AI Contribution Ratio: Metrics & Insights

• AI Contribution Ratio is a family of domain-specific attribution measures quantifying the proportional impact of AI relative to human or system performance.
• It encompasses diverse formulations ranging from performance increments in human–AI collaboration to information-theoretic and operational ratios used in auditing and economic analysis.
• Different sectors apply tailored AI Contribution Ratios—each with distinct methodologies and limitations—to evaluate AI’s role in content generation, decision-making, and productivity.

AI Contribution Ratio denotes a family of attribution measures that estimate how much of an observed result is attributable to artificial intelligence rather than to a human, another agent, or a broader production system. The literature does not treat it as a single standardized quantity. Instead, closely related formulations appear in human–AI collaboration, AI-assisted content generation, macroeconomic analysis, autonomy auditing, cooperative multiagent systems, blockchain consensus, educational assessment, and sectoral studies such as translation. This suggests that the term functions less as a universal metric than as a domain-specific class of ratios whose numerator identifies an AI-attributable increment and whose denominator identifies either final output, total system activity, or total contribution capacity (Ganuthula et al., 13 Feb 2025, Xie et al., 2024, Occhipinti et al., 2024, Mairittha et al., 12 Dec 2025).

1. Conceptual scope and principal formulations

Several arXiv works define or motivate ratios that can be read as AI Contribution Ratios, even when the phrase itself is absent. In collaborative assessment, the ratio is the marginal gain from AI assistance relative to final collaborative performance. In AI-assisted text generation, it is the residual information in the output not explained by human input. In autonomy auditing, it is the fraction of decisions executed by AI without mandatory human substitution. In macroeconomic analysis, it is the relative weight of AI capital against labour (Ganuthula et al., 13 Feb 2025, Xie et al., 2024, Mairittha et al., 12 Dec 2025, Occhipinti et al., 2024).

Context	Representative expression	What is apportioned
Human–AI collaboration	$C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$	Final task performance
AI-assisted content generation	$$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$	Output information content
AI autonomy auditing	$$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$	Operational decision load
Macroeconomic production	$\theta=\frac{K_{AI}}{L}$	AI capital relative to labour

These formulations are not commensurable. $C_{AI}$ is task-relative and can become negative; $\phi_{\text{AI}}$ is information-theoretic and output-centric; $\alpha$ is operational and architectural; $\theta$ is a factor ratio in production. A plausible implication is that any use of “AI Contribution Ratio” requires explicit specification of unit of analysis, denominator, and attribution rule before cross-study comparison is meaningful.

2. Human–AI collaboration and authorship-based ratios

One major line of work treats AI contribution as the increment obtained when a human uses AI relative to a human-only baseline. The AIQ framework defines collaborative intelligence as a human capacity to leverage AI and organizes it into eight dimensions: Strategic AI Understanding, Prompt Engineering, Critical Evaluation, Integration Intelligence, Adaptive Learning, Ethical Judgment, Context Sensitivity, and Creative Synthesis. Although the paper does not provide a formal ratio, its assessment structure supports an AI Contribution Ratio defined as $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$, where $P_H$ is human-only performance and $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$0 is human-plus-AI performance. The same framework supports task-level and dimension-conditioned variants over multiple tasks and weighted aggregates (Ganuthula et al., 13 Feb 2025).

In that formulation, AI contribution is not identical to AI standalone capability. It is “AI’s added value as leveraged by a particular human.” The distinction matters because the same tool can yield very different $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$1 values depending on prompting, evaluation, and integration skill. The framework also allows negative values: if $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$2, then $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$3, which the synthesis interprets as harmful AI contribution associated with automation bias or uncritical trust (Ganuthula et al., 13 Feb 2025).

A second line of work formalizes contribution in information-theoretic terms. For AI-assisted content generation, the human contribution ratio is defined as $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$4, where $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$5 is the self-information of the output and $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$6 is its conditional self-information given human input. The complementary AI contribution ratio is therefore $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$7 (Xie et al., 2024).

That metric was evaluated on paper abstracts, news, patent abstracts, and poems using polishing, summary, title, and subject conditions. It discriminated systematically across those conditions: in the news domain with Llama-3, average human contribution was about $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$8 for polishing and about $$\phi_{\text{AI}}=\frac{I(\boldsymbol{y}\mid \boldsymbol{x})}{I(\boldsymbol{y})}$$9 for generation from subject-only input. The same study reports that with Llama-3, generation from summary produced average human contribution of about $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$0 for paper abstracts and about $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$1 for news articles, showing that the ratio depends on the information density of the prompt relative to the generated output (Xie et al., 2024).

Educational process tracing provides a third authorship-oriented formulation. NIRVANA defines a Human Contribution Ratio, $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$2, from human additions $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$3, deletions of human text $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$4, pasted ChatGPT words $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$5, and deletions of pasted ChatGPT text $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$6. A complementary word-level AI contribution ratio is therefore $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$7. The same dataset defines a Human Edit Ratio, $$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$8, which separates final-word provenance from process effort. On that basis, four writing profiles were identified: Lead Authors ($$\alpha=\frac{\text{Number of Decisions Made by AI Alone}}{\text{Total Number of Decisions}}$$9, $\theta=\frac{K_{AI}}{L}$0), Collaborators ($\theta=\frac{K_{AI}}{L}$1, $\theta=\frac{K_{AI}}{L}$2), Drafters ($\theta=\frac{K_{AI}}{L}$3, $\theta=\frac{K_{AI}}{L}$4), and Vibe Writers ($\theta=\frac{K_{AI}}{L}$5, $\theta=\frac{K_{AI}}{L}$6) (Jelson et al., 8 Apr 2026).

Taken together, these works distinguish at least three different attribution targets: final performance, final information content, and observable writing process. They also show that a high AI share in final wording need not imply low human effort, and that a high human share in final wording can still coexist with substantial idea-level AI influence.

3. Autonomy, oversight, and the boundary between assistance and substitution

A separate literature treats AI contribution as operational autonomy. In the AFHE framework, a system $\theta=\frac{K_{AI}}{L}$7 processes tasks through an AI module $\theta=\frac{K_{AI}}{L}$8 and a human module $\theta=\frac{K_{AI}}{L}$9 according to $C_{AI}$0. The central metric is the AI Autonomy Coefficient, $C_{AI}$1, with $C_{AI}$2. Here, the numerator includes tasks accepted and executed without mandatory human substitution; asynchronous audit does not reduce $C_{AI}$3 in the same way as per-task operational fallback (Mairittha et al., 12 Dec 2025).

The framework distinguishes legitimate Human-in-the-Loop from Human-Instead-of-AI. Ideal HITL is characterized by a high steady-state $C_{AI}$4, specifically $C_{AI}$5, with human intervention reserved for a low percentage of high-risk or novel cases. By contrast, a system marketed as an AI product is said to satisfy the HISOAI condition if $C_{AI}$6. The paper’s case study reports a legacy system with $C_{AI}$7 and human time cost accounting for over $C_{AI}$8 of operational resource usage; that system is classified as HISOAI. In the successor AFHE system, an example deployment target of $C_{AI}$9 is enforced, and after three re-engineering cycles the system reaches $\phi_{\text{AI}}$0 before clearance for deployment (Mairittha et al., 12 Dec 2025).

This operational perspective redefines contribution as a property of system architecture rather than of a specific output. It also changes the meaning of “human contribution.” Under AFHE, humans are meant to perform ethical oversight, boundary pushing, and strategic tuning, rather than routine fallback labor. A plausible implication is that autonomy-style ratios are best suited to auditing claims of automation and scalability, whereas authorship-style ratios are better suited to output attribution.

4. Economic and organizational interpretations

At the macroeconomic level, the closest analogue to an AI Contribution Ratio is the AI-capital-to-labour ratio, $\phi_{\text{AI}}$1, where $\phi_{\text{AI}}$2 denotes the stock or intensity of generative AI capital and $\phi_{\text{AI}}$3 denotes effective human labour input. The recessionary-pressure framework posits a threshold $\phi_{\text{AI}}$4 beyond which a self-reinforcing cycle of recessionary pressures could be triggered. Below that threshold, AI acts mainly as productivity-augmenting capital; above it, job displacement, wage compression, weak demand, and “productivity overhang” dominate (Occhipinti et al., 2024).

The same paper also introduces an output-based interpretation, $\phi_{\text{AI}}$5, where $\phi_{\text{AI}}$6 is the portion of total output attributable to AI-operated processes or AI-generated labour-equivalent services. However, its preferred framing is factor-based rather than output-based. The key claim is therefore not that AI contribution is inherently beneficial, but that the macroeconomic effect depends on whether AI capital displaces labour beyond the institutional capacity of the economy to absorb the change (Occhipinti et al., 2024).

Broader measurement reports provide additional ratio families without consolidating them into one ACR. In the AI Index 2025, AI’s share of computer science publications rose from $\phi_{\text{AI}}$7 in 2013 to $\phi_{\text{AI}}$8 in 2023. Global private investment in generative AI reached $\phi_{\text{AI}}$9 billion in 2024 and represented more than $\alpha$0 of all AI-related private investment. The share of surveyed organizations using any AI in at least one function rose from $\alpha$1 in 2023 to $\alpha$2 in 2024, while generative AI use rose from $\alpha$3 to $\alpha$4. In the United States, AI-related job postings rose from $\alpha$5 of all job postings in 2023 to $\alpha$6 in 2024 (Maslej et al., 8 Apr 2025).

These are not direct attribution ratios for a single output. Rather, they are ratio-based indicators of AI’s share in research activity, investment, organizational adoption, and labour demand. This suggests that in policy and industry analysis, “AI contribution ratio” often denotes a family of share measures tracking AI’s weight within a larger system rather than its causal contribution to a single result.

5. Cooperative allocation, fairness, and multidimensional contribution

In cooperative multiagent reinforcement learning, contribution is formalized via Shapley values. For a cooperative game $\alpha$7, the Shapley value of agent $\alpha$8 is $\alpha$9. Here, agents are players, coalitions are subsets of active agents, and $\theta$0 is the expected global reward obtained by coalition $\theta$1. The method yields per-agent contribution scores interpretable as average marginal contributions to the team reward. Because exact computation is expensive, the paper uses Monte Carlo approximation; in Predator–Prey experiments, $\theta$2 samples produced approximation errors of about $\theta$3 in one setting and about $\theta$4 in another. In Harvest, one agent’s Shapley value was near zero, and removing that agent left total global reward essentially unchanged at about $\theta$5 reward units (Heuillet et al., 2021).

In blockchain consensus, AICons extends Shapley-based attribution to a multidimensional utility over model accuracy, energy consumption, and network bandwidth. Its utility is $\theta$6, with the energy term defined as $\theta$7 and the bandwidth term as $\theta$8. After normalization and scalar aggregation, the resulting node contribution score is used to distribute rewards. The reported outcome is an evenly distributed reward-contribution ratio across nodes, together with higher throughput: AICons handles $\theta$9 more transactions per second than the state-of-the-art schemes considered in the study (Xiong et al., 2023).

In group workload investigation, AI contribution is again derived from a richer evidential structure. The framework organizes artifacts into three dimensions—Contribution, Interaction, and Role—with nine benchmarks. Objective metrics are normalized and aggregated, inequality is surfaced via the Gini index, and an LLM produces interpretable advisory judgments. The proposed scalar summary is $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$0, with an optional share form $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$1. The paper is explicit that such outputs are advisory rather than binding and must remain human-reviewed (Slapek et al., 10 Nov 2025).

Across these settings, contribution is not inferred from superficial output similarity. It is computed from cooperative marginal value, multidimensional utilities, normalized evidence, or inequality-aware aggregation. A plausible implication is that fairness-oriented ACRs are most mature where the system exposes explicit coalition structure, shared reward, or auditable artifact trails.

6. Sectoral applications, intellectual variants, and persistent limitations

Sector-specific studies show that AI contribution can also be tracked without a single scalar. In the translation industry, AI contribution to translation is operationalized through technological stages, research intensity, and practical impact rather than a universal ratio. The scientometric review retrieved $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$2 records and, after deduplication, analyzed $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$3 unique research articles. It identifies a progression from rule-based machine translation to statistical machine translation, neural machine translation, and LLMs. The review reports, for example, that “machine translation” has centrality $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$4 and “chatgpt” centrality $C_{AI}=\frac{P_{HA}-P_H}{P_{HA}}$5, while also emphasizing persistent difficulties in low-source languages, multi-dialectical and free word order languages, and cultural and religious registers (Shormani, 2024).

Other work uses the language of contribution in an explicitly intellectual rather than numerical sense. The study of generative linguistics and AI argues that Chomskyan generative linguistics has a “huge” contribution to AI through formal grammars, syntax, semantics, Universal Grammar, the computational system of human language, programming languages, and LLM evaluation. In that usage, “contribution ratio” refers to relative theoretical centrality, not to a computed scalar. The paper locates especially strong influence in formal language theory, parsing, automata, programming language design, and the syntactic and semantic phenomena used to evaluate neural LLMs (Shormani, 2024).

The literature also converges on several limitations. AIQ-derived performance ratios are context-dependent, sensitive to tool evolution, and can be distorted by poor task design; higher values are not always better because ethical or contextual appropriateness may require keeping AI use at zero in some settings (Ganuthula et al., 13 Feb 2025). Information-theoretic authorship ratios require access to the actual human input and depend on the evaluator model’s probability distribution; they are not normative legal standards and do not fully capture selection, curation, or multi-turn co-creation (Xie et al., 2024). Educational provenance ratios such as HCR and HER are genre-specific, one-shot, and can miss idea-level AI influence when students consult AI without copying its text (Jelson et al., 8 Apr 2026).

An enduring controversy follows from these limitations. One tradition measures contribution as AI-added performance, another as AI-origin information, another as autonomous task completion, another as factor intensity, and another as cooperative marginal value. These are not interchangeable. The most defensible use of the term therefore treats AI Contribution Ratio as a family resemblance concept: a domain-bounded attribution operator that must declare its object of attribution, evidence model, and normative purpose before its value can be interpreted.