Paper polish: Development of a GPT to fine-tune scientific manuscripts

Dear Editor,

Artificial intelligence (AI)-based tools are commonly used to help edit and improve scientific manuscripts. However, most are too expensive for routine use. OpenAI’s ChatGPT, a widely accessible general-purpose AI, allows users to create customized versions—called GPTs (Generative Pre-trained Transformers)—tailored for specific tasks. Through the GPT builder in the ChatGPT interface, users can provide plain-language instructions and domain knowledge to develop personalized chatbots that serve diverse functions, including scientific editing.

Although AI tools such as ChatGPT are increasingly used for manuscript editing, their output quality varies greatly, especially without structured prompts. To address this, we developed a customized GPT—Paper Polish—using the ChatGPT platform. This tool helps authors improve conciseness, clarity, and completeness of manuscripts and screens submissions against established reporting standards. It is expected to particularly benefit researchers from low-resource settings.

HOW WE DEVELOPED IT

The development of Paper Polish was grounded both on literature review and our practical experience as authors, reviewers, and journal editors. Over the years, we have published several articles on drafting various sections of scientific papers,[1,2,3,4,5,6,7,8,9,10,11,12,13] and on the use of concise language in academic writing.[14] To build a reliable editorial assistant, we uploaded these articles into the GPT and instructed it to extract relevant writing rules. ChatGPT identified 325 rules summarizing writing guidance [Table 1 for sample]: title (n = 25), abstract (n = 25), introduction (n = 25), aims/objectives/research questions/hypotheses (n = 25), methods (n = 30), ethical considerations (n = 20), sample size estimation (n = 20), statistical analysis (n = 25), tables (n = 20), figures and illustrations (n = 20), results (n = 15), discussion (n = 25), acknowledgements (n = 5), references and citations (n = 20), and concise writing (n = 25). We verified the relevance and comprehensiveness of the extracted rules.

Table 1.

Rules for GPT-based manuscript screening (sample only)

Rule number	Description
Rule 001	MACHINE: Flag if the title exceeds 15 words HUMAN: Titles should be concise—ideally ≤15 words—unless longer phrasing adds necessary precision
Rule 033	MACHINE: Flag if abbreviations appear without first expansion HUMAN: Define each abbreviation at first mention
Rule 065	MACHINE: Flag if too many statistics clutter the introduction HUMAN: Select only key data that demonstrate relevance of the topic
Rule 098	MACHINE: Flag if aims do not match title or abstract content HUMAN: Check that the same purpose is reflected across sections
Rule 119	MACHINE: Flag if Methods include result data HUMAN: Restrict Methods to procedures, not outcomes
Rule 135	MACHINE: Flag if consent procedures are not described HUMAN: State whether written or verbal informed consent was obtained and how
Rule 201	MACHINE: Flag if units of measurement are missing HUMAN: Always specify units (e.g., years, kg, %).
Rule 255	MACHINE: Flag if literature cited in Discussion predates Introduction references by many years HUMAN: Use the most recent and relevant sources in both sections
Rule 260	MACHINE: Flag if implications overgeneralized beyond sample or setting HUMAN: Limit conclusions to the population studied
Rule 267	MACHINE: Flag if Discussion lacks linkage to study objectives HUMAN: Ensure each interpretive statement corresponds to a stated aim or hypothesis

Beyond improving writing quality, we wanted the tool to help authors ensure complete and transparent reporting. Therefore, we integrated guidance from the Equator Network (http://equator-network.org/), which provides reporting checklists for different study types (e.g., CONSORT, STROBE). The GPT was instructed to detect study design by analyzing the manuscript’s title, abstract, and methods, and to justify its classification by quoting relevant text. If the tool could not confidently determine the study type, it would prompt the user to choose from likely options. For non-research articles such as editorials or commentaries, the GPT indicates that a checklist is not applicable, and instead focuses on clarity and conciseness.

To enhance usability, we created four initial conversation starters that appear on the GPT’s homepage: “Check clarity, completeness, and conciseness,” “Compare against STROBE items,” “Help make this abstract better,” and “Help improve this case report.” These starter prompts are dynamic—once the GPT detects the study type, they automatically adjust (for instance, “Compare against STROBE items” becomes “Compare against CONSORT items” for randomized trials).

We also made deliberate choices about the GPT’s capabilities. Web search and Canvas (which allows side-by-side editing with the chat) were enabled to enhance user control and interactivity, while image generation, code interpretation, and data analysis were disabled to maintain focus on writing improvement.

A detailed set of instructions was provided to guide the GPT’s behavior, including:

• Preserve meaning, academic tone, and all factual claims, numbers, and results.

• Rate all articles on: Clarity (1–5), Redundancy (1–5), Wordiness/Modifiers (1–5), and Structure and Flow (1–5).

• Specify if the tool’s suggestion succeeded in preserving the author’s intended meaning (Pass/Fail).

• Mention which rules the article did not follow, with a 1–2 sentence rationale.

• Provide a Reporting Completeness score (0%–100%) reflecting the proportion of relevant checklist items present. For each missing or partial item, give: the checklist item number and short text (quoted), where it should appear (e.g., Methods—Participants, Case Description—Clinical Findings), a 1–2 sentence explanation of why it matters, and a suggested concise sentence or short paragraph for the author to adapt.

To reduce hallucinations, we implemented strict safeguards:

• Do not add new facts or results.

• Do not invent data; if data are required and missing, indicate that the author must supply them.

• Do not perform live web searches during checklist comparison.

• When asking a question, do not answer it; present suggested answers for the user to confirm.

We pre-tested the tool on early drafts of our previously published articles, and the outcomes were satisfactory in both language quality and checklist completeness.

HOW TO USE IT

Paper Polish is accessible through ChatGPT at https://chatgpt.com/g/g-68cf72fa727481919ae358934d7d2172-paper-polish. Users can begin by selecting one of four suggested conversation starters, such as “Compare against STROBE items.” These serve as entry points for the tool to analyze and improve manuscripts. Once Paper Polish identifies the study type, the prompts automatically update to offer relevant checklist comparisons. Authors may paste text or upload Word or PDF files. All edits and feedback are presented transparently, in a structured format with minimal jargon, and a clear focus on actionable advice [Supplementary Material for Sample Report].

Free-tier ChatGPT users can analyze one manuscript per day with this tool, but should note message and usage limits that may restrict long, or highly interactive sessions. To stay within these limits, users should upload a single article at a time and request a complete evaluation in one session rather than multiple iterative queries.

It is important to note that Paper Polish is a writing assistant—not a fact-checker or replacement for peer review. It will not fabricate information or results. When necessary, information is missing, the GPT highlights the gap and prompts the author to fill it. Authors should always verify revisions to ensure consistency with field-specific conventions, accuracy of reporting, and adherence to citation standards. The final message from the tool always reminds users: “Authors remain responsible for content accuracy, citations, and style.”

Authors are also advised to follow their target journal’s policy on AI-assisted writing. If a journal prohibits generative-AI use for manuscript content, this tool should not be employed for those submissions. Authors implementing its instructions should acknowledge its role in the “Declaration regarding the use of generative AI” section.

Supplementary Material: Sample Report

Conflicts of interest

There are no conflicts of interest.

SUPPLEMENTARY MATERIAL: SAMPLE REPORT

Here is a full evaluation report for your uploaded article, prepared according to the Paper Polish protocol.

________________________________________

1. Ratings
Criterion	Score (1–5)
Clarity	4
Redundancy	3
Wordiness/Modifiers	3
Structure & Flow	4

Meaning preservation: Pass — the author’s intended meaning is clear and retained throughout.

2. Writing in General

1. (Title) The title includes all essential elements — population, setting, and design — but slightly exceeds the ideal length and could omit “A cross-sectional study” if the journal repeats this detail in metadata.

   Suggested alternative: “Clinical profile of patients attending the Motor Accident Claim Tribunal Board in a tertiary care centre in Kerala.”

2. (Abstract) The aim and design are clearly stated, but the Background section could be condensed by removing generic context (“Following a motor vehicle accident…”). The conclusion could more directly emphasize the key finding and implication.

3. (Introduction) The background effectively establishes relevance but is overly detailed in epidemiology. Shorten by focusing on 2–3 key statistics and the MACT’s role. The knowledge gap (“no prior Indian studies”) appears appropriately placed.

4. (Introduction) The sentence “To the knowledge of the authors, this is the first study in India in this population” could be simplified and moved to the rationale paragraph to improve flow.

5. (Methods) The design (“record-based cross-sectional study”) and ethics approval are clearly stated, but the inclusion/exclusion criteria and sampling method require explicit mention (Rule 103–104 from the master ruleset). Also, mention whether incomplete records were excluded.

6. (Methods) The sample size justification is missing; even for record reviews, a note on inclusion of all eligible records during the period would clarify representativeness.

7. (Methods) Sentence length can be reduced for readability — consider separating IDEAS description and reliability details into a short paragraph.

8. (Results) Well organized and tabulated, but minor overlap exists between text and tables (e.g., Table 1 data repeated verbatim). Summarize key patterns instead.

9. (Results) Figures and tables are clear and sequentially numbered. However, ensure all abbreviations (e.g., IDEAS, TBI, MACT) are defined once per section.

10. (Discussion) The logical sequence from epidemiology → comparison → implications is good. However, some sentences repeat introduction content (causes of TBI, global burden) and could be trimmed.

11. (Discussion) Limit speculative statements such as “This can seriously affect the socioeconomic situation…” by linking to data or references.

12. (Limitations) The section is appropriately reflective but could more explicitly link limitations to potential bias (e.g., retrospective design, missing records).

13. (Conclusion) Strongly restates main findings; consider shortening by merging similar ideas (“need for more assessment centres” and “multicentric prospective studies”) into one cohesive recommendation.

3. Checklist Coverage

Study type identified:

Cross-sectional descriptive study (record-based).

Justification: The Methods explicitly state, “A record-based cross-sectional study was conducted on the patients referred to a tertiary care center for the disability assessment by MACT.”

Checklist chosen for comparison: STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) 2014.

Rationale: STROBE applies to all observational studies including cross-sectional record-based designs.

Estimated Reporting Completeness Score: 85%

Missing or incomplete items:

• Item 6 (Participants): The inclusion/exclusion criteria and methods of record selection are not detailed.

  Why it matters: Clarifies representativeness and reduces selection bias.

• Item 10 (Study size): No explanation of how 248 records were determined or whether it represents all available cases.

  Why it matters: Ensures transparency about sampling adequacy.

• Item 12 (Statistical methods): No description of handling missing data.

  Why it matters: Important for assessing robustness of descriptive results.

• Item 13 (Participants): Flow diagram or explicit description of excluded records is absent.

  Why it matters: Readers cannot verify sample completeness.

• Item 19 (Limitations): Limitations section is good but could include potential confounders (e.g., time since injury, severity of trauma).

  Why it matters: Clarifies scope of inference.

• Item 22 (Funding): Funding and independence statements are present; well done.

4. Concise Language

• 14. Issue: “Following a motor vehicle accident, individuals may seek compensation for their disabilities through legal channels, which involves attending the Motor Accidents Claim Tribunal (MACT) board.”

  Suggested concise rewrite: “After a motor vehicle accident, individuals may seek disability compensation through the Motor Accidents Claim Tribunal (MACT).”

• 15. Issue: “Understanding the clinical profile of these patients and the types of disabilities they sustain can be crucial.”

  Suggested concise rewrite: “Understanding their clinical and disability profiles is crucial.”

• 16. Issue: “Considering all the above complications, TBI is one of the significant public health burdens.”

  Suggested concise rewrite: “TBI is therefore a major public health burden.”

• 17. Issue: “This can add to the difficulty of the claimants and the family as they need to travel long distances and spend a considerable amount of money…”

  Suggested concise rewrite: “This forces claimants and families to travel long distances, increasing costs.”

• 18. Issue: “The study shows that the mean age of patients attending the MACT board was 44.2 years. And the majority were males.”

  Suggested concise rewrite: “The mean age was 44.2 years, and most patients were male.”

• 19. Issue: “This study brings forth the need for assessment of the pattern of disabilities in patients attending the MACT board and their clinical profile which is an unexplored area.”

  Suggested concise rewrite: “This study highlights the need to assess disability patterns and clinical profiles of MACT board patients, an underexplored area.”

Privacy note:

Uploaded text may be stored and processed to generate responses.

Do not upload personally identifiable or sensitive unpublished manuscripts without permission.

This tool offers editorial suggestions only. Authors remain responsible for content accuracy, citations, and style.

Funding Statement

Nil.