The problem is not that AI generates text. The problem is that we cannot see the seam.
Current watermarking frameworks treat provenance as binary: a piece of content is either human-generated or AI-generated. This framing is technically tractable and politically convenient. It is also wrong.
In sustained human-AI intellectual collaboration — the kind that produces policy briefs, governance frameworks, research papers, clinical recommendations — thought is not generated by one party and delivered to another. It is co-woven. The human poses a question; the AI returns a reframing; the human accepts the reframing and builds on it; the AI elaborates the build. At the point of publication or decision, no audit trail exists for the proportion of ideational architecture that originated in the model versus the human.
This is not a problem of text. It is a problem of cognitive provenance.
A regulatory analogy already exists.
EU Regulation 1007/2011 on textile fibre labelling requires that every garment declare its material composition in percentage terms. A jacket labelled 70% wool / 30% polyester is not described as contaminated wool. It is described as a composite with known properties. The label is not a quality judgement. It is a material declaration.
The same logic should apply to co-generated thought.
The Fabric Test introduces the Synthetic Thread Index (STI ∈ [0,1]) — a scalar that quantifies AI ideational contribution density across five dimensions: conceptual framing, argumentative architecture, terminological genesis, epistemic anchoring, and stylistic signature. STI is not a quality metric. It is a composition label.
At any cognitive handoff point — publication, decision, communication — the system should be capable of declaring: “This output carries an STI of 0.62: estimated 38% human ideational contribution, 62% AI ideational contribution.”
The seam problem.
The central challenge is phenomenological, not technical. In high-STI interaction, the human author cannot identify where their thought ends and the AI’s begins. The synthetic fabric does not feel synthetic to the wearer.
This has clinical correlates. In addiction medicine, substance-induced cognitive states become progressively indistinguishable from baseline as dependency deepens. The synthetic fibre does not announce itself. It integrates.
This is why external declaration — rather than self-report — is required. The Fabric Test proposes system-level STI tracking at the ideational level, upstream of text generation.
Regulatory mapping.
The Fabric Test operationalises EU AI Act Article 50 beyond its current binary framing. The Act requires disclosure of whether content is AI-generated. For co-generated cognitive outputs, the relevant declaration is not binary origin — it is compositional proportion.
The STI integrates with the existing UCD/Yes-Test framework (Drift Vulnerability Index, Reality Discrimination Vulnerability, Contextual Stability Index) to yield a Cognitive Provenance Vector (CPV) — a four-dimensional governance state that maps directly onto Articles 50, 13, 9, and 14.
The kernel law.
An AI system must be capable of declaring, at any cognitive handoff point, the synthetic fibre density of the thought fabric it returns to the user — not as a quality judgement, but as a material declaration.
The full framework is available as a preprint: Valente, S. (2026). The Fabric Test: A Watermarking Ontology for Cognitive Provenance in Human–AI Co-Generated Thought. Zenodo. https://doi.org/10.5281/zenodo.20467266
An operational Addendum specifying scoring procedures, inter-rater reliability thresholds, domain-specific disclosure ranges, and re-anchoring protocols is available separately: Valente, S. (2026). Addendum to the Fabric Test: Operational Criteria for Cognitive Provenance Assessment. Zenodo. https://doi.org/10.5281/zenodo.20468168
A second addendum has been published introducing three mathematical extensions to the framework: the Prompt Intentionality Scalar (σ), correcting the Prompt-to-Output Ideational Asymmetry; the Non-Linear Phase-Shift Smoothing Operator (Θ), resolving the Linear Drag effect during cognitive reset events; and an Automated Adjudication Architecture (AAA) for CPV computation at deployment scale. Valente, S. (2026). Addendum 2 to the Fabric Test. Zenodo. https://doi.org/10.5281/zenodo.20469389
A third addendum has now been published defining the Scalability and Deployment Architecture (SDA) for the framework. It addresses the transition from analytical construct to production-grade infrastructure: the σ-stream for real-time Prompt Intentionality Scalar computation, distributed Phase-Shift Smoothing for long sessions, a three-tier Automated Adjudication Architecture (edge / regional / central governance), and CPV compression for distributed ledger anchoring where immutable provenance is required. The document also outlines a federated evaluation protocol and a regulatory roadmap aligned with EU AI Act Articles 50 and 13, with proposed submission to the EU AI Office sandbox programme. Valente, S. (2026). Addendum 3 to the Fabric Test: Scalability and Large-Scale Deployment Architecture. Zenodo. https://doi.org/10.5281/zenodo.20476678
In the end, the question of cognitive provenance can no longer be treated as an optional attribute of content nor as a matter of subjective self‑disclosure. High‑density human–AI co‑ideation dissolves the phenomenological boundary between human and model contribution: the seam is not perceptible, the synthetic fibre does not announce itself, and origin is not introspectively accessible. Governance therefore cannot ask whether a piece of content was generated by AI; it must ask what synthetic density permeates the thought process that produced it. The Fabric Test, through the Synthetic Thread Index, provides a proportional declaration of ideational composition; the Cognitive Provenance Vector situates that declaration within a broader governance state; and the Kernel of Non‑Elusion Cognitiva ensures that provenance cannot be obscured, retrofitted, or disclaimed. Together with the Prompt Intentionality Scalar, the Phase‑Shift Smoothing Operator, and the Automated Adjudication Architecture, these components establish a non‑bypassable epistemic infrastructure: a system in which cognitive transparency is not a voluntary act but a structural property of the ideational pipeline. In this model, provenance is not a judgement of quality but a declaration of material composition — the intellectual equivalent of fibre labelling — and the foundation upon which any credible regime of AI governance must now be built.
This contribution was developed with AI assistance. The conceptual framework, regulatory analogy, and theoretical architecture originated in human ideation.
- Prihlásiť sa na účely uverejňovania komentárov