rishanthrajendhran/cc2026-deleaked-outlines

CC-2026 De-Leaked Role-Labelled Outlines (Human vs AI)

60,506 web documents, each reduced to a role-labelled outline and then source-blind de-leaked (paraphrased to strip surface authorship), with a human / AI label. Built to study whether AI-vs-human authorship survives when the giveaway surface phrasing is removed.

What each record is

field	description
id	document id (sha256)
source	label — human or ai (doubly confirmed, see below)
format	document format (9 classes, e.g. Nonfiction Writing, Knowledge Article)
role_format	slug of the per-format role set used for extraction
topic	Common-Crawl topic (24 classes, e.g. Finance & Business)
url, date	source provenance
word_count, token_count	source-document length
editlens_bucket, editlens_score	EditLens AI-edit signal (0=human … 3=AI)
pangram4_prediction, pangram4_fraction_ai/ai_assisted/human	Pangram-4 labels on the source doc
source_text	the original document (self-contained)
extracted_outline	role-labelled outline: {document_description, global_themes[], items[{role_name, content, verbatim}]}
deleaked_outline	the same outline after source-blind paraphrase, same structure

How it was made

Source corpus: jjrussell10/cc-2026-editlens-ai-50k-pangram4. "Doubly confirmed" = EditLens and Pangram-4 agree: source=human iff EditLens bucket 0 and Pangram "Human"; source=ai iff EditLens bucket 3 and Pangram "AI".

Two-stage pipeline, both Gemini 3.1 Pro (gemini-3.1-pro-preview, high thinking):

1. Extraction — 6-shot, per-format consolidated role set → the role-labelled outline (extracted_outline).
2. De-leak — a source-blind "canonical paraphrase" (v5) that rewrites the outline to read as machine-generated, removing lifted phrasing, while preserving structure (items 1:1) and meaning (deleaked_outline). The paraphraser never sees the source or its human/AI provenance.

The de-leaked outlines read as ~0.93 fraction-AI / 0% human under Pangram-4 (surface authorship removed).

Composition

• Source: 29,210 human · 31,296 AI (≈ 48 / 52)
• Formats: Nonfiction Writing 27,049 · Knowledge Article 11,867 · News Article 8,087 · Personal Blog 7,100 · Transcript/Interview 2,423 · Academic Writing 1,661 · User Reviews 1,208 · Creative Writing 705 · Personal About Page 406
• Length: median source ~1.1k tokens; 24 topics.

Intended use & what the outlines encode

Research on AI-vs-human authorship detection under surface de-leaking. Key findings from this data (format-controlled held-out AUC, human vs AI):

signal	before de-leak	after de-leak
item length / count (gameable)	~0.50	~0.50
structural + stylometric	0.848	0.841
role frequency + ordering (all roles)	0.898	0.898
idea content (embeddings)	0.947	0.925

The de-leak strips only the surface phrasing channel; the rhetorical-structure (role) and content channels survive almost intact — so a detector trained on these outlines cannot cheat on lifted phrasing but can still learn the real signal. Roles are the robust, topic-independent tell (~0.90); the content signal is ~0.89 even when human and AI are matched on the same topic (so it is not merely a topic detector).

Caveats

• ~0.7% of the source corpus is missing — dropped by Google content-safety filters during extraction/paraphrase.
• source_text is Common-Crawl 2026 content (public); this dataset is a derivative.
• The de-leaked text is deliberately machine-styled; do not treat it as natural human prose.
• source labels are model/heuristic-derived (EditLens + Pangram-4), not human-annotated ground truth.