alexreversegpt/ai-writing-tells

AI Writing Tells: 100 AI-to-Human Rewrite Pairs

100 short passages written in recognisable LLM house style, each paired with a human rewrite and a label for the specific habit ("tell") being removed.

The pairs are deliberately small — one to three sentences — so the edit is legible. In every row the rewrite keeps the claim and changes the prose: no row is a synonym swap, and none of them make the text longer for its own sake.

Built and released by the team behind ReverseGPT, an AI writing assistant that rewrites AI-drafted text into prose that reads like a person wrote it.

Why this exists

Most public data on machine-generated text is labelled at the document level: human vs ai, one flag per document. That is useful for classifiers and useless for editing — it tells you a paragraph is off without telling you which habit gave it away.

This set labels the habit. Each row isolates one pattern (a hedge, a tricolon, a not just X, but Y construction) so a model or a person can learn the edit rather than the verdict.

Dataset structure

One JSON object per line in train.jsonl:

field	type	description
id	string	Stable row id, rgpt-001 … rgpt-100
domain	string	Where the passage would appear (see below)
category	string	The tell being demonstrated
tell	string	One line naming the specific habit in this row
ai_text	string	The passage in LLM house style
human_text	string	The rewrite
edits	string[]	The operations that turn ai_text into human_text

Example row:

{
  "id": "rgpt-008",
  "domain": "product_copy",
  "category": "rule_of_three",
  "tell": "Alliterative tricolon standing in for a claim.",
  "ai_text": "Our solution is fast, flexible, and future-proof.",
  "human_text": "It renders a 200-page PDF in under a second. Nothing else we tested came close.",
  "edits": ["break_triad", "add_number"]
}

Categories

14 categories, 7 rows each (9 for superlative_vague):

category	the habit
hedge_and_filler	Throat-clearing before the sentence says anything
rule_of_three	Tricolons standing in for a claim
thesaurus_verbs	delve, leverage, utilise, foster, harness
not_just_but	not just X, it is Y escalation
empty_opener	"In today's fast-paced landscape"
summary_closer	"In conclusion", restating what was just said
abstract_nouns	landscape, realm, ecosystem, journey, tapestry
uniform_rhythm	Consecutive sentences of identical length and shape
passive_hedge	Passive voice hiding who acted
list_reflex	Bullets and numbered steps for two short facts
balanced_contrast	Both-sides framing that never picks a side
formula_benefit	"By doing X, you can significantly improve Y"
redundant_restatement	Two words doing one word's job
superlative_vague	Intensifiers with no quantity behind them

Domains

blog (22), product_copy (19), email (16), docs (11), academic (11), social (10), support_reply (6), cover_letter (5).

Edit vocabulary

edits is drawn from a closed set, so it can be used as a multi-label target: cut_hedge, cut_opener, cut_closer, cut_preamble, cut_redundancy, break_triad, break_formula, unlist, vary_rhythm, plain_verb, concrete_noun, name_actor, active_voice, add_number, add_specific, add_stake, add_instruction, add_deadline, add_mechanism, add_sequence, add_source, show_evidence, show_method, state_ask, state_question, take_side, recommend, rank_claims, downgrade_claim, offer_remedy, flag_risk, own_it, shorten.

Usage

from datasets import load_dataset

ds = load_dataset("reversegpt/ai-writing-tells-100", split="train")

# Supervised rewriting
pairs = [(r["ai_text"], r["human_text"]) for r in ds]

# Or as a 14-class single-label classification task over ai_text
labels = sorted({r["category"] for r in ds})

Both directions are usable. ai_text -> human_text trains or evaluates a rewriter; ai_text -> category trains a tell classifier; human_text alone is a small negative set for anything that claims to detect machine style.

How it was built

Written by hand, not sampled from a model, which is the point: generating the ai_text side with an LLM would reproduce one model's habits at one point in time. These are the patterns that survive across models and across years, drawn from editing work on real drafts.

The editing rules behind the human_text column are the ones set out in How to Humanize AI Text Without Losing Your Meaning — revise for rhythm, specificity and structure while leaving the underlying claim intact. For where automated rewriting sits relative to doing this by hand, see StealthGPT vs ReverseGPT: Humanizer Features and Fit.

Limitations

• English only, and mostly US/UK business and academic register.
• 100 rows. This is an evaluation and probing set, not a training corpus. Use it to measure a rewriter or to seed few-shot prompts, not to fine-tune from scratch.
• One reference rewrite per row. There are many good rewrites of each passage; human_text is one, so exact-match scoring is the wrong metric. Prefer human judgement or a preference model.
• Not a detector benchmark. The ai_text column imitates LLM house style; it is not sampled from a specific model, so it should not be used to make claims about any detector's accuracy on real model output.
• The human_text numbers, names and anecdotes are illustrative, written to show what specificity looks like. They are not factual records.

Licence

CC BY 4.0. Use it commercially, modify it, redistribute it — keep the attribution.

Citation

@misc{reversegpt2026aiwritingtells,
  title  = {AI Writing Tells: 100 AI-to-Human Rewrite Pairs},
  author = {ReverseGPT},
  year   = {2026},
  url    = {https://huggingface.co/datasets/reversegpt/ai-writing-tells-100},
  note   = {Released by ReverseGPT, https://reversegpt.ai}
}