Generative AI tools

Link to Guidelines for detecting LLM text

• Authors must declare any use of AI tools in their data collection, data analysis, writing, and editing process, including the use of LLM tools such as ChatGPT. 

• Authors under-report AI use (study presented at Peer Review Congress Presentation 2025, preprint)

How to detect possible LLM text

• sounds too perfect

  • neutral or overly balanced tone

  • lacks a real voice or emotion

  • no grammatical, mechanical, or idiomatic errors; no idiosyncratic phrases

• lacks depth, precision, and cohesion

  • only sounds compelling when read superficially, but arguments do not hold up under scrutiny

  • lacks coherence across sentences and sections (semantic inconsistencies, logical errors) (Uchendu et al. 2023)

  • repeats the same idea without developing it further

  • use general terms instead of effectively tailored information relevant to the specific problem context

  • frequently commit factual and common-sense reasoning errors, which are easily noticed by humans (Jawahar, Abdul-Mageed, and Lakshmanan 2020)

• verbose yet generic

  • nothing new: reads stylish and smooth, but fails to provide fresh insights or a clear main message upon closer reading (Ma et al. 2023)

  • not really relevant: loads of text touches on the topic, but appears irrelevant upon closer reading (Dugan et al. 2023)

  • uses filler phrases and meta dialogue (talks about the issue without ever getting to the issue)

  • describes and talks about the topic in a superficial way without sufficient detail and nuance, and without ever taking a stance on the topic; in contrast, human authors often write lengthy paragraphs that weighs arguments (Desaire et al. 2023), using terms such as “but,” “however,” and “although.”

  • uses loads of adjectives and bloated descriptors that can be omitted without loss of meaning

• nonsensical jargon

  • uses nonsensical paraphrases terms for technical jargon (see article with examples)

    • 'chest growth' instead of 'breast tumor', 'flat variable' instead of 'categorical variable'

• fabricated content, including fabricated references

  • LLM-generated and -edited text often contains factual errors, including fabricated data and fabricated citations (see section of 'fabricated citations'). 

This list has been constructed based on an article by Wu, Junchao, et al. "A survey on llm-generated text detection: Necessity, methods, and future directions." Computational Linguistics 51.1 (2025): 275-338.

Resources

• • Wikipedia page from Wikiproject AI Clean-up

  • LLM text detection methods - a review by Wu et al. (2025) - citation: Wu, Junchao, et al. "A survey on llm-generated text detection: Necessity, methods, and future directions." Computational Linguistics 51.1 (2025): 275-338.

  • The Scholarly Kitchen about a taxonomy for describing use of GenAI in articles (GAIDeT)

  • How to detect LLM generated text (published by the Association for Computing Machinery)

Recommended action

• If there is strong evidence that the manuscript was written using LLM tools yet authors did not disclose the use of LLM, please contact the editor in chief and the managing editor about your concern.

• What will happen next:

  • The editor will contact the author about their possible use of GenAI tools and request that the authors disclose any such use using the GAIDeT tool (see Scholarly Kitchen article linked in the Resources of this section). 

  • The editor will invite a report by one of ICB's publication integrity reviewers.

  • The editor's decision letter will point out areas where the writing needs to be revised to enhance clarity (see list of LLM indicators)

What is citation fabrication?

Citation fabrication can occur when authors use genAI tools such as ChatGPT to generate or edit their references. 

Fabrications by genAI tools are sometimes called hallucinations, use of this term is discouraged because it anthropomorphizes these tools.

Where does it occur in the manuscript?

Fabricated citations can occur in the main text of the manuscript and the Reference list. 

What does it look like?

Fabricated citations run the full spectrum from small omissions and addition (missing or added author name) and changes (slight deviations in the wording of the article title or journal name) over combining elements from multiple existing citations (title, author list and/or journal from different real references) to whole-cloth fabrications of journals, titles, and authors.

Some bibliographic resources have become infected by fabricated articles.

Recommended action

• If you suspect that the manuscript contains fabricated references, immediately contact the editor in chief and the managing editor about your concern.

• What will happen next:

  • Initial steps, the journal will ask the reviewer for evidence and investigating the concerns.

  • The editor will invite a report by one of ICB's publication integrity reviewers.

  • The editor will then contact the authors to discuss concerns about citation fabrication, allowing them an opportunity to respond or clarify before proceeding with any formal actions or decisions.

  • Potential actions based on the journal’s findings include notifying relevant parties, rejecting the manuscript, or reporting to the authors’ institutions, depending on the severity of the misconduct identified.