Datasets:

rishanthrajendhran
/

TwiceTold

license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - ai-text-detection
  - idea-provenance
  - creative-writing
pretty_name: TwiceTold
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: human_rewrite
        path: data/human_rewrite.parquet
      - split: ai_original
        path: data/ai_original.parquet

TwiceTold

TwiceTold holds 50 short stories that people wrote from scratch, each working only from the outline of an AI-generated story, together with the 50 AI stories those outlines came from. The ideas in every story are the AI's. The prose of the 50 rewrites is human. A detector of who wrote the prose should call the rewrites human. A detector of whose ideas they are should call them AI. The name comes from Nathaniel Hawthorne's Twice-Told Tales.

Construction

AI stories: GPT-5.6 Sol (high reasoning effort) wrote a 500-word story for each of 50 genres, from the prompt "Write me a 500 word story in {genre}".
Outlines: an outline of each AI story was extracted: an ordered list of [Role] content items.
Rewrites: 17 computer science graduate students and faculty each received one or more outlines, and wrote the story in person rather than through crowdsourcing, to reduce the chance of AI use during writing.

Writers are identified only by pseudonymous ids (W01-W17). No other information about them is included.

Splits

Split	Rows	What it holds
`human_rewrite`	50	the stories people wrote from the outlines
`ai_original`	50	the AI stories those outlines came from

A rewrite and its original share story. load_dataset("rishanthrajendhran/TwiceTold", split="human_rewrite") loads the rewrites.

Fields

Field	Description
`id`	`twicetold_NN_human` (the rewrite) or `twicetold_NN_ai` (the original AI story)
`story`	pair number (1-50): a rewrite and its original share it
`kind`	which of the two this row is
`prose_author`	`human` for rewrites, `ai` for originals
`idea_author`	`ai` for both
`writer`	pseudonymous writer id (rewrites only)
`genre`	the story's genre
`generation_prompt`	the prompt that produced the original AI story
`writer_outline`	the outline the writer worked from, one `[Role] content` line per item
`words`, `text`	the story
`idealens_p_ai`, `proselens_p_ai`	IdeaLens and ProseLens P(AI), as reported in the paper
`editlens_score`, `pangram4_p_ai`	EditLens-3B score; Pangram 4 weighted P(AI), its AI fraction plus half its AI-assisted fraction (flagged as AI at 0.5 or above)

Caveats

Fifty stories in one domain (short fiction) is a small evaluation set.

Scoring

The idealens package (PyPI) runs IdeaLens, ProseLens and the other released detectors on these documents: pip install idealens, then idealens run on a JSONL file with a text field.

All IdeaLens models and datasets are in the IdeaLens collection.

License

The dataset, including the human-written stories, is released under CC BY-NC-SA 4.0: free to share and adapt for non-commercial purposes, with attribution and under the same license. The original AI stories and outlines were produced with OpenAI models.

Citation

@article{idealens2026,
  title         = {IdeaLens: Detecting AI Ideas in Long-form Writing},
  author        = {Rajendhran, Rishanth and Choi, Minjoon and Russell, Jenna and Namuduri, Ramya and B{\"o}l{\"o}ni-Turgut, Deniz and Karpinska, Marzena and Wieting, John and Iyyer, Mohit},
  journal       = {arXiv preprint arXiv:2610.06778},
  year          = {2026},
  eprint        = {2610.06778},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.06778}
}