Keeping it personal: How to preserve your voice when using AI

“It has not escaped our notice,” wrote James Watson and Francis Crick in one of the most famous conclusions ever to appear in the scientific literature, “that the specific pairing we have postulated immediately suggests a possible copying mechanism for the genetic material.”

This line is famous precisely because of its understated style and calculated restraint—they’re announcing one of the biggest discoveries in biology while expressing it with typical British reserve.

Now imagine the year is 2025, and the duo turns to ChatGPT:

Prompt:

“Please rewrite this conclusion to be clearer and more direct: ‘It has not escaped our notice that the specific pairing we have postulated immediately suggests a possible copying mechanism for the genetic material.’”

ChatGPT output:

“The base-pairing structure we have described provides a clear mechanism for how DNA might replicate itself, which has significant implications for understanding genetic inheritance.”

ChatGPT’s sentence falls flat. The deliberate understatement—“has not escaped our notice”—is gone, as is the carefully chosen passive construction that creates suspense. The measured tone of Watson and Crick’s original statement, along with the slightly arch academic voice reminiscent of the wood paneling in the Cambridge pub where they first announced their discovery, lets readers discover the importance themselves. Sure, the ChatGPT version is arguably more “readable,” but at the expense of all the authority and drama of the original. It’s like turning a perfectly timed pause in a speech into a simple declarative statement.

Voice—the alchemy of words, tone, syntax and point of view that distinguish one writer’s creation from another—does real work in scientific communication. Voice helps readers follow the thread of an argument, sense where certainty ends and speculation begins, and grasp why a writer thinks a finding matters. In this way, voice isn’t just another writing technique but the imprint of the human mind on the page. When artificial intelligence sacrifices voice to the altar of clarity, something essential is lost. At the same time, excessive voice, especially in scientific writing, can get in the way of the clear expression of complicated ideas. No one would argue that scientific literature, as a genre, is a paragon of clarity. So how can scientists leverage AI’s workmanlike prose while preserving enough of their voice to make the writing recognizably their own?

Large language models (LLMs) are essentially really complicated autocomplete systems tuned to please the maximum number of users. As such, they default by design to middle-of-the-road word choices and unimaginative syntax. But their training data are so extensive that they can be steered in wildly different stylistic directions. Case in point—an early viral post instructing ChatGPT to write a biblical verse in the style of the King James Bible that explains how to remove a peanut butter sandwich from a VCR. As the output shows, taking some time up front to instruct an LLM how to write something can pay off later.

H

ere’s one strategy to preserve your voice when working with AI, which I’ll demonstrate by walking through a little experiment. I gave the AI the following task: Write the first paragraph of a column called “Does AI make scientists worse writers?” that examines whether AI assistance could erode core writing and thinking skills, being sure to reference the introduction of the calculator as an analogy. Then, I asked AI (Claude Sonnet 3.5 in this case) to do the task twice. In one chat, I left the AI to write in its default style, and in a different chat, I carefully specified the writing style I wanted (based on an analysis of this column’s voice; see “How to create custom style instructions in Claude” below for full methods). The difference in voice is striking:

No style guidance

As artificial intelligence tools like ChatGPT become ubiquitous in academic settings, a growing chorus of writing instructors and senior researchers are raising alarm about their impact on scientific communication skills. The parallel to calculators’ effect on mathematical thinking is striking—while these tools undeniably boost productivity, they may also be eroding scientists’ ability to develop and articulate complex ideas independently. Early observations from writing centers at major research universities suggest that manuscripts drafted with AI assistance often lack the logical flow and deep analytical reasoning that characterize strong scientific writing. Yet others argue that AI writing tools, when used thoughtfully, could actually enhance scientists’ communication abilities by serving as collaborative thought partners and freeing up cognitive resources for higher-level synthesis. As the scientific community grapples with this technological shift, the key question becomes not whether to use AI writing assistance, but how to harness it while preserving the essential skills that underpin scientific discourse.

Custom style guidance

The first time I reviewed a manuscript clearly written with AI assistance, the telltale signs jumped out immediately—flawless grammar paired with strangely generic phrasing, technical precision that somehow felt hollow. As both a journal editor and writing instructor, I’ve watched AI tools transform scientific writing from a craft honed through practice into something more automated. Yet the implications run deeper than just smoother prose. When we outsource the writing process, do we also outsource the thinking process that traditionally accompanied it? Just as calculators reshaped how students develop mathematical intuition, AI writing assistants may be fundamentally changing how young scientists learn to structure arguments, synthesize ideas, and develop their scholarly voice.

The first reads like an academic review—comprehensive but detached. While technically sound, it lacks the immediacy that draws readers in. The second version better follows the column’s style by anchoring the discussion in concrete experience: a journal editor encountering AI-written papers. The personal framing (“The first time I reviewed …”) creates a natural way to explore the deeper implications. Rather than cataloging concerns, it leads readers through the thought process: from surface observations about language to questions about cognition. The calculator analogy appears in both versions, but in the second it emerges naturally from the narrative rather than feeling dropped in. The first version is indistinguishable from boilerplate text generated by AI. The second feels distinctly human, or at least a lot closer.

T

o train the model in your own voice, consider excerpts of strong passages from your grants and papers, so long as you have permission to input them into the AI system you’re using. A note of caution: Although this example demonstrates how to preserve voice, using simple prompts to generate full paragraphs from scratch (complete with hallucinated experiences of me as a journal editor) isn’t how I recommend using AI. This use runs counter to my core advice to tap AI as a collaborative partner or to intervene in discrete phases of the writing process, such as by suggesting structure or offering feedback on a draft. But the experiment shows how custom instructions can help AI mimic rather than flatten your preferred writing style.

Beyond custom instructions, additional strategies can help preserve voice while working with AI. Rather than using AI as a writer, use it as an analyzer. Have it flag places where clarity could be improved, and ask the AI to explain why—then fix the issue your own way. When you do need AI to generate text, make targeted requests. I’ve found that vague requests such as “improve” or “clarify” often lead to flattening; instead, frame your prompts around maintaining specific elements of style. Ask the AI to “suggest alternative phrasings while maintaining my measured tone” or “help me make this more concise without losing the deliberate understatement.” Finally, when using AI-generated text, systematically revise to restore voice. Reintroduce strategic tentativeness where appropriate (“we suggest” rather than “this shows”), restore personal constructions that serve a purpose (“we reasoned” rather than “it was determined”), ensuring your draft retains the stance and meaning you intended.

Voice in scientific writing isn’t just a matter of literary flourish—it’s a way to transmit to the reader the writer’s intention. When Watson and Crick wrote “it has not escaped our notice,” they weren’t just being stylish; they were using voice to signal the weight of their discovery while maintaining scientific restraint. It might even have been, in today’s parlance, a sort of humblebrag of their success where others had failed. In an age when AI makes grammatically perfect prose trivially available, these subtle markers of human judgment may be what ultimately distinguish the most effective papers from the merely competent.

Claude, like many AI assistants, can be given specific writing instructions. I like Claude for this particular use because it offers “Projects”—persistent spaces where you can upload reference documents (“project knowledge”) and set custom instructions that apply to all interactions. Think of it as creating a specialized version of the AI that’s tweaked to your needs. (Note: After I prepared this column, Claude released a new feature to configure custom writing styles. Although I found the method below to give better results, the built-in feature is much easier to implement and will probably improve over time.)

Here’s how to create style-specific instructions:

Step 1: Create a style analysis prompt

I first asked Anthropic’s prompt generator to create a systematic way to analyze writing style. I asked the generator “to produce a prompt that will take as input some writing samples, analyze them for stylistic elements, and then produce a second prompt that I can use to replicate that writing style. Use a few-shot approach with exemplars from the provided text.” The resulting prompt (shown below) breaks down the analysis into specific elements such as tone, sentence structure and narrative voice.

You are tasked with analyzing a piece of writing, identifying its stylistic elements, and creating a prompt that can generate text in a similar style. Follow these steps carefully:

1. Analyze the following text:

<text_to_analyze>

{{TEXT_TO_ANALYZE}}

</text_to_analyze>

2. Examine the writing style of the text. Consider elements such as:

– Tone (formal, informal, humorous, serious, etc.)

– Sentence structure (simple, complex, varied, etc.)

– Vocabulary (simple, advanced, technical, colloquial, etc.)

– Literary devices (metaphors, similes, alliteration, etc.)

– Narrative voice (first person, third person, etc.)

– Pacing and rhythm

– Any unique or distinctive features

3. Identify 3-5 key stylistic elements that define this piece of writing.
4. Select 2-3 short exemplars (phrases or sentences) from the text that best demonstrate these key stylistic elements. These will be used as few-shot examples in the prompt.
5. Develop a prompt that would guide an AI to produce writing in a similar style. The prompt should:

– Describe the overall tone and style

– Specify the key stylistic elements to emulate

– Include the selected exemplars as few-shot examples

– Provide clear instructions on how to apply these elements in new writing

6. Present your analysis and the resulting prompt in the following format:

<analysis>

Key Stylistic Elements:

1. [Element 1]
2. [Element 2]
3. [Element 3]

(etc.)

Selected Exemplars:

1. “[Exemplar 1]”
2. “[Exemplar 2]”
3. “[Exemplar 3]” (if applicable)

</analysis>

<generated_prompt>

[Your generated prompt here]

</generated_prompt>

Ensure that your generated prompt is clear, specific, and provides enough guidance for an AI to effectively mimic the style of the original text.

Step 2: Analyze your writing style and create style instructions

I added previous columns from this series to Claude’s project knowledge and fed it the Style Analysis prompt. In the {{TEXT_TO_ANALYZE}} section, I simply asked the bot to reference the columns I had already uploaded to the project knowledge. Claude analyzed these writing samples and generated specific Style Instructions, shown below:

Write in the style of an accessible technical article for an educated professional audience. The writing should:

Tone and Structure:

• Maintain a scholarly yet approachable voice that bridges academic and practical perspectives
• Begin with broad concepts before diving into specific applications
• Use first-person perspective strategically to share insights from experience
• Address readers directly when providing practical guidance

Technical Content:

• Explain complex concepts using clear, precise language
• Include concrete examples to illustrate abstract ideas
• Acknowledge both benefits and limitations of discussed approaches
• Connect theoretical frameworks to practical applications

Writing Techniques:

• Break down complex topics into digestible sections
• Use analogies to make technical concepts more accessible
• Incorporate strategic repetition to reinforce key points
• Balance technical terminology with plain language explanations

Format your writing similar to these examples: “Rather than diving straight into implementation, consider first what makes this approach valuable. A blank page can be intimidating, but an unclear direction can be equally challenging.”

“When applying this framework, think of it as building blocks rather than a rigid structure. Start with the foundation [explain basic concept], then gradually add complexity as needed [provide specific application].”

Remember to:

• Lead with the practical value before diving into technical details
• Include specific, actionable guidance
• Acknowledge potential challenges while offering solutions
• Maintain a teaching tone without being condescending

Step 3: Implement the instructions

The final step was to paste these Style Instructions into the “project instructions” field. These will now act as a permanent style guide for all chats in that project, guiding Claude to maintain the voice of this column in future interactions.