Written by a human or an AI?.
Can you tell lines by Shakespeare, Keats, Whitman and Dickinson from an AI writing in their style?
Original study
- Porter, B., & Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports, 14, 26133. Source
- Köbis, N., & Mossink, L. D. (2021). Artificial intelligence versus Maya Angelou: Experimental evidence that people cannot differentiate AI-generated from human-written poetry. Computers in Human Behavior, 114, 106553. Source
Our adaptation
An adaptation with short excerpts of the human-versus-AI poetry experiment of Porter and Machery (2024). Two excerpts are shown from each of four poets (in Turkish Yunus Emre, Karacaoğlan, Namık Kemal, Tevfik Fikret; in English Shakespeare, Keats, Whitman, Dickinson); for each poet the pool holds two genuine excerpts from public-domain works published before 1929 and two written in the poet's style by Claude (Anthropic). Each excerpt is first rated for liking from 1 to 5, then judged 'human' or 'AI'. The key is shown on the result screen.
Differences from the original study
- Four-line excerpts are used instead of whole poems.
- The AI excerpts were written by a single model and selected by us.
- Turkish and English pools differ and should be separated by the language column.
Measured variables
The columns of the open data file. List fields are expanded into numbered columns (for example rt_60, est_3); JSON columns contain only numbers and fixed stimulus names.
| Column | Type | Description |
|---|---|---|
r | JSON | Excerpts: [[pool excerpt, judgment 0 human / 1 AI, liking 1–5, time ms, correct 1/0] × 8]. |
score | integer | Number of correct judgments (0–8). |
of | integer | Number of excerpts (8). |
Columns present in every file (14)
row_id | string | Row code. Regenerated at random in every release; it does not identify a person and cannot be matched across releases or experiments. |
experiment | string | Short name of the experiment (URL slug). |
experiment_version | integer | Version of the answer format. For experiments whose format changed, only the current format is published. |
date | YYYY-MM-DD | YYYY-Www | YYYY-MM | Day the answer was given (Istanbul time). If fewer than 5 answers share the same day, language and device, it is coarsened to the ISO week, and to the month if that is still too few. |
date_precision | day | week | month | Precision of the date column. |
lang | tr | en | (boş) | Interface language. Language recording started in the first week of October 2026; for earlier answers it is known only in experiments whose stimuli depend on the language, and empty otherwise. |
device | desktop | mobile | tablet | (boş) | Device class reported by the browser (class only; browser details are neither stored here nor published). |
source | site | embed | Where the answer came from: the balabs site or the experiment embedded on another site. The embedding site is not published. |
mode | free | daily | challenge | race | session | (boş) | Play mode in scored experiments: free play, daily round, challenge, race room, experiment session. Empty for unscored experiments and for older answers that did not store it. |
color_vision | normal | rg | by | contrast | (boş) | In color-based experiments, the color vision mode the participant chose: normal, red-green, blue-yellow or high contrast. Stimuli are generated along different axes per mode, so compare within a mode. Empty for other experiments. |
first_play | 1 | Every row is a participant’s first play of this experiment (always 1). Replays are stored only as leaderboard scores and are not part of the open data. |
seen_before | yes | no | unsure | (boş) | Participant’s own report of whether they had seen this experiment or its known answer before (in experiments that ask). |
duration_s | number | Time from the start of the experiment to submitting the answer, in seconds (measured in the browser, 0.1 s). |
play_score | number | Leaderboard score of the first play (scored experiments; defined on the method card). Empty for unscored experiments. |
Exclusion criteria
The server checks that the excerpts belong to the player's language and that exactly two come from each poet, and computes hits itself.
- Only each participant's first answer to this experiment; replays are not included.
- Simulated rows, bots, banned and sample accounts are not included.
- Answers before 3 October 2026 are not included (the first day the open data notice was live).
Scoring
Leaderboard score: number of correct judgments (0–8); higher is better.
Leaderboard measure: correct guesses (higher is better). The play_score column in the open data is this score for the first play.
Version notes
- October 2026
The interface language (lang) started being stored with each answer; for earlier answers it is known only in experiments whose stimuli depend on the language, and empty otherwise.