How a legal risk becomes a channel name

Public Register2024-2025

Meta counsel / litigation posture

litigation position · Kadrey v. Meta court filings

“Training on copyrighted material is fair use. Meta took precautions not to seed downloaded files and complied with lawful principles of transformative use.”

Private Artifact2023

Meta Llama staff, memo escalated to Mark Zuckerberg

internal approval memo · WorkChat / approval memo

“A dataset we know to be pirated… in no case would we disclose publicly that we had trained on libgen.”

The Split:The public register is fair use and rigorous engineering precautions. The private register acknowledges a known pirated library and issues a firm internal directive never to let the public know. The concealment instruction is the preference falsification.

The Tell:When a memo discusses disclosure strategy before it discusses the work itself, the audience has already split.

Pipe: courtUnsealed in Kadrey v. Meta discovery briefing.

Step 2 of 5rename

Once concealment is chosen as policy, the language must be sanitized. Plain nouns like 'pirate' become sterile enterprise terms like 'external corpus' to prevent keyword matches in litigation.

renameai·2023 / unsealed 2024

Rename 'pirate' to 'external' in all documentation

Public RegisterJuly 2023

Meta AI Whitepaper & Comms

public research team · Llama 2 Technical Overview

“Our training data was sourced from a variety of publicly available online sources, filtered to exclude personal data and prioritize high-quality text.”

Private Artifact2023

Meta Research Manager

engineering director · WorkChat thread

“Let's make sure we scrub references to pirate sites from internal tracking spreadsheets and documentation. Refer to these batches as 'external corpus' or 'public dataset ingest' only.”

The Split:Public whitepapers refer benignly to 'publicly available online sources'. Privately, managers ordered staff to purge words like 'pirate' and replace them with sterilizing enterprise synonyms like 'external corpus'.

The Tell:Look for shifts from descriptive nouns ('pirated book database') to abstract bureaucratic categories ('external data asset').

Pipe: courtUnsealed in Kadrey v. Meta discovery exhibits.

Step 3 of 5rename

OpenAI makes the identical structural move. When legal risk tightens, the Slack channel named #excise-libgen is hastily renamed #project-clear.

renameai·2023 / unsealed 2024

Rename #excise-libgen to #project-clear

Public RegisterDecember 2023

OpenAI Spokesperson

official press statement · Statement to Ars Technica

“We support creators and respect copyright laws. Our models are trained on publicly available data, fair use materials, and through commercial partnerships.”

Private Artifact2023

OpenAI Infrastructure Lead

engineering team · Internal Slack

“Renaming this channel from #excise-libgen to #project-clear. We need to be careful about keywords that could trigger subpoena matches.”

The Split:In public, OpenAI affirmed copyright respect and partnership. In private Slack, engineers actively renamed a channel explicitly titled '#excise-libgen' to the neutral alias '#project-clear' specifically to evade subpoena discovery search terms.

The Tell:When an internal channel changes name right before litigation or during data cleanup, the new name is a concealment wrapper.

Pipe: courtUnsealed in Authors Guild v. OpenAI (S.D.N.Y. ECF 782).

Step 4 of 5euphemism

In academic publications, the pirated material is given an innocuous placeholder label—Books1 and Books2—to create the illusion of legitimate proprietary datasets.

euphemismai·2020

Books1 and Books2 in the GPT-3 paper

Public RegisterJune 2020

Tom Brown, Benjamin Mann et al.

OpenAI Research Authors · Language Models are Few-Shot Learners (GPT-3 Paper)

“Books1 and Books2 are two internet-based book corpora. Together they constitute 16% of the pre-training dataset by token count.”

Private Artifact2020

OpenAI Training Data Manifest

internal data catalog · Internal Data Catalog / GitHub commit log

“Books2: 294,000 titles, sourced directly from Library Genesis / Bibliotik torrent dumps, de-duplicated and filtered.”

The Split:The academic paper invented the labels 'Books1' and 'Books2', leaving the impression of licensed or benign corpora. Internal manifests showed Books2 was an exact ingest of shadow library torrents totaling nearly 300,000 pirate-scanned books.

The Tell:Numbered dataset names (Corpus1, Corpus2) with zero publisher citations are almost always masks for unauthorized scrapes.

Pipe: courtAuthors Guild discovery disclosures and investigative reporting.

Step 5 of 5values-after

When the messages eventually leak, the final move is universal across industries: 'These communications do not reflect who we are.' After-the-fact values language is a second register of corporate speech, not a confession.

values-afteraviation·2020

These communications do not reflect who we are

Public RegisterJanuary 9, 2020

The Boeing Company

official corporate response statement · Press Release on Release of Employee Communications

“These communications do not reflect the company we are and need to be, and are completely unacceptable. We have taken appropriate steps to hold employees accountable and reinforce our safety culture.”

Private Artifact2015-2018

Boeing Senior Management Memos

internal executive directives · Internal Management Review Memos

“Any request for additional flight simulator hours will trigger executive escalation. The commercial agreement with Southwest guarantees $1 million per plane if simulator training is mandated. Protect the schedule.”