25 minutes

Concealment moves

Start here if you want the rename, the euphemism, and the values statement.

How a legal risk becomes a channel name, then a cleanup sentence. Five pairs. About twenty-five minutes.

Read full-screen

Step 1 of 5
The memo reaches MZ with the disclosure rule already written. That is the first move: decide the audience before you decide the work.
In no case would we disclose publicly
euphemism
Public
Meta counsel / public posture · 2024-2025
litigation position
Kadrey filings and public statements
“Training on copyrighted material is fair use. Meta took precautions not to seed downloaded files.”
Private
Meta Llama staff, memo escalated to MZ · 2023
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.”
Exhibit Meta_Kadrey_00211699 · 3:23-cv-03417
The public register is fair use and precaution. The private register is a known-pirate library plus an instruction not to say so. The instruction is the preference falsification.
The tell · When a memo discusses disclosure strategy before it discusses the work, the audience has already split.
How this became public: Unsealed in Kadrey v. Meta discovery briefing.
Read this pair full-screen Next pair
Step 2 of 5
Same files. The adjective is the product. Pirate is for staff. External is for a later room.
Replace pirate with external
euphemism
Public
Meta public posture · 2024-2025
press and litigation language
Kadrey coverage
“The company describes training data as publicly available or licensed where required, and argues the use is fair.”
Private
Meta staff, suggested edit · 2023
WorkChat edit on a memo
WorkChat
“A suggested edit: replace “pirate” with “external.””
Exhibit Meta_Kadrey · 3:23-cv-03417
The noun is the preference. Pirate names the legal and moral fact. External names a vendor relationship. Same files; different audience.
The tell · Watch for the adjective swap that removes the crime while keeping the dataset.
How this became public: Quoted in unsealed Kadrey briefing.
Read this pair full-screen Next pair
Step 3 of 5
OpenAI’s version is a channel title. Counsel’s documented act is a rename. Google’s “history off” habit belongs in this family — a link-out, not a pair we authored — the same move: make the record quieter. Then the paper rename:
excise-libgen → project-clear
rename
Public
OpenAI privilege position · 2025
counsel in authors’ litigation
Authors Guild v. OpenAI briefing
“OpenAI argued the Slack discussing deletion was privileged legal advice and should stay sealed.”
Private
Jason Kwon and channel participants · 2022
in-house lawyer; research staff
Slack channel title
“The deletion was discussed in a Slack channel first titled excise-libgen. A documented lawyer move in a key batch was to rename it project-clear.”
Docket 1:23-cv-08292
A channel named after the thing being deleted is the private register. A channel named after cleanliness is the audience-facing one. Judge Wang held that a blander title is not privilege.
The tell · When counsel’s first visible act is a rename, read the old name.
How this became public: Magistrate Ona T. Wang, ECF 782, Nov. 24, 2025, Authors Guild v. OpenAI.
Read this pair full-screen Next pair
Step 4 of 5
Libgen1/2 become Books1/2 between the disk and the GPT-3 paper. After-the-fact values language is the last register, not a confession.
Libgen1/2 → Books1/2
rename
Public
Brown et al. · 2020-05
GPT-3 paper authors
Language Models are Few-Shot Learners
“Two internet-based books corpora (Books1 and Books2) appear in the GPT-3 training mix as ordinary data sources.”
Private
OpenAI internal naming, later described in court · circa 2018-2020
research ops
Internal dataset names
“Those dumps became internal sets first called Libgen1 / Libgen2, later relabeled Books1 / Books2 in the GPT-3 paper.”
Docket 1:23-cv-08292
Books is a genre. LibGen is a library with a legal history. The paper needed the first word. The disk had the second.
The tell · If a training-set name gets quieter between the folder and the paper, ask what the folder was called.
How this became public: Described in the authors’ MDL privilege opinions.
Read this pair full-screen Next pair
Step 5 of 5
After-the-fact values language is a second register, not a confession. You now have the sequence: hide, rename, rename again, then say it was never us.
Do not reflect the company we are
values after
Public
Boeing · 2020-01
corporate statement after the messages became public
Company statement
“These communications do not reflect the company we are and need to be, and they are completely unacceptable.”
Private
Unnamed Boeing employees · 2015-2018
staff on the MAX
Internal messages later released by the company
“Would you put your family on a MAX simulator trained aircraft? I wouldn’t. / This airplane is designed by clowns, who in turn are supervised by monkeys.”
After-the-fact values language is a second register, not a confession. It reassigns the private sentences to a few bad speakers so the public sentence can survive.
The tell · “Does not reflect who we are” arrives after the tape. It is not evidence about the tape.
How this became public: Statement issued the week Boeing released the messages.
Read this pair full-screen Next pair

Concealment moves

In no case would we disclose publicly

Replace pirate with external

excise-libgen → project-clear

Libgen1/2 → Books1/2

Do not reflect the company we are