The Machine in the Manuscript

In December 2022, the world of academic publishing changed overnight. The public release of ChatGPT introduced a tool that could generate coherent, well-structured prose on virtually any topic—including scientific research. Within months, manuscripts co-written with AI began appearing in peer-reviewed journals, ethical guidelines were hastily drafted, and a fierce debate erupted over what constitutes authorship, originality, and intellectual contribution in the age of generative AI.

Two years on, the dust has settled into an uneasy consensus: AI tools are here to stay, but their use in scientific writing is governed by an evolving set of rules that every researcher must understand. This article provides a thorough, practical guide to using AI in your research writing—ethically, effectively, and transparently.

The Current Landscape of Publisher Policies

Major publishers have converged on a set of core principles, though the specifics vary:

Springer Nature

Springer Nature's policy, updated in 2024, states that:

• ●AI tools cannot be listed as authors because they cannot take responsibility for the work.
• ●Authors must disclose the use of AI in the Methods or Acknowledgments section, specifying which tool was used and how.
• ●Authors bear full responsibility for the accuracy and originality of AI-assisted content.

Elsevier

Elsevier's guidelines mirror Springer Nature's core position but add a nuance: AI-generated text must be clearly attributed within the manuscript, and the use of AI for data analysis or interpretation is subject to stricter scrutiny than its use for language editing.

Wiley

Wiley explicitly prohibits the use of AI to generate original research content (e.g., fabricating data, running experiments). However, it permits AI use for "improving the readability and language" of a manuscript, provided this is disclosed.

The ICMJE Position

The International Committee of Medical Journal Editors (ICMJE), whose recommendations are followed by thousands of biomedical journals, took a clear stance in 2023: authorship requires the ability to be held accountable for the work. Since AI tools cannot be held accountable, they cannot be authors. Period.

Where AI Helps—and Where It Doesn't

Legitimate and Valuable Uses

Language polishing. For non-native English speakers, AI tools can smooth awkward phrasing, correct grammatical errors, and improve sentence flow. This is perhaps the most defensible use case and can significantly reduce the barrier to publication for researchers working in a second or third language.

Brainstorming and outlining. Stuck on how to structure your Discussion section? AI can generate multiple structural options in seconds. You choose the one that best fits your narrative and then write the content yourself.

Literature summarization. AI can help you quickly digest large volumes of literature, identify key themes, and generate preliminary notes for your Introduction. However, you must verify every citation against the original source.

Code generation and debugging. For computational researchers, AI tools like GitHub Copilot and ChatGPT are powerful assistants for writing analysis scripts, debugging code, and generating data processing pipelines. This use is generally accepted and doesn't raise the same ethical concerns as text generation.

Editing and proofreading. Using AI to catch typos, inconsistencies, and formatting errors is functionally equivalent to using Grammarly or a spell-checker. Few would argue that this constitutes an ethical violation.

Problematic and Prohibited Uses

Generating original research text. Asking ChatGPT to "write the Introduction section of a paper about X" and submitting the output as your own work is, by any reasonable definition, a form of plagiarism. The fact that the text is generated rather than copied doesn't change the underlying ethical issue: you are presenting words and ideas that are not your own.

Fabricating or augmenting data. This is the most serious violation. Using AI to generate synthetic data, "fill in" missing data points, or create fake figures is scientific fraud. It is grounds for retraction, career sanctions, and legal consequences.

Interpreting results. The Discussion section of a scientific paper is where the author's expertise, judgment, and scientific intuition are most critical. Outsourcing this to an AI produces generic, often incorrect interpretations that lack the nuanced understanding of the field that reviewers expect.

Generating citations. Large language models are notorious for "hallucinating" references—generating plausible-sounding but entirely fabricated citations with realistic author names, journal titles, and DOIs. Submitting a manuscript with fabricated references is a form of academic misconduct. Every single citation generated by AI must be independently verified.

The Hallucination Problem

It bears repeating: AI models do not "know" things. They generate statistically probable sequences of tokens based on their training data. This means that they can produce text that is fluent, confident, and completely wrong.

In a 2023 study published in Nature, researchers tested multiple LLMs on their ability to generate accurate scientific references. The results were alarming:

• ●GPT-3.5 fabricated references in over 30% of cases.
• ●GPT-4 performed better but still hallucinated approximately 5–10% of citations.
• ●Even when the cited paper existed, the AI frequently misattributed findings to the wrong paper.

The practical implication is clear: never trust AI-generated citations without manual verification. Cross-reference every citation against PubMed, Google Scholar, or the journal's own website. This is non-negotiable.

Detecting AI-Generated Text

As AI writing has proliferated, so have detection tools. Universities and publishers are deploying software like Turnitin's AI Writing Indicator, GPTZero, and Originality.AI to flag potentially AI-generated submissions.

However, these tools are imperfect. They produce both false positives (flagging human-written text as AI-generated) and false negatives (failing to detect AI-generated text that has been lightly edited). Some researchers have reported that their own, entirely human-written manuscripts were flagged by AI detectors—a situation that underscores the importance of maintaining documentation of your writing process.

Protecting Yourself

If you use AI tools in any part of your writing process:

1. ●Disclose it. Always. Even if you only used ChatGPT to rephrase a single paragraph. Transparency is your best defense.
2. ●Keep records. Save your chat logs, prompts, and the AI's raw output. This creates an audit trail that demonstrates you used AI as a tool, not as a ghostwriter.
3. ●Substantially revise AI output. If you use AI to generate a first draft of a section, rewrite it thoroughly in your own voice, with your own interpretations and emphasis. The final text should be unmistakably yours.

A Framework for Ethical AI Use

Based on the evolving consensus among publishers, ethicists, and research institutions, here is a practical framework:

Level 1: Always Acceptable

• ●Spell-checking and grammar correction
• ●Code generation and debugging
• ●Literature search and summarization (with verification)

Level 2: Acceptable with Disclosure

• ●Language polishing and sentence restructuring
• ●Generating structural outlines
• ●Drafting non-critical sections (e.g., Acknowledgments)

Level 3: Requires Careful Justification

• ●First-draft generation of Methods or Results sections (must be heavily revised)
• ●Translation from another language

Level 4: Never Acceptable

• ●Generating original interpretations or discussion points presented as the author's own
• ●Fabricating or augmenting data
• ●Generating references without verification
• ●Submitting AI-generated text without any human revision

The Human Edge

For all their power, AI tools have fundamental limitations that human editors do not share:

• ●AI cannot evaluate whether your conclusions are supported by your data. A human editor with domain expertise can.
• ●AI cannot detect logical inconsistencies between your Results and Discussion sections. A human can.
• ●AI cannot assess whether your framing is appropriate for the norms of your specific subfield. A human who has worked in that field can.
• ●AI cannot take responsibility for the accuracy of your manuscript. You—and your human collaborators—can.

This is why the role of the professional scientific editor has not been diminished by AI. If anything, it has become more important. As AI-generated text proliferates, the distinguishing mark of a high-quality manuscript is the evidence of rigorous human thinking—the kind that no algorithm can replicate.

How SciScribe Solutions Approaches AI

We use AI internally as a productivity tool for preliminary grammar checks and formatting tasks. We are transparent about this.

However, every manuscript that passes through our hands is edited by a human specialist with relevant domain expertise. Our editors don't just correct English—they evaluate the logic of your arguments, the consistency of your data presentation, and the appropriateness of your citations. This is work that requires scientific judgment, not statistical prediction.

When you work with SciScribe Solutions, you get the efficiency of modern tools and the irreplaceable judgment of human expertise. That combination is what produces manuscripts that get published.

AI is a tool. Like all tools, its value depends entirely on the skill and integrity of the person wielding it. Use it wisely, use it transparently, and use it as a complement to—never a replacement for—your own scientific thinking.