Statistical watermarking for AI-generated text, as described by Kirchenbauer et al. [2] and extended by Zhao et al. [3], biases token selection at generation time to embed a statistically detectable signal. The EU AI Act Article 50 [1] requires that outputs of AI systems capable of generating text be identifiable as AI-generated. Watermarking satisfies this requirement at scale. Its binary output, content either passes detection or it does not, is well-matched to the regulatory intent: disclosure that a text was produced by an AI system.

It is structurally mismatched to a distinct governance requirement that enterprises face simultaneously: establishing and demonstrating the degree of human intellectual involvement in AI-assisted content. Consider two professionals. The first receives a client brief, prompts an AI to draft a memorandum, reviews it briefly, and signs it. The second spends hours developing an original analytical framework, uses AI to test alternative formulations, discards most, substantially rewrites the rest, and publishes a synthesis that bears little textual resemblance to any AI output. Both documents may carry the same watermark state B=1. Their intellectual origins are structurally different. The watermark is silent on the distinction.

The NIST AI Risk Management Framework [4] addresses risk identification and mitigation but provides no constructs for quantifying human intellectual contribution. ISO/IEC 42001 [5] treats content provenance as a documentation obligation, not a measurable property. The EU AI Act explicitly limits Article 50 to identification of AI-generated content, not characterization of human contribution within it. Academic watermarking literature [2],[3] addresses technical properties of embedding and detection, not authorship attribution. WIPO [6] identifies measurement of human contribution as a pending policy problem. The US Copyright Office [7] treats degree of human control as a case-by-case factual question without a formal instrument.

This paper introduces the Contribution Ratio Framework (CRF):

• The Contribution Ratio (CR), a formal function in [0,1] defined over three authorship phases (Section III).
• IEJ Decomposition, partitioning collaborative content work into Ideation, Expression, and Editorial Judgment phases (Section IV).
• The Provenance Collapse Property, formally characterizing information loss when CR is encoded as a binary watermark signal (Section V).
• A four-level CR Maturity Model with concrete organizational criteria (Section VI).
• Enterprise implementation guidance with a policy template and three industry scenarios (Sections VII and VIII).

Kirchenbauer et al. [2] partition the token vocabulary into green-list and red-list subsets using a hash of preceding context, then bias the model toward green-list tokens at generation. The resulting statistical bias is detectable with knowledge of the partitioning key and is imperceptible under normal reading conditions. Zhao et al. [3] extend this with cryptographic robustness guarantees, demonstrating resistance to paraphrasing attacks under specified conditions. Both bodies of work define the watermark as a binary property: a text either passes the statistical detection threshold or it does not.

The US Copyright Office's March 2023 guidance established that copyright requires human authorship and that AI-generated content without human creative control does not qualify [7]. WIPO's 2023 analysis [6] finds that human authorship is the near-universal threshold criterion across member states, with "sufficient" human contribution unresolved across jurisdictions. WIPO identifies this measurement gap as a pending policy problem. The CRF is designed to fill it.

No existing framework provides: (1) a formal definition of human intellectual contribution in collaborative human-AI content; (2) a measurable decomposition of that contribution into auditable components; or (3) a characterization of the relationship between binary watermark signals and continuous authorship attribution. The CRF addresses all three.

Let A denote a content artifact produced through human-AI collaboration. Let P = {I, E, J} denote the set of authorship phases (defined in Section IV). For each phase p in P, let h(p, A) in [0,1] denote the human intellectual contribution to that phase, where 1 represents entirely human and 0 represents entirely AI. Let w = (w_I, w_E, w_J) satisfy w_I + w_E + w_J = 1 and w_I ≥ w_E ≥ w_J > 0 (strict positivity ensures each phase contributes to CR for non-empty P).

The ordering constraint w_I ≥ w_E ≥ w_J encodes the authorship principle that Ideation carries the highest weight. This reflects the legal and philosophical consensus that originating a thesis or framework is a more fundamental authorship act than selecting wording, which is itself more fundamental than copy-editing decisions. Organizations select their specific weighting vector subject to this constraint.

When a human provides only a prompt and accepts AI output unchanged: h(I, A) may be positive (topic specification contributes to Ideation), but h(E, A) = 0 and h(J, A) = 0. CR(A) is bounded above by w_I, the Ideation weight alone, regardless of prompt sophistication.

When a human writes a complete draft and uses AI only for grammar correction, accepting all corrections: h(I, A) = 1, h(E, A) = 1, h(J, A) approaches 1. CR(A) approaches 1. The binary watermark assigns B = 1 to both cases. The CR distinguishes them; the watermark cannot.

The Contribution Ratio is computed over three authorship phases. IEJ Decomposition provides the operational vocabulary for assigning h(p, A) in each phase and for specifying policy requirements at the phase level.

The IEJ phases are auditable. An organization can specify phase-level policy requirements: for example, "for client-facing analysis signed by a named advisor, h(I, A) must be at least 0.8," meaning the advisor must have originated at least 80% of substantive analytical claims before AI involvement in structuring or drafting. This standard is more precise and legally defensible than a binary disclosure policy.

The three phases are non-overlapping but not temporally sequential. A human may originate an idea (Ideation), ask AI to draft it (reducing h(E, A)), then substantially rewrite the draft (restoring h(E, A)) before making final editorial decisions (J). The phases describe categories of intellectual act, not workflow steps. An audit process can reconstruct IEJ contributions from process documentation even when phases interleave across multiple sessions.

Statistical watermarking systems produce a binary output. Detection applied to a content artifact returns B = 1 if the watermark pattern is detected above a statistical threshold, and B = 0 otherwise [2],[3]. This binary representation satisfies its regulatory purpose: EU AI Act Article 50 requires that AI-generated content be identifiable, a binary property [1]. It is structurally mismatched to authorship attribution, which requires a continuous measure.

The Provenance Collapse Property establishes a structural distinction between two governance instruments that organizations currently conflate. An enterprise complying with EU AI Act Article 50 by using a watermarked AI system has satisfied the binary provenance disclosure requirement. It has not, by that fact alone, established any authorship standard. The two obligations require separate policy instruments: the watermark satisfies disclosure; the CR policy satisfies authorship accountability.

First, professional services firms whose deliverables are signed by named experts face client and regulatory expectations about the intellectual origin of those deliverables that the watermark cannot address. Second, regulated industries in which AI-assisted documents enter compliance, legal, or clinical processes require documentation of human intellectual involvement that the binary signal cannot provide. Third, organizations publishing under named authors' brands face reputational exposure proportional to the gap between the implied CR (high, because a named expert signed) and the actual CR (potentially low, if the content was substantially AI-generated).

A watermark asserts: "this text was generated by an AI system of a specified class." It does not assert the degree of human intellectual contribution to the ideas expressed, the identity or expertise of any human reviewer, whether the human substantively modified or endorsed the content, or whether the content meets any professional, legal, or institutional authorship standard. Organizations treating watermark compliance as equivalent to authorship accountability conflate two structurally distinct properties.

The CR Maturity Model provides a four-level staged adoption path. Each level builds on the prior and is defined by concrete policy criteria, governance capability, and organizational indicators.

AI is used in content creation with no explicit authorship policy. Use is ad hoc and undocumented. When asked "what was the human contribution to this deliverable?", the organization cannot answer systematically. Watermark compliance may exist if using a watermarked model, but is not connected to any authorship governance instrument. The organization receives the same binary watermark output for a human-originated analysis and a fully AI-generated draft: no internal process distinguishes them.

The organization has a policy requiring disclosure that AI was used. The disclosure is binary: AI was or was not involved. Named authors must acknowledge AI assistance. This level satisfies the minimum EU AI Act Article 50 transparency intent at the organizational policy layer. It does not define what degree of AI involvement requires disclosure, does not distinguish AI-generated from AI-assisted content, and does not assign a CR to any specific artifact.

The organization has adopted IEJ Decomposition vocabulary and requires authors to document, at creation time, which phases involved AI assistance. Documentation may be qualitative: "ideation was entirely human; expression was AI-drafted and substantially revised; editorial judgment was fully human." Level 2 gives the organization vocabulary to distinguish the passive-review scenario from the active-origination scenario. It does not yet produce a numeric CR but creates the audit foundation for one.

The organization has defined minimum CR thresholds for each content type and enforces them through policy. Authors document phase-level contributions at creation time, retaining documentation as an audit record. Content below threshold requires additional human revision before publication. Level 3 organizations can respond to a client or regulatory authorship audit with a documented, specific answer: "the Ideation contribution was 0.85, computed as..." This constitutes a qualitative advance beyond any governance level currently specified in existing standards.

Organizations designing a CR policy face three primary decisions. First, content classification: which artifact types require CR governance? A risk-proportionate approach applies the CRF to content where authorship has legal, regulatory, or reputational consequence: signed client deliverables, regulatory submissions, published research, and contractual documents. Internal operational content (scheduling communications, routine summaries) typically does not require CR governance.

Second, threshold setting: what is the minimum CR for each classified content type? The threshold should be calibrated to the governing obligation (copyright eligibility, professional standards, client commitments) rather than set arbitrarily. Legal services firms may require CR(A) ≥ 0.70 for partner-signed client advice. Research organizations may require h(I, A) ≥ 0.80 regardless of overall CR.

Third, weighting vector selection: how should the organization weight the I, E, and J phases? The ordering constraint is fixed by the CRF; specific values are determined by governance context. A conservative default is (w_I = 0.55, w_E = 0.30, w_J = 0.15), weighting ideation heavily. Organizations with strong editorial processes may use values closer to (0.45, 0.35, 0.20).

A Level 2 implementation requires no automated tools. The minimum viable approach uses a structured author declaration completed at content creation. The declaration records: the primary source of the thesis and argument structure (I); the proportion of final published text that was human-written versus AI-generated and retained without substantive modification (E); and the editorial decisions made regarding inclusion and exclusion of AI-generated material (J). This documentation creates an auditable CR basis.

A Level 3 CR policy implementation requires: one Policy Lead responsible for defining CR thresholds by content type and the organizational weighting vector (typically a senior legal, compliance, or governance professional); one Process Owner integrating CR documentation into content workflows and training authors on IEJ vocabulary; one Audit Coordinator retaining CR documentation records and coordinating audit responses. External legal counsel review of the CR policy is advisable before client-facing adoption. Automated CR scoring is a Phase 2 capability, following manual Level 2 operation that validates the policy design.

A mid-size law firm deploys a generative AI system for client advisory work. Partners use it to draft memoranda on complex regulatory questions. A corporate client requests the firm's authorship policy following a governance audit: specifically, what proportion of the legal analysis was originated by the advising partner versus generated by the AI system.

Without a CR policy, the firm cannot answer. The watermark on the document confirms AI involvement but provides no information about the partner's intellectual contribution to the legal analysis. The firm's liability exposure depends on whether the partner can demonstrate professional judgment at the Ideation phase: did the partner originate the legal framework for the advice, or accept an AI-generated analysis?

An investment research firm publishes equity analysis reports under named analyst signatures. Analysts use AI to generate initial sector summaries used as background reading before forming their own investment thesis. A regulatory inquiry under MiFID II analytical independence provisions requires the firm to demonstrate that investment recommendations represent the analyst's independent professional judgment, not AI-generated conclusions.

A healthcare organization uses an AI system to assist physicians in drafting clinical notes from voice recordings and EHR data. A payer audit requires the organization to confirm that clinical decisions documented in the notes represent the physician's medical judgment. The EU AI Act classifies AI systems used in clinical decision support as high-risk, requiring documentation that AI output was subject to human oversight with ability to override or modify [1]. The watermark confirms AI involvement but does not document the physician's degree of oversight at the clinical judgment phase.

The CRF complements existing regulatory obligations without requiring their modification. Organizations subject to EU AI Act Article 50 should continue using watermarked AI systems for disclosure compliance; the CRF addresses the adjacent obligation that Article 50 explicitly does not cover. The two instruments together constitute a complete authorship governance response: watermark compliance satisfies the disclosure obligation; CR policy satisfies the authorship accountability obligation. NIST AI RMF Govern and Map functions [4] can accommodate the CRF as a supplementary construct, and ISO/IEC 42001 [5] management systems can integrate CR documentation as an operational record. Neither requires modification.

Three limitations should be addressed in future work. First, h(p, A) measurement is currently self-reported; automated measurement of human intellectual contribution remains an open research problem. Second, the weighting vector w is specified by policy rather than derived from a principled theory of authorship, creating potential inconsistencies across governance contexts. Third, the CRF defines CR as a property of a completed artifact and does not model dynamic multi-session collaborative workflows.

Future research directions include: automated IEJ phase detection; principled derivation of weighting vectors from professional authorship standards; extension to multi-author artifacts; and empirical calibration of CR threshold values against copyright eligibility determinations in specific jurisdictions.

The Provenance Collapse Property is a formal result derived from the binary nature of statistical watermarking as defined in [2] and [3]. It holds for any binary provenance signal regardless of detection performance, distinguishing it from empirical claims about watermark robustness. Prior work has noted qualitatively that binary disclosure is insufficient for collaborative authorship attribution; this paper provides the first formal characterization of the information loss and the first structured measurement framework as a response.

This paper introduced the Contribution Ratio Framework as a formal response to a gap that statistical watermarking, the EU AI Act Article 50, and existing AI governance standards leave unaddressed: the continuous measurement of human intellectual contribution to AI-assisted content. The three formal contributions are: the Contribution Ratio (CR), a function in [0,1] over the IEJ Decomposition phases; IEJ Decomposition, providing operational vocabulary for authorship policy specification and audit; and the Provenance Collapse Property, establishing that binary watermark signals and continuous authorship attribution are structurally distinct and require separate governance instruments.

The practical implication is immediate: watermark compliance satisfies the EU AI Act disclosure obligation but does not constitute an authorship accountability policy. Organizations that have implemented one have not implemented the other. Three enterprise scenarios in legal services, financial services, and regulated healthcare demonstrate that IEJ Decomposition and CR threshold policy produce governance outcomes that watermark-only disclosure cannot: specific, phase-level characterization of human intellectual contribution that can withstand regulatory and client audit.

The CR Maturity Model provides a four-level path from undifferentiated AI use to quantified, auditable authorship standards. A Level 2 implementation requires no specialized tools. A Level 3 implementation requires a defined policy, a documentation workflow, and audit coordination, all within reach of governance and compliance teams without technical AI expertise. The Contribution Ratio, IEJ Decomposition, and Provenance Collapse Property are offered as foundational vocabulary for what will become a necessary governance domain as AI involvement in professional content creation becomes pervasive, regulated, and consequential.

© 2026 Arjun Jaggi and Aditya Karnam Gururaj Rao. All rights reserved. Academic citation permitted with attribution; commercial use and derivative frameworks require written permission. The terms "Contribution Ratio," "IEJ Decomposition," and "Provenance Collapse Property" originate with this work and are subject to the above license.