Skip to content

Latest commit

 

History

History
140 lines (98 loc) · 6.2 KB

README.md

File metadata and controls

140 lines (98 loc) · 6.2 KB
Raw

Python 3.12.1 MIT License arXiv

The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates

This repository contains the implementation for the models' data labeling and experiments in The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates.

@misc{latona2024ai,
      title={The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates}, 
      author={Giuseppe Russo Latona and Manoel Horta Ribeiro and Tim R. Davidson and Veniamin Veselovsky and Robert West},
      year={2024},
      eprint={2405.02150},
      archivePrefix={arXiv},
      primaryClass={cs.CY}
}

Please consider citing our work, if you found the provided resources useful.

1. The Idea and the Repository in a Nutshell

Journals and conferences worry that peer reviews assisted by artificial intelligence (AI), particularly large language models (LLMs), may negatively influence the validity and fairness of the peer-review system, a cornerstone of modern science. In this work, we address this concern with a quasi-experimental study of the prevalence and impact of AI-assisted peer reviews in the context of the 2024 International Conference on Learning Representations (ICLR), a large and prestigious machine-learning conference. In this study, we conducted three analyses:

  1. Prevalence of AI-assisted reviews: We quantify the number of reviews labeled as AI-assisted (according to GPTZero LLM detector).
  2. Effect of AI-assisted reviews on scores: We quantify the difference between AI-assisted and human review scores on submissions submitted to ICLR2024.
  3. Effect of AI-assisted reviews on paper acceptance: We study whether receiving an AI-assisted review boosts the likelihood of acceptance.

These three analysis and the results we obtained are illustrated in

2. Data Collection and Dataset Description

Our data collection process consists of two steps:

  • We download papers, reviews, and acceptance decisions from 2018-2024 from the Openreview API.
  • We label all reviews as either AI-assisted reviews or human written reviews using GPTZero.

By following these steps we obtain the following statistics relative to the datasets we obtained:

Year Reviews Submissions Acceptance LLM reviews
2018 2921 1007 36.0% 57
2019 4734 1569 31.5% 95
2020 7783 2593 26.5% 123
2021 11488 3009 29.1% 216
2022 13161 3422 32.0% 164
2023 18575 4955 24.3% 176
2024 28028 7404 30.5% 4887
Total 86690 23959 --- ---

We structure this dataset in the following folder:

analyses/
  -data/
    - submissions_iclr_2018_2024.csv
    - reviews_iclr_2018_2024_annotated.csv
    - decisions_2024.csv
README.md

We summarize the structure of each dataset:

data/submissions_iclr_2018_2024.csv: This dataset contains all the information about papers submitted to ICLR from 2018 until 2024.

| id | creation_date | year | title | abstract | pdf | keywords | authors | authorids |

  • id: submission unique identifier
  • creation_date: date of the submission
  • year: Year of publication.
  • title: Title of the paper.
  • abstract: Summary of the paper's content.
  • pdf: Path to the PDF file of the paper.
  • keywords: Keywords related to the paper's content.
  • authors: Names of the authors.
  • authorids: Email addresses or other identifiers of the authors.

data/reviews_iclr_2018_2024.csv: This dataset contains all the information about reviews submitted as an evaluation of reviews at ICLR from 2018 until 2024.

| id | review_id| year | review | rating | confidence | ai | human | mixed | isAI |

  • id: submission unique identifier
  • review_id : review unique identifier
  • year: Year of publication.
  • review: content of the paper
  • rating: score given by the reviewer to the submission
  • confidence: confidence of the reviewer in the score provided .

We provide the dataset annotated accordingly to GPTZero (explanation below) data/reviews_iclr_2018_2024.csv which does provide the labels that we use for our analysis. Specifically, this dataset contains the following additional three columns

  • ai: likelihood of the review to be AI-assisted according to GPTZero.
  • mixed: likelihood of the review to be mixed (parts human and AI-assisted written) according to GPTZero.
  • isAI: if 1 == AI-assisted else human

3. Environment Setup

With the repository cloned, we recommend creating a new conda virtual environment:

conda env create -n aireviewlottery
conda activate aireviewlottery

and install

pip install -r pip_requirements.txt

4. Getting Started: Labeling data with GPTZero

To obtain the datasets described above and run the analysis you can navigate to analyses/0_reviews_labeling. Then to start the labeling process you can simply run

python gptzero_labelling.py --api_key "your_api_key"

This label it is easily customizable as it allows to subselect on the portion of the data you might want to label. For instance by setting the bash filter parameter to an integer value between 2018-2024 you can label only the reviews submitted in the specified year.

python gptzero_labelling.py --api_key "your_api_key" --filter 2022

The script contains other parameters that allows to change the directories where data are stored. Code is fully commented to provide easier understanding of the different steps.

5. Study 1: Prevalence Analysis