Techniques for supercharging academic writing with generative AI

Full text

nature biomedical engineering
https://doi.org/10.1038/s41551-024-01185-8
Comment
Techniques for supercharging
academic writing with generative AI
Zhicheng Lin
Generalist large language models can elevate
the quality and eciency of academic writing.
To many researchers, academic writing evokes a Sisyphean ordeal: it
robs precious time and mental bandwidth that could be better spent
doing actual science. Franz Kafka expressed it eloquently: “How time
flies; another ten days and I have achieved nothing. It doesn’t come
off. A page now and then is successful, but I can’t keep it up, the next
day I am powerless.” Although digital writing tools such as Grammarly,
QuillBot or Wordtune can ease the burden of writing by assisting
with basic language tasks — such as spelling and grammar checking,
paraphrasing, and providing suggestions on style, tone, clarity and
coherence — these tools often lack nuance and fall short when more
substantive writing assistance is needed. Professional writing services
offer advanced editing, rewriting and even writing from scratch, but
they are not accessible to those with limited financial resources and
to those who need it most, such as non-native English researchers in
economically disadvantaged regions. This exacerbates a communica-
tion bottleneck that hampers scientific progress1,2.
Tools leveraging generative artificial intelligence (AI), and par-
ticularly generalist large language models (LLMs), are truly useful as
writing assistants. LLMs are widely available, versatile and come across
as collaborative, patient and largely non-judgemental
3
. However, using
them for effective and responsible writing demands more expertise
than the use of specialized digital writing tools, including choosing
the appropriate AI systems and crafting effective prompts4–6. With
this Comment I aim to help researchers to tap into the potential of AI
to elevate the quality and efficiency of academic writing, and to make
the writing process feel less like drudgery and more joyful. In the follow-
ing sections I introduce a collaborative framework for AI engagement,
delineate effective routines and prompting techniques, and discuss
major caveats stemming from AI ethics and AI policy. Although writ-
ten with biomedical and behavioural scientists in mind, the methods
introduced here are applicable to scholarly writing in general.
A writer–AI collaborative framework
LLMs power ChatGPT, Gemini, Claude and other popular AI chatbots.
They are trained using extensive corpora, with some also including
multimodal datasets (visual, audio and textual inputs). They excel at
understanding and responding to language-based tasks through user
queries (that is, prompts). Although the best way to engage with these
tools in science and education is being intensely debated3,4,7–11, they are
undeniably set to transform academic research3,12.
In this section, I outline a collaborative framework for AI engage-
ment in academic writing. The framework addresses (1) the short-term
and long-term rationale for engagement, along with their underlying
mechanisms and potential limitations; (2) the role of AI throughout
the writing process, conceptualized through a two-stage model for
writer–AI collaborative writing; and (3) the nature of AI assistance
in writing, through a model of types and levels of writing assistance.
Rationale. The importance of engaging with AI can be distilled into
short-term and long-term goals. Examples of short-term goals are
enhancing writing productivity, quality and creativity for specific
projects. AI can make writing faster and better, enabling greater focus
on content and ideas — thereby reducing communication costs and
enhancing creativity and quality
9,13–15
. There are two central mecha-
nisms underpinning these gains: cognitive offloading and imaginative
stimulation. Cognitive offloading works similarly to how calculators
or statistical programs assist by crunching numbers. By offloading to
the AI some of the mental work (from mechanical elements such as
spelling and grammar to more nuanced tasks such as word choices,
naming and summaries), writers can free up cognitive resources for
other tasks. Cognitive offloading therefore improves performance
and reduces the likelihood of errors.
Imaginative stimulation works through the synergistic collabo-
ration between writers and AI (not unlike the inventive camarade-
rie among Renaissance artists, scientists and patrons, which fuelled
a period of explosive creativity and discovery). LLMs can present
new perspectives, ideas and framing that can spark creativity. For
example, in brainstorming sessions, an LLM can serve as a bouncing
board, offering counterarguments to challenge prevailing notions. In
writing-evaluation sessions, LLMs can provide constructive feedback to
help refine the exposition, and may offer interdisciplinary perspectives,
challenging the writer to think beyond conventional paradigms. Such
interaction, which should be iterative, can broaden the intellectual
space for the writer and promote dynamism that may lead to more
well-rounded arguments and to novel insights and solutions.
Long-term goals involve the educational aspects of engaging
with LLMs for academic writing16,17, rooted in the age-old educational
philosophy of ‘learning by doing’. LLMs modernize and maximize this
approach via two distinctive advantages: adaptive interactivity and
instant availability. LLMs offer a personalized and interactive learning
experience that caters to the individual style and learning pace of each
writer. The adaptability of LLMs is not dissimilar from the one-to-one
tutoring model, which has long been recognized for its effectiveness in
personalizing education to the learner’s needs. By providing real-time
feedback to the writer (as in traditional writing workshops), the LLM
further helps the writer recognize and correct mistakes, enhances
their writing skills, and facilitates a deeper understanding of language.
Instant availability, on the other hand, provides an immediate
on-demand learning environment, making learning flexible, hands-on
and accessible. This echoes the ethos of online open-access educational
resources, which democratize learning by making learning materials
readily available to all. Still, access to the most powerful LLMs (such
as GPT-4) requires a subscription and may not always be available in
all regions.
Check for updates

nature biomedical engineering
Comment
them. (The coaching role has roots dating back to ancient peda-
gogical traditions, where tutors guided learning through the
Socratic method of questioning and dialogue. Modern writing cen-
tres have expanded this coaching capacity, with tutors reviewing
and advising on drafts. The AI tool as a writing coach may be con-
sidered the digital evolution of these time-honoured educational
approaches.)
On the other hand, in direct assistance, the LLM serves instead as
a co-writer, contributing original text for the manuscript. Throughout
history, creative collisions between authors have fuelled literary and
scientific feats (most notoriously, the Goncourt brothers co-writing
famed novels, and the cross-continental correspondence between
Darwin and Wallace catalysing modern evolutionary theory). By digi-
tally extending the collaborative scope, LLMs can propel this tradition
forward.
The degree of assistance can be categorized into five levels. Each
successive level exhibits a greater degree of interactive engagement
and cognitive contribution from the LLM, as defined in Table 1 along
with example prompts. In the first level, ‘basic editing’, the LLM serves
as a digital extension of a traditional proofreader. It ensures language
accuracy and lexical variety. In the second level, ‘structural editing’, the
LLM is used as a linguistic artisan that reshapes the text to enhance its
clarity and coherence. In the third level, ‘creating derivative content’,
the LLM synthesizes existing content into concise summaries, suc-
cinct abstracts or engaging titles. In the fourth level, ‘creating new
content’, the LLM transcends the role of an editor; the tool becomes a
collaborative partner that brainstorms and contributes original ideas
or alternative perspectives that enrich the narrative and broaden the
intellectual horizon. And in the fifth level, ‘evaluation or feedback’, the
LLM acts as a critical reviewer and mentor, offering feedback, analysing
strengths and weaknesses, and providing constructive suggestions for
refinement and improvement. Across all these levels, assistance can
The use of LLMs in academic writing also comes with its draw-
backs and limits. First, although LLMs can produce articulate and
authoritative-sounding text, users ought to verify the accuracy of the
generated content through fact-checking and cross-referencing3,11,18.
Responsible use also entails providing appropriate references and
adhering to ethical guidelines and disclosure practices11.
Second, overdependence on LLMs for writing risks could deterio-
rate writing skills. AI assistance in writing should be used with a critical
eye and an intent to learn through active thinking, making decisions
about the tool’s output, deciding on proper implementation, and
noticing and understanding the changes made by the LLM, rather than
blindly accepting its output.
Moreover, the quintessential elements of great writing — insight,
originality and creativity — cannot yet be fully replicated by LLMs.
Even though AI can assist the writing process by facilitating cognitive
offloading and imaginative stimulation, and by helping with summari-
zation, explanation, brainstorming and other tasks3,14,15, good writing
cannot rescue poor scholarship.
Two-stage collaborative writing. How can LLMs best help the process
of academic writing? The writing process first involves developing
the essence of the story, its value proposition and its structural out-
line19–22. By way of example, the story of this Comment is that academic
writing has always been difficult, and that LLMs can elevate its qual-
ity and efficiency. This Comment’s value proposition lies in provid-
ing a timely and much-needed conceptual framework for writer–AI
collaborative writing, along with practical techniques to effectively
harness AI in academic writing. The structural outline of the piece
includes the writer–AI collaborative framework, a table summarizing
the levels of writing assistance and a box with a list of prompts. In this
AI-inspired stage, the writer solicits imaginative and creative sugges-
tions from the LLM. This interaction expands the writer’s conceptual
horizons and fuels creative thinking, which can lead to better ideas and
structures.
The second stage of the writing process involves the iterative
drafting of a manuscript. In this AI-assisted stage, the writer taps into
the LLM’s linguistic capabilities for assistance, typically by eliciting
critical feedback from the LLM, which acts as a second set of eyes to
identify potential areas for refinement and clarification. The writer
can capitalize on the LLM’s linguistic prowess to improve spelling
and grammar accuracy, and to enhance word choice, clarity and flow.
Incidentally, statistical and drawing software (which can also empower
writing) may also leverage AI-based assistance (particularly in problem
solving3 and code writing9,23).
Therefore, the LLM can be woven into the entire fabric of the
writing process, assisting in both high-level creative tasks (such as
brainstorming and critical feedback) and in low-level compositional
tasks (such as language polishing and line editing). In this iterative,
human-centric dialogue, the writer integrates or discards the LLM’s
suggestions as deemed fit.
Types and levels of writing assistance. A critical point — one that is
also central to discussions in AI ethics — is the nature of AI assistance in
writing
3,14
. In other words, what types and levels of assistance in writing
should AI provide?
On the one hand, the assistance can be indirect, where the
LLM functions as a writing coach, providing suggestions and feed-
back without directly altering the text itself. This requires users to
decide whether to accept the suggestions and how to implement
Table 1 | Five levels of writing assistance by LLMs, with
example prompts
Level Prompt
(1) Basic editing, such as checking
spelling and grammar, or
suggesting synonyms.
“Check the spelling and grammar in
this paragraph, and suggest synonyms
for any repetitive words.”
(2) Structural editing, such as
paraphrasing, translating, or
improving the structure of the text,
or its low or coherence.
“Paraphrase this lengthy sentence
to improve its clarity and low, and
translate it to French.”
(3) Creating derivative content, such
as summarizing, creating titles and
abstracts, rewriting or generating
analogies.
“Summarize this document and
create a short, catchy title for a journal
submission.”
(4) Creating new content, such
as completing, continuing or
expanding text, or brainstorming
ideas.
“Continue the text to explain the key
question being addressed. Show why
it is important, drawing parallels or
analogies where you see it.”
(5) Evaluation or feedback, such as
assessing the quality of the writing
or inding weaknesses in it.
“Review this introduction and highlight
any logical gaps or areas that need
further development.”
The type of assistance — direct (levels 1–4) or indirect (level 5) — can be invoked by
prompting accordingly. For example, for levels 1–4, adding the prompt “Do not revise or
rewrite the text, but only offer suggestions for the user to implement” results in indirect
assistance; conversely, for level 5, the prompt “Revise the text to implement the suggestions
provided” leads to direct assistance. These ive levels exemplify the utility of the LLMs that are
publicly available as of January 2024.

nature biomedical engineering
Comment
manifest either directly, by providing textual revisions and content
that can be used in the manuscript, or indirectly, by offering only sug-
gestions that must be implemented by the writer.
Routines and techniques
What effective routines and prompting techniques can best help
unlock the potential of AI throughout the writing process? The
educational utility of AI can be more effectively realized through
indirect assistance (prompt 5 in Table 1), whereas shorter-term and
project-specific benefits are best achieved through direct assistance
(prompts 1–4 in Table 1). The type of assistance can be selectively
activated through prompting6, by instructing the LLM either to offer
suggestions without altering the text (indirect assistance) or to imple
-
ment these suggestions by revising the text or adding new content
(direct assistance).
The implementation of the various levels of assistance is also
outlined in Table 1. To do so effectively, one must consider how AI
assistance fits organically into the writing routine. Academic writing
takes many forms, and is shaped by discipline, genre, message, audi-
ence and the writers themselves
24–27
. Writing a brief empirical report on
a new phenomenon differs from crafting an elaborate opinion piece on
a contentious subject. Some writers draft quickly and revise fiercely,
whereas others iterate section by section. Yet all writers must start with
a point, however vague, that conveys a message (stage 1), even if they
end up deviating from it in a later draft (stage 2). As F. Scott Fitzgerald
put it, “You don’t write because you want to say something; you write
because you’ve got something to say.”
Accordingly, an effective routine consists of three iterative pro
-
cesses: outlining the manuscript (broadly defined to also include
the central message), writing the content and editing the draft. Each
process benefits from various levels of assistance that can be triggered
by the prompting techniques provided in Box 1. To maximize their
utility, the prompts are designed to be specific for each task yet in the
context of academic writing.
The optimal degree of assistance depends on the project and on
the author’s needs. By way of example, for this Comment I had a clear
concept and narrative, hence I did not use LLMs for outlining the manu-
script. However, I did extensively use ChatGPT 4 and Claude 2 to edit the
text (‘Basic editing’ in Box 1), and occasionally employed completion
and continuation techniques when stuck (‘Completing or continuing
text’ in Box 1), which greatly enhanced efficiency. When addressing
comments from the journal’s editor and the reviewers on ethics, I first
catalogued all the essential ethics points to be included. I then sought
feedback from ChatGPT on this organization (item 1.2 in Box 1), which
led me to reorder the points slightly. While writing, I merged certain
points for tighter organization. I also consulted Perplexity AI with
regard to relevant content (I did not use this LLM’s output directly, yet
it provided helpful assurance).
Regulations of the use of AI in scholarly publishing are in flux, given
the abrupt emergence of generative AI. The types and levels of writing
assistance by LLMs discussed in this Comment do not violate current
guidelines stipulated by most associations on scholarly publishing,
publishers and journals. However, caution may be warranted28. As of
this writing, the Lancet group and Elsevier only permit AI to “improve
readability and language of the work”. Although I expect this stance
to change (Science originally forbade the use of AI-generated text
on 26 January 2023, but abolished this restriction on 16 November
2023), writers should pay close attention to the relevant policies of
their target journals.
Outlining the manuscript. The outline of a manuscript serves as a
roadmap to guide the writing and editing that follows. Although it
does vary in elaboration and detail, the outline must encapsulate the
central message while addressing these key considerations: what value
does the manuscript offer to the audience25,26? How should the manu-
script be structured to effectively convey the key points through sto-
rytelling20–22? Although much of this work must come from the writer
(through thinking, reading, talking, discussing and drafting), LLMs can
assist with the content of the outline and its actual writing. LLMs can
help ensure that all critical points are covered and that the value to the
intended audience is clear through brainstorming (assistance level 4)
and feedback (assistance level 5). For example, as illustrated by the
prompts in Box 1, LLMs can help generate ideas for the outline, identify
key points, elaborate on an idea or concept, or clarify the value proposi-
tion. Also, LLMs can offer critical feedback on key points, the outline
itself, the value proposition and the narrative flow. The actual writing
of the outline can be LLM-assisted through rewriting (assistance level 3)
and editing (assistance levels 1 and 2) to ensure a logical narrative flow.
Box 1 provides examples of prompts to use when passing a hastily
written outline to an LLM for language polishing and for structuring
a compelling story arc.
Writing the content. Outlines rarely remain fixed; they undergo adjust-
ments during the iterative writing and editing processes. Yet when an
outline is ready, the writing process becomes more structured. LLMs
can assist in creating both derivative content (based on existing text
from the writer or from another source; assistance level 3) and new
content (by following the writer’s instructions; assistance level 4). For
creating derivative content, LLMs can be used to transform text by
rephrasing it to vary the exposition or to avoid plagiarism, and to sum-
marize lengthy text to help the writer better understand the material
or to extract key points from it. For creating new content, LLMs can
assist by completing or continuing a piece of text. For example, when
a writer is uncertain about which words or phrases to use in a sentence,
the LLM can supply multiple options; if the writing is ‘stuck’, the LLM
can ‘take over’ to provide text that keeps the narrative moving; and if
the writer does have specific ideas for continuing, the LLM can also
help expand the text to implement those ideas. Moreover, LLMs can
assist in the writing of specific parts of a manuscript. For instance,
they can generate topic sentences and transition sentences to improve
flow. And when the main text of a manuscript is ready, they can craft
potential abstracts or titles for it. For all of these types of assistance,
examples and instructions can be added to the prompt to refine the
LLM’s output (Box 1).
Editing the draft. Great papers may emerge from messy first drafts but
rarely from sloppy rewriting or editing. LLMs can facilitate this editing
process by handling basic tasks, supporting more advanced tasks and
acting as an advisor, a peer reviewer or a non-specialist reader (Box 1).
Basic editing tasks involve checking spelling and grammar, sug-
gesting synonyms and antonyms, and enhancing vocabulary (Table 1).
For example, the prompts in Box 1 can turn an LLM into a digital proof-
reader tailored to the needs of the writer, helping with copyediting and
improving the clarity and flow of the text, and educating the writer in
the process.
Advanced editing tasks include summarizing key points and organ-
izing them logically, extracting key ideas from the text to suggest
examples and analogies that enhance its readability, revising text to fit a
specific style, analysing the style of the text to emulate it, and rewriting

nature biomedical engineering
Comment
Box 1
Example prompts to invoke the assistance of LLMs in academic
writing
The prompts are meant to guide the LLM by providing it with relevant
context, examples or explicit instructions and even emotional cues,
and to constrain the LLM’s output by instructing it to explore multiple
options or to take on certain roles.
1. Outlining
1.1 Brainstorming
“Generate potential sections and sub-sections for a manuscript
exploring {title or topic}.”
“Identify key points to be discussed under the topic of {topic}.”
“Expand on the idea of {idea}, detailing {aspects of discussion}.”
“What value could the exploration of {topic} bring to the audience
of {audience}? Suggest dierent angles of value proposition for the
manuscript.”
1.2 Evaluation and feedback
“Evaluate the key points listed under the sections of {sections} in
the provided outline for relevance, depth, and alignment with the
manuscript’s central message of {message}. Outline: {outline}.”
“Review the provided outline for the manuscript on {title or topic} and
suggest improvements for clarity, coherence, and completeness.
Outline: {outline}.”
“Assess the value proposition presented in the outline for the
manuscript on {title or topic}. Suggest any additional angles or
reinements to better appeal to the intended audience of {audience}.
Outline: {outline}.”
“Analyse the narrative low of the provided outline for the manuscript
on {title or topic}. Provide suggestions for enhancing the storytelling
aspect to eectively convey the key points and value proposition.
Outline: {outline}.”
1.3 Rewriting and editing
“Edit the jotted outline for the manuscript on {title or topic}. Rewrite
to make it concise. Outline: {jotted outlines}.”
“Propose a narrative structure that eectively communicates the key
points and value proposition of {title or topic}, ensuring a coherent,
logical low from introduction to conclusion. Key points: {key points}.
Value proposition: {value proposition}.”
2. Writing content
2.1 Transforming text
“Act as a top editor in top journals. I will provide you with text. Your
task is to paraphrase the text. Show 3 versions. Conirm with OK.”
“Act as a top editor in top journals. Summarize the text provided. The
summary {requirements; for example, needs to be in about 100 words
and cover all the main points}. Text: {text}.”
2.2 Completing or continuing text
“Act as a top editor in top journals. I will provide you with text that
is missing some parts, marked with {...}. Your task is to ill in the {…}
parts. Show 5 variations of completion. Include only the illed-in text
in the output format. Conirm understanding with OK.”
“Prompt 1: Act as a top editor in top journals. I will provide you with
text. Your task is to continue the text based on the instruction I
provide. Show 3 versions of continuation. Conirm with OK.
Prompt 2: Instruction: Continue from {end of text}. Explain {what to
explain}. Show {points to make}, drawing parallels or analogies where
you see it. End by {what to end}. Text: {text}.”
2.3 Writing parts
“Act as a top editor in top journals. Provide three alternative
topic sentences for each of the following paragraphs. Ensure
{requirements; for example, that they encapsulate the main idea of
each paragraph and align with the overall theme of the manuscript}.”
“Create three alternative transition sentences between the following
two paragraphs:
Paragraph 1: {paragraph}
Paragraph 2: {paragraph}”
“Act as a top researcher writing for top journals. Write three
alternative {abstracts, titles, etc.} for the attached manuscript to
be submitted to {journal name}. The {abstract, title, etc.} needs to
{requirements}. Two examples {abstract, title, etc.} are given below.
Example 1: {example}
Example 2: {example}”
3. Editing the dra
3.1 Basic editing
“Act as an expert editor for top scientiic journals (Nature, Science) to
improve the clarity and low of the writing. Take a deep breath—this is
very important for my career! I will give you text later, and your job is
to oer three revisions with a brief explanation of the changes based
on the following instructions.
The irst version is a copyedited version, with the following rules:
only make needed changes (for example, if it is just a matter of
subjective choice/preferences, don’t make the changes)
when the text provided is long, copyedit through each paragraph
(that is, don’t condense the long text unless necessary)
The second and third versions are the ones you revise to maximize
clarity and low by following the key recommendations suggested by
the following books, but make the two versions distinct:
1. “The Elements of Style” by William Strunk Jr. and E.B. White
2. “Style: Lessons in Clarity and Grace” by Joseph M. Williams
and Joseph Bizup
3. “On Writing Well” by William Zinsser
4. “The Sense of Structure: Writing from the Reader’s Perspec-
tive” by George Gopen
Your output consists of the three versions of texts and a brief
explanation of the changes for each version (omit the text provided

nature biomedical engineering
Comment
introductory or concluding sentences or paragraphs to align them with
the context of the manuscript and the writer’s taste.
When acting as an advisor or a peer reviewer, an LLM can provide
thoughtful and constructive feedback by evaluating the writing to
identify its strengths and weaknesses, and by offering concrete ways
to improve it. By acting as a non-specialist reader, it can identify jargon
in the text and enhance the accessibility of the writing.
Ethical and policy considerations
The use of generative AI in academic research has raised pressing ethi-
cal questions that are being intensely debated. In academic writing,
because writing assistance — from spell checkers and grammar check-
ers to professional copyeditors — has long been widely accepted, it
seems justified for AI to be used for the same types of task and need.
In particular, for low-level contributions (levels 1 and 2; Table 1),
the output of LLMs is akin to that of common writing-assistance
software.
However, for high-level contributions (levels 3–5), using the out-
put of LLMs verbatim raises ethical considerations and potentially legal
challenges
11
. Extensive use of LLMs for academic writing and not prop-
erly attributing their assistance risks presenting AI-generated content
as original work, thus raising ethical issues of attribution and, possibly,
plagiarism. Currently, most publishers forbid listing an AI system as an
author28, yet extensive use challenges the concept of authorship when
AI contributes most of the text29. It is likely that authorship guidelines
and practices will evolve as AI technology rapidly integrates into aca-
demia. Guidelines for the ethical use of AI in academic writing will need
to consider that restricting the scope of AI use is hardly enforceable.
Still, understanding and managing ethical and policy considerations
is integral to the responsible and productive use of AI.
Another issue is that, at present, the copyright holder of output
generated by LLMs remains legally ambiguous. Microsoft, Google
and other providers of generative AI tools offer their users protection
against potential legal risks, whereas OpenAI transfers the owner-
ship and rights of its output to users. Yet the fact that by default most
LLMs retain the prompts for training purposes raises concerns about
copyright (when the prompt includes copyrighted material) and confi-
dentiality or privacy (when prompts include patient data or data from
human participants). Users can opt out of data usage for model training
by disabling ‘Chat history & training’ in ChatGPT or ‘AI Data Usage’ in
Perplexity AI, or by using LLMs designed to respect data privacy (such
as ChatGPT Team or ChatGPT Enterprise).
Moreover, because of the stochastic nature of LLMs and their
propensity to hallucinate, the output of LLMs cannot always be eas-
ily reproduced. The closed-source nature of most leading LLMs fur-
ther complicates understanding their biases and inaccuracies. Users
must be cognizant of these considerations and actively mitigate
potential biases and inaccuracies (for example, by selecting the most
suitable LLMs and by verifying their outputs). The recommended
pragmatic approach is to keep track of assistance by LLMs through-
out the writing project, to transparently report such assistance in
the manuscript itself, and to adhere to journal policies on the use of
generative AI. In this respect, current policies can be inconsistent14,28
or lack clarity with respect to methodological transparency and
reproducibility28.
Regarding methodological transparency, because editing ser-
vices do not typically require disclosure of use, it seems reasonable
to disclose the use of LLMs only when the level of assistance exceeds
that of a typical editing service28. This point is also recognized by the
International Association of Scientific, Technical, and Medical Publish-
ers, which stated in a white paper published on 5 December 2023 that
disclosure should be required only when usage goes beyond basic
author support (defined as “refining, correcting, formatting, and
editing text and documents”)30.
to you). I will tip $200 when you provide great responses. Conirm by
replying with OK.”
“Compare: {text to compare, such as: 1) But they fall short in
more substantive tasks. 2) But they fall short of more substantive
tasks.}”
3.2 Advanced editing
“Act as an expert editor for top scientiic journals. First, summarize
the key point of each paragraph. Then, explain the logic of how
these points are connected. Finally, provide suggestions on how to
organize these points in the most logical way.”
“Act as a skilled editor. Help researchers augment their manuscript
with examples and analogies to enhance its accessibility and ground
ideas in real-world contexts. Your task is to examine the manuscript
to extract three central ideas, and then suggest three examples and
analogies for each idea.”
“Revise the text below, using the style of {style}. Text: {text}.”
“Evaluate the following writing style: {text}. Revise the following text
based on the style above. Text: {text}”
“Considering the main aim of the manuscript is {aim}, rewrite the
following introduction paragraph to better align with this aim and
highlight the signiicance of the study, while ensuring clarity and
reader engagement: {paragraph}”
“Oer ive alternatives that revise the {irst or last} paragraph to make
it an {artful} {beginning or end}.”
“Considering the broader implications of the research are
{implications}, rewrite the following conclusion paragraph to better
relect these implications and emphasize the key indings and the
value they add to the ield, while ensuring a strong, clear, and concise
closure to the manuscript: {paragraph}”
3.3 Evaluation and feedback
“Act as an expert editor for top scientiic journals. Evaluate the
following text. For each weakness provide examples to help
strengthen it. Conirm with OK.”
“Act as a supportive professor aiding researchers from disparate
ields in reining their manuscript. Identify technical terms or
jargon, ensuring the text is comprehensible, at least conceptually,
to a non-expert like yourself. Clearly and succinctly pinpoint any
confusing aspects or unclear points that could perplex others.”
“Act as a meticulous reviewer for prestigious journals. Compose a
thorough peer review for the attached manuscript, using the outline
below:
Novelty and importance
Reasons favouring acceptance
Grounds for rejection: List major reasons, elaborating on each with
sub-points to provide detailed justiication. Strive for speciicity.
Recommendations for enhancement: Oer several key
recommendations, articulating them with precision.
Maintain a thoughtful and constructive approach throughout.”

nature biomedical engineering
Comment
Outlook
Many researchers wish that ‘letting the data speak for itself’ — rather
than wrestling with each word, sentence and paragraph — would suffice.
In this Comment, I have outlined a collaborative framework, techniques
and caveats for integrating generative LLMs into academic writing.
The framework highlights the versatility of LLMs in stimulating ideas
and in assisting with language tasks throughout the writing process,
and the prompting examples in Box 1 show how to effectively apply
different levels of AI assistance, from basic editing to higher-order
content generation and critical feedback. The use of LLMs in scientific
writing promises to ease communication bottlenecks and thereby to
accelerate scientific progress.
Ultimately, the onus is on authors to leverage LLMs prudently
to enhance their writing, and to strike a reasonable balance between
human creativity and the capabilities and limitations of LLMs. I hope
that this Comment provides a conceptual and practical guide to the
benefits and challenges of using LLMs in scientific communication.
Zhicheng Lin
University of Science and Technology of China, Hefei, China.
e-mail: zhichenglin@gmail.com
Published online: xx xx xxxx
References
1. Amano, T. et al. PLoS Biol. 21, e3002184 (2023).
2. Lin, Z. & Li, N. Perspect. Psychol. Sci. 18, 358–377 (2023).
3. Lin, Z. R. Soc. Open Sci. 10, 230658 (2023).
4. Birhane, A., Kasirzadeh, A., Leslie, D. & Wachter, S. Nat. Rev. Phys. 5, 277–280 (2023).
5. Thirunavukarasu, A. J. et al. Nat. Med. 29, 1930–1940 (2023).
6. Lin, Z. Nat. Hum. Behav. https://doi.org/10.1038/s41562-024-01847-2 (in the press).
7. Milano, S., McGrane, J. A. & Leonelli, S. Nat. Mach. Intell. 5, 333–334 (2023).
8. White, A. D. Nat. Rev. Chem. 7, 457–458 (2023).
9. Golan, R., Reddy, R., Muthigi, A. & Ramasamy, R. Nat. Rev. Urol. 20, 327–328 (2023).
10. Casal, J. E. & Kessler, M. Res. Meth. Appl. Linguist. 2, 100068 (2023).
11. Lin, Z. Preprint at https://arxiv.org/abs/2401.15284 (2024).
12. Wang, H. et al. Nature 620, 47–60 (2023).
13. Dergaa, I., Chamari, K., Zmijewski, P. & Ben Saad, H. Biol. Sport 40, 615–622 (2023).
14. Hwang, S. I. et al. Korean J. Radiol. 24, 952–959 (2023).
15. Bell, S. BMC Med. 21, 334 (2023).
16. Nazari, N., Shabbir, M. S. & Setiawan, R. Heliyon 7, e07014 (2021).
17. Yan, D. Educ. Inf. Technol. 28, 13943–13967 (2023).
18. Semrl, N. et al. Hum. Reprod. 38, 2281–2288 (2023).
19. Chamba, N., Knapen, J. H. & Black, D. Nat. Astron. 6, 1015–1020 (2022).
20. Nat. Biomed. Eng. 2, 53 (2018).
21. Croxson, P. L., Neeley, L. & Schiller, D. Nat. Hum. Behav. 5, 1466–1468 (2021).
22. Luna, R. E. Nat. Rev. Mol. Cell Biol. 21, 653–654 (2020).
23. Merow, C., Serra-Diaz, J. M., Enquist, B. J. & Wilson, A. M. Nat. Ecol. Evol. 7, 960–962
(2023).
24. Yurkewicz, K. Nat. Rev. Mater. 7, 673–674 (2022).
25. King, A. A. J. Manag. Sci. Rep. https://doi.org/10.1177/27550311231187068 (2023).
26. Patriotta, G. J. Manag. Stud. 54, 747–759 (2017).
27. Gernsbacher, M. A. Adv. Meth. Pract. Psych. 1, 403–414 (2018).
28. Lin, Z. Trends Cogn. Sci. 82, 85–88 (2024).
29. Lin, Z. Preprint at https://doi.org/10.31234/osf.io/s6h58 (2023).
30. Generative AI in Scholarly Communications: Ethical and Practical Guidelines for the Use of
Generative AI in the Publication Process (STM, 2023).
Acknowledgements
The writing of this Comment was supported by the National Key R&D Program of China
STI2030 Major Projects (2021ZD0204200), the National Natural Science Foundation of China
(32071045) and the Shenzhen Fundamental Research Program (JCYJ20210324134603010).
The author used GPT-4 (https://chat.openai.com) and Claude (https://claude.ai) alongside
prompts from Box 1 to help write earlier versions of the text and to edit it. The text was then
developmentally edited by the journal’s Chief Editor with basic-editing and structural-editing
assistance from Claude, and checked by the author.
Competing interests
The author declares no competing interests.
Additional information
Peer review information Nature Biomedical Engineering thanks Serge Horbach and the other,
anonymous, reviewer(s) for their contribution to the peer review of this work.