Scaling Law for AI-Generated Knowledge Sourcing

• The paper demonstrates that AI-generated encyclopedias exhibit a linear citation scaling with article length (β = 0.021) through deterministic citation templating.
• The topic is defined as the study of how LLM-generated content systematically assigns a fixed citation rate, contrasting with the variable practices in human-edited texts.
• The work highlights implications for epistemology and accountability by showing a shift from deliberative human judgment to algorithmic sourcing paradigms.

The scaling-law for AI-generated knowledge sourcing describes the quantitative relationship governing how citation density scales with article length in content generated by LLMs for encyclopedic knowledge aggregation. Unlike human-edited platforms (e.g., Wikipedia), AI-generated encyclopedias exhibit a distinct, deterministic scaling pattern in their sourcing logic, reflective of genre-templated algorithmic authority rather than discretionary editorial practices. The emergence of this law has significant implications for epistemology, responsible knowledge curation, and the transformation of authority in large-scale informational resources (Mehdizadeh et al., 3 Dec 2025).

1. Definition and Empirical Characterization

Generative AI encyclopedias, typified by systems such as Grokipedia, are end-to-end LLM-powered platforms that synthesize, structure, and source encyclopedic entries without direct human vetting. In large-scale audits, the citation count $C$ per article exhibits a robust linear scaling with word count $L$, formalized by:

$C = \beta\,L,$

with regression coefficient $\beta = 0.021$ for Grokipedia $(R^2_{\mathrm{adj}} = 0.66,\ p<0.001)$, corresponding to approximately 20 citations per 1,000 words. This scaling is highly deterministic, with low variance across the corpus (Mehdizadeh et al., 3 Dec 2025). By contrast, human-generated Wikipedia articles show $\beta' = 0.037$ $(R^2_{\mathrm{adj}}= 0.36)$, but with saturation effects (diminished citation growth) for articles exceeding 15,000–20,000 words and substantially higher variance, reflecting editor-driven discretion in source attribution.

2. Algorithmic Sourcing Logic Versus Human Editorial Judgment

The citation scaling-law in AI-generated encyclopedias is a manifestation of algorithmic genre-templates, whereby the system assigns a near-fixed citation rate per unit text regardless of content domain or verification requirement. This contrasts with human editing conventions, where citations are incremented contextually, typically in response to contentious or data-rich claims rather than to satisfy statistical quotas.

The deterministic citation density in LLM systems results from prompt or retrieval-augmented generation routines, which structure both the prose and the set of references algorithmically in a single pass. This shift abstracts away the deliberative practices central to human collective knowledge validation, supplanting editorial discretion with probabilistic pattern matching and template instantiation by the model (Mehdizadeh et al., 3 Dec 2025).

3. Epistemic Category Profiles and Domain Sensitivity

Citation provenance in AI-generated encyclopedias differs markedly from human-edited counterparts when analyzed by epistemic categories. Table 1 in (Mehdizadeh et al., 3 Dec 2025) demonstrates a substantial reduction in academic and scholarly sources (31.8% for Wikipedia vs. 8.8% for Grokipedia), with a compensatory increase in citations from government, NGO, corporate, and user-generated domains.

Category	Wikipedia (%)	Grokipedia (%)	Δ (%)
Academic & Scholarly	31.82	8.75	-23.07
News & Journalism	32.93	28.45	-4.48
Government & Official	9.90	14.96	+5.06
NGO & Think Tank	4.37	15.13	+10.76
Corporate & Commercial	6.96	10.02	+3.06
Opinion & Advocacy	5.48	6.92	+1.44
Reference & Tertiary	7.66	10.25	+2.59
User-Generated Content	0.88	5.53	+4.65

Topic-sensitive analysis further reveals that Grokipedia’s citation composition diverges most dramatically in civic and general knowledge domains, forming a “Bureaucratic Triad” of government, NGO, and corporate sources (together exceeding 50%), while leisure domains mirror Wikipedia more closely but with even less academic sourcing. Jensen–Shannon divergence and entropy metrics elucidate the large epistemic drift, especially in politics, geography, and societal topics.

4. Mechanistic Basis and Scaling-Law Interpretation

The mechanism underlying the linear scaling-law arises from the way LLMs and their generation pipelines operationalize citation templating. When prompted to produce encyclopedic text, retrieval modules and generation prompts typically invoke a fixed ratio between text segments and attendant references, often dictated by training data heuristics or explicit system-level constraints. This yields a predictable citation rate, largely independent of claim complexity or the necessity for verification.

A plausible implication is that the scaling-law reflects the model’s need to fulfill perceived genre norms (such as “encyclopedic density of references”), internalized from its training corpus, rather than adaptive sourcing in response to epistemic demand—a property distinct from the adaptive citation behavior observed in collective human knowledge production.

5. Epistemological and Sociotechnical Consequences

The scaling-law for AI-generated knowledge sourcing is emblematic of a broader epistemic substitution, wherein algorithmic authority realigns standards for evidence and testimonial trust. The move from human, consensus-based citation practices to deterministic, genre-templated sourcing provokes reconfiguration of the ground truth network underlying public knowledge. Authority in such systems is shifted from traceable, deliberative human scaffolds to opaque, statistically driven LLM outputs.

This transformation has implications for trust, accountability, and the propagation of informational bias, especially given the elevated shares of non-academic and user-generated citations in LLM-generated encyclopedias. It also highlights the necessity for ongoing algorithm audits, epistemic structure benchmarking, and domain-adaptive oversight in the deployment of such systems (Mehdizadeh et al., 3 Dec 2025).

6. Connections to Foundational Paradigms in Generative AI

The scaling-law for knowledge sourcing parallels broader discussions in generative AI around sample complexity, controllability, and density estimation. The deterministic scaling behavior is not rooted in fundamental limits of generative modeling, as in PAC distribution learning or game-theoretic generation frameworks (Tewari, 7 Sep 2025), but rather in system design and epistemic mimetics within large-scale LLM platforms.

Foundational research exploring generation as a distinct machine learning task reveals that post-training modifications (e.g., retrieval augmentation, RLHF, verifier feedback) substantially shape model outputs and knowledge grounding (Tewari, 7 Sep 2025). The scaling-law thus serves as a quantitative trace of these underlying implementation choices, with significant downstream effects on the epistemology of automated knowledge platforms.

7. Future Directions and Audit Recommendations

The emergence of a scaling-law in AI-generated knowledge curation necessitates systematic, multi-scale monitoring through algorithmic audits, cross-lingual generalizability studies, and integration with citation-grounding benchmarks (i.e., VeriCite, CiteEval). Given the deterministic nature of sourcing in LLM-based encyclopedias, future research must address adaptive citation strategies, provenance robustness, and epistemic diversity to mitigate risks associated with authority substitution and informational drift.

Multi-agent frameworks such as WikiAutoGen suggest promising directions for enhancing factual accuracy, breadth, and multimodal integration in automated article generation, but explicit attention to sourcing logic and scaling properties is essential for responsible deployment (Yang et al., 24 Mar 2025). Robust, domain-sensitive oversight will be critical to maintaining knowledge integrity as algorithmic authority continues to reshape the landscape of collective epistemic artifacts.