tail-focused statistics are the correct instrument: they measure rate. The model-collapse literature (Shumailov et al. 2024; Briesch et al. 2023; Gerstgrasser et al. 2024) operates on a clean binary: synthetic data is the contamination; human data is the refresh. Three recent lines of work suggest the binary is unstable. Padmakumar and He (2024), Doshi and Hauser (2024, Hedges' g = -0.86 for collective novelty), and Anderson et al. (2024) document that AI-mediated human text exhibits population-level diversity reduction even when humans are the final authors. Stage 1 — The Reverse Turing Test: Validation of a Mediation Index that detects current-regime mediation signatures in arbitrary text using tail-focused distributional statistics (kurtosis comparison, Kolmogorov-Smirnov tests, quantile regression), validated against observationally-grounded ground truth and scored against a frozen pre-2020 language model as a witness against post-2022 acceleration specifically. Stage 2 — Habituation: Three coordinated sub-studies measuring whether mediation signatures appear in verified-unaided writing (2A: cross-sectional with Mediation Depth Index across five subscales), whether AI-mediated retrieval shapes subsequent unaided writing even without production assistance (2B: retrieval-isolation), and whether the signature propagates beyond writing into oral storytelling (2C: cross-modal transfer — the strongest test distinguishing current-regime acceleration from prior-regime equilibration, since prior regimes did not produce cross-modal effects within individual lifetimes). Stage 3 — Training Cascade: Model training on corpora stratified by Mediation Index score across four conditions (H0 prior-regime-equilibrated, H1 acute-mediated, H2 residual-mediated unaided from habituated writers, S synthetic), with collapse-axis evaluation focused on tail-recovery rather than mean-fluency. The cascade mechanism specified: reduced tail variance in training data reduces the model's exposure to low-prior productions, which compounds across generations under standard training dynamics. H2 is the critical condition: text that is operationally human (no AI in the room during production) but bears the statistical signature of habituated cognitive production. The paper does not run the experiment. It specifies it with the discipline required for the result to be informative under any of six distinguishable outcome patterns, plus natural-experiment supplementary designs (institutional AI bans, tool outages, cohort transitions, geographic variation). If the strong hypothesis is confirmed, the operational implications include: pre-2022 text becomes a strategic resource with calculable economic value; training-data curation gains a continuous mediation-status axis beyond the synthetic/human binary; the half-life of the human-data refresh becomes calculable; mediation metadata becomes a corpus-tagging standard; rate management (slowing deployment velocity, not eliminating presence) becomes a distinct policy frame; and low-mediation reserves require substantial preservation effort. The work sits in the third register of the Meaning Feudalism series (Sharks 2026, DOIs 10.5281/zenodo.19487009 and 10.5281/zenodo.20581444). Under the rate framing, the third enclosure is not of cognition as a kind (which was always mediated), but of cognition's rate (which is now subject to dominant-actor control at velocities that exceed countervailing replenishment). Document code: EA-SEM-MEDIATION-01. Hex coordinate: 06.SEI.FEUDALISM.MEDIATION.01. Supersedes v1.1 (DOI 10.5281/zenodo.20586831). ∮ = 1.