---
id: "claim-unlicensed-data-performance"
type: "claim"
source_timestamps: ["§ Lessons for Gen AI Companies", "¶17"]
tags: ["ai-performance", "training-data", "open-source"]
related: ["entity-eleuther-ai", "contrarian-unlicensed-data-unnecessary", "question-unlicensed-data-necessity", "quote-eleuther-performance"]
confidence: "medium"
testable: true
sources: ["tail2"]
sourceVaultSlug: "hbr-seg-tail2"
originDay: 2
articleStem: "hbr-tail-126-genai-copyright"
sourceUrl: "https://hbr.org/2025/07/can-gen-ai-and-copyright-coexist"
sourceTitle: "Can Gen AI and Copyright Coexist?"
---
# Unlicensed Data May Not Be Necessary for AI Performance

**Claim (confidence: MEDIUM — experimental and not independently confirmed).**

Citing research from [[entity-eleuther-ai]], the authors argue that the prevailing industry assumption — that massive amounts of unlicensed copyrighted text are strictly necessary for frontier LLM performance — may be unjustified. EleutherAI released **Common Pile v0.1**, an 8 TB dataset composed entirely of open-source or licensed content, and reported that models trained on it performed **just as well** as models trained on unlicensed copyrighted data (see [[quote-eleuther-performance]]). If correct, the marginal benefit of scraping unlicensed data does not justify the legal and financial risk quantified in [[claim-piracy-financial-risk]]. This is the empirical backbone of the contrarian thesis in [[contrarian-unlicensed-data-unnecessary]].

**Enrichment calibration:** The existence and purpose of Common Pile (a license-clean dataset from a credible open-source lab known for The Pile and GPT-Neo(-X)) is accurate. The **performance-equivalence** claim, however, rests largely on EleutherAI's own preliminary experiments rather than independent, peer-reviewed benchmarking. A domain expert would demand careful benchmark comparisons across tasks and model scales before accepting parity as established — see the open question [[question-unlicensed-data-necessity]]. Treat parity as a plausible, actionable hypothesis, not a proven result.