---
id: "concept-model-retraining-removal"
type: "concept"
source_timestamps: ["§ Lessons for Rightsholders", "¶9"]
tags: ["llm-architecture", "data-management", "ip-protection"]
related: ["contrarian-data-removal-possible", "action-demand-retrain-removal", "prereq-llm-training-lifecycle"]
definition: "The technical window of opportunity to remove copyrighted material from an LLM's corpus when developers train a major new version from scratch."
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?"
---
# Model Retraining Data Extraction

A common misconception is that once copyrighted content is ingested into an LLM's training corpus it is permanently "baked in" and impossible to extract (see the contrarian correction in [[contrarian-data-removal-possible]]). The authors point out that when AI companies release major new versions of their models (e.g., ChatGPT 3 → ChatGPT 4), they do not merely update the old model; they typically **retrain the new model entirely from scratch** on a newly compiled full training corpus. Understanding the difference between fine-tuning an existing model and training a base model from scratch is the prerequisite here — see [[prereq-llm-training-lifecycle]].

This architectural reality opens a **strategic window**: during a from-scratch retrain it is technically feasible for the AI company to leave a specific rightsholder's material out of the corpus, or for a rightsholder to obtain and enforce a court order compelling removal. The corresponding play is [[action-demand-retrain-removal]].

**Enrichment counterpoint (carry forward):** Some ML researchers argue neural representations are highly *entangled*, so cleanly excluding a specific work's influence can be difficult even during retraining; the adjacent field of **machine unlearning** offers partial, not guaranteed, solutions at scale. Courts may eventually have to define what counts as adequate removal or destruction once a model has already learned from infringing data. Notably, remedies already observed (e.g., destruction of pirated libraries and derivative copies in the Anthropic settlement) show courts *can* order data destruction — so the strategy is grounded, but treat the "neat retrain window" as strategic guidance rather than a settled, frictionless standard.