Using Claude and ChatGPT Effectively for Scientific Writing
A practical guide to using large language models for scientific writing tasks — covering what they genuinely help with, where they introduce risk, and how to prompt them to get useful output rather than generic text.
What AI writing assistants are actually good at
AI writing assistants (Claude, ChatGPT, Gemini) are not research tools — they don’t search literature, don’t have access to your data, and can generate plausible-sounding but incorrect scientific claims. What they are good at is improving the expression of ideas you already have.
The tasks where they add genuine value:
- Restructuring text: turning a dense paragraph into a cleaner flow, or reorganizing sections of a Methods
- Clarity revision: identifying sentences that are ambiguous, overlong, or unclear to a non-specialist reader
- Abstract drafting: given a clear summary of your study, drafting an abstract structure to edit from
- Cover letter drafting: journal cover letters follow a predictable structure that AI handles well
- Response to reviewers: drafting the polite framing for point-by-point responses (you supply the scientific substance)
- Translation support: for non-native English speakers, improving fluency and idiomatic phrasing
- Standardizing terminology: flagging inconsistent use of terms across a long manuscript
The tasks where they introduce significant risk:
- Generating factual claims about the literature (hallucination risk)
- Citing papers (will fabricate citations)
- Describing methods they don’t know (may generate plausible but incorrect protocols)
- Interpreting your results (may confabulate explanations)
The most important principle: you provide the science
The safe and effective pattern is:
You → AI → You
You write a draft with correct scientific content. The AI helps you improve its expression. You review the AI’s output and verify that nothing has changed scientifically — only stylistically.
Never ask an AI to write scientific content from scratch based on vague instructions. The output will be generic and may contain factual errors that are difficult to spot if you’re not already expert in the area.
How to write prompts that get useful output
For revision and clarity
Don’t paste text and say “improve this.” Instead, specify the problem:
Weak prompt:
Improve this paragraph.
Better prompt:
This paragraph is from the Discussion section of a paper on enzyme kinetics. The audience is specialist biochemists. The paragraph is too long and the main point (that we observe substrate inhibition) is buried in the third sentence. Rewrite it so the main finding is stated first and supporting detail follows. Do not change any of the factual content.
The key elements to include:
- Context (what section, what field, what audience)
- The specific problem you want fixed
- Constraints (don’t change facts, keep under X words, maintain passive/active voice, etc.)
For abstract drafting
Prompt:
Draft an abstract for the following study. Use this structure: one sentence on the problem/gap, one sentence on what we did, two sentences on key results (listed below), one sentence on the main conclusion. Keep it under 250 words. The factual content is mine — only help with expression.
Problem: [your description] Methods: [your description] Results: [your actual results] Conclusion: [your interpretation]
Then edit the draft heavily. The AI’s abstract will give you a structure to work from; your job is to make it accurate and precise.
For response to reviewers
Prompt:
I am writing a response to Reviewer 2’s comment: “[paste the reviewer comment]”
My scientific response is: [your actual response — the scientific substance]
Draft the polite framing for this response — thank the reviewer, state that you have addressed the concern, and lead into the scientific response. Keep it professional and not sycophantic.
You supply the science; the AI handles the diplomatic scaffolding.
Practical prompting patterns
Specify the audience
“Explain this to a specialist in X” and “Explain this to a non-specialist” produce very different outputs. Be explicit.
Ask for options
“Give me three alternative ways to open this paragraph” gives you choices to pick from rather than committing to one revision. Often the right answer is a combination.
Ask it to identify problems, not just fix them
“What are the weakest points in this paragraph?” or “What questions might a skeptical reviewer raise about this methods section?” can reveal issues you’ve stopped seeing after reading the text too many times.
Set hard constraints
“Do not add any information not already in the text” is a critical constraint when accuracy matters. Without it, the AI may introduce “helpful” elaborations that are wrong.
Use it to check your own text
“Does this sentence say what I think it says? I intend to convey [X].” AI can help identify unintended ambiguity.
Workflow: revising a manuscript section
- Write the section yourself with all scientific content in place — methods, numbers, citations, interpretations
- Read it yourself first and mark what you think is unclear or needs improvement
- Paste to AI with a specific prompt targeting the issues you identified
- Read the AI output carefully — does it preserve all the factual content? Has it changed any numbers, added any claims, or removed important caveats?
- Accept selectively — use tracked changes or side-by-side comparison. Don’t accept whole-section rewrites; accept sentence-level improvements you can verify
- Final read of the revised version — does the scientific meaning match your original intent?
Disclosure and ethics
Check your journal’s policy. Policies vary widely:
- Some journals require disclosure of AI assistance in the methods or acknowledgements
- Some prohibit AI-generated text entirely (particularly for peer review)
- Most allow AI for language editing with appropriate disclosure
The default position from most major publishers: AI tools may assist with writing, but the authors are fully responsible for accuracy. Hallucinated citations or incorrect facts generated by AI and not caught before submission are the authors’ responsibility.
Never list an AI as an author. Authorship requires accountability; AI tools cannot be held accountable.
Tools compared for scientific writing
| Tool | Strengths | Notes |
|---|---|---|
| Claude | Handles long documents well; follows complex multi-part instructions accurately; strong for revision tasks | Best for longer manuscripts and detailed revision prompts |
| ChatGPT (GPT-4o) | Strong general performance; widely integrated into other tools | Similar capability to Claude for most writing tasks |
| Paperpal | Purpose-built for academic writing; built-in grammar checking and journal-specific formatting | Less capable for complex revision tasks but more automated |
| Gemini | Good Google Docs integration; reasonable writing assistant | Useful if your workflow is Google Docs-based |
Related content
- Tutorial: AI-Assisted Writing and Revision: Drafting, Editing, and Getting Unstuck
- Tool: Paperpal — purpose-built academic writing assistant
- Tool: Perplexity — for background research before writing
- Glossary: Hallucination — understanding why you need to verify AI output
- Glossary: Prompt Engineering