At Scale, we're navigating the frontiers of human-AI collaboration. This post is the first in our series about writing effectively with LLMs.
“All I have is a voice.”1
― W.H. Auden
How do we collaborate with LLMs while maintaining our own unique voices? Highlighting LLM-generated or LLM-revised text here, I will explore this question and demonstrate my own process.
When GPT-3.5 launched in 2022, it seemed suddenly as if AI had figured out this writing thing and human writers could become obsolete.2 But it didn’t take long for people to realize that these tools have limitations, one of these limitations being that large language models (LLMs) sound, well, like LLMs. Today, LLMs are mainstream, with a majority of American adults having tried them.3, 4 Though detectors are limited in their efficacy, those who use LLMs can often tell when content comes straight from an LLM.5, 6 This matters because readers remain positively biased toward human-created writing, while typically responding neutrally or negatively if they feel the content they are reading is LLM-generated (especially in sensitive areas like healthcare).7, 8 However, when humans collaborate with LLMs, readers typically cannot tell and allow writers to benefit from the positive human-created writing bias while benefitting from using LLMs.9, 10
At a time when, for example, over half of long posts on LinkedIn are likely LLM-generated, having an authentic, human voice allows writers to harness readers’ positive bias toward human-created work.11 Throughout this post, I will demonstrate how I collaborate with LLMs while maintaining my own voice, weaving LLM content (highlighted) with my own. As I write, I treat the LLM as a writing partner – brainstorming, outlining, and revising. By showing my own process, I hope to demonstrate how effective writing with LLMs combines the efficiency of these tools with the irreplaceable value of human voice.
Real Differences
To effectively harness AI while preserving our unique voices, it is worth taking a look at the recognizable ways LLM-generated text differs from human writing. These differences are varied, but often grammatical (overused structures), lexical (overused words), and stylistic (predictable patterns), all impacting voice.12, 13, 14 If you look at the first sentence of this paragraph, for instance, I took an LLM-generated sentence and made it more aligned with my voice. The original: To effectively harness AI while preserving our unique voice, we must first understand what separates machine-generated text from human writing. “We must first understand” is too formal for my voice so I replaced it with “it is worth taking a look at.” I also swapped out “machine” in favor of “LLM” for consistency.
There is, of course, variation between models’ writing styles. Some are designed with warmer, more personal styles or to be more clinical.15 There is also at least one model designed with a little bit of attitude.16 The overarching goal of most LLMs, however, is to generate content in a predictable style that is the least disagreeable for the greatest number of users. However, if you have (or are trying to have) your own audience, it benefits you to write for them in a way that sounds authentically like you.
Understanding Your LLM Writing Partner
Regardless of personality, most commercial LLMs prioritize efficiency, and providing the most information in the least amount of space does make sense, particularly for certain business contexts.17 However, our voices are defined by strategic inefficiencies. When we communicate naturally, even in professional settings, we take deliberate detours to our points, create meaningful pauses, repeat key information for emphasis, and shift our diction or syntax. Because of this, when collaborating with an LLM, we need to guide it away from its efficiency bias. In an effort to be helpful, LLMs will revise your voice out of your work unless you carefully prompt it not to do so and carefully revise when it does. Remember: asking LLMs to mimic your voice works against their fundamental design; more often than not, you will need to tweak what the LLM produces or create a sort of collage, using various strings of text as I am doing here.
That being said, LLMs are effective at creating categories of style—they can write formally or informally, technically or conversationally. What they can't do is easily replicate the myriad unique combinations of choices that compose an individual writer’s voice. They are more capable of mimicking general patterns but less so the subtle variations and intentional breaks from those patterns that make your writing sound like you.
An Extended Conversation
When I collaborate with LLMs, I lead. We begin by brainstorming and bantering about the topic. I throw out ideas and see what the LLM notices are particularly strong or novel ideas. For outlining, I might request different structural approaches or present my own framework for feedback, constantly nudging the LLM toward my vision. As I draft, I often establish ideas and directions with partial sentences or paragraphs letting the LLM continue, then critically evaluating its suggestions, revising, and editing as needed. I request the changes in boldface to clearly see them, and I push back when the writing is poor, the feedback is weak, or the responses are sycophantic.18, 19 Throughout the entire process, I ensure the writing remains mine; I do not take anything from the LLM uncritically, and most of what it provides is cut up and seamlessly aligned with my organic text.
Does Voice Still Matter as Models Improve?
As models improve, it is tempting to think that infusing one’s own voice into their content would be less important. Since I began this piece several months ago (forgive me, we’ve been busy at Scale) several new models have emerged that are even better than the ones I started with. You can imagine my dismay when on the cusp of completing this piece mostly using Claude 3.5 Sonnet and Gemini 2.0 Pro Experimental (along with GPT o3-mini), that Claude 3.7 Sonnet and Gemini 2.5 Pro Experimental were released. Since these models are better writers than their predecessors, I questioned whether this effort was all for naught, so I asked the latest models: is this still worth doing? Their responses were refreshingly insightful:
Because LLMs can be sycophantic in their effort to be helpful, I wanted to make sure they weren’t just telling me what I wanted to hear. So I asked them: “Were you being sycophantic? I need you to reflect critically on whether or not this is worth doing.” Its answers, though more nuanced, remained in the affirmative. In spite of models getting closer to sounding human, it remains important as ever that it sounds like a specific human. As the technology advances, so must our approach to collaboration with it.
Strategies and Refinements
This section was the hardest for me to write. I had already worked through several versions and was running out of ideas about how best to further demonstrate my main points. This made it an ideal time to lean on an LLM for support. My plan was to have it give me a rough outline and some text to work with that I could then rearticulate through my own perspective and voice.
I told the LLM my problem: “I am writing an essay on how to maintain voice with LLMs. It is largely finished except for one middle section specifically focused on examples that require screenshots of LLM usage. I need you to help me with it. You must keep in mind the focus of this is specifically around creating content with LLMs while maintaining an authentic, authorial human voice. Here is my draft with an indication of the section I need help with.” I also made sure to label where this section should be: “[this is where the example section that we are working on should go]”
It gave a thorough outline and when I asked it to draft my first paragraph, it suggested this:
The process described above – brainstorming, outlining, drafting, revising – involves specific tactics when interacting with the LLM to guide it towards your voice, or at least closer to it. It's crucial to remember that you're not just prompting for content, but for content delivered in a manner that aligns with your style. Here are a few common strategies I employ, along with examples demonstrating both their utility and their inherent limitations.
This paragraph is well-written, informative, and sets up this section perfectly. There is nothing explicitly wrong with it, but when the writing sounds flat and uninspired like this, it is easy for a reader’s attention to drift. Voice makes information breathe and inspires readers to keep reading. So what were my options? My impulse was to simply edit or rewrite it in my own voice, while adding my own ideas to flesh it out. But what if I wanted the LLM to do that for me? How could I push it to sound more human, more like me?
Since I had a solid example of what I wanted this writing to sound like, I gave that sample (the first half of this post) to two LLMs and asked them to write short, but detailed paragraphs describing only the voice. I then took the answer from one LLM and gave it to the second, asking “is there anything in this description that your description is missing? Give me a short paragraph that is instructive for an LLM to write like that and includes all the details.”
The instructions:
The results, though a bit longer, are generally better and sound closer to how I sound. While they don’t replicate the voice I’ve been attempting to illustrate, they do sound less like the flat, generic output we started with:
Though neither of these two responses truly sound like me, they are notably better than the original.
I feel most aligned with option 1, but in order to sound like me, first I’ll strip away any excess content. What words are included here that I would never use myself? What phrases feel too familiar, like maybe I’ve heard them before? What just feels extra?
Regarding the process I've described earlier—this flow encompassing brainstorming, outlining, drafting, and revising—my interaction with the LLM at each stage involves rather specific tactics. The aim, really, is to consciously guide the model's output, nudging it closer, if not perfectly aligned, with my own authorial voice. I find it crucial to remember here that we're not just prompting for raw information, but for information articulated in a way that resonates with our unique stylistic signature. With that reflective understanding in mind, let me walk through a few common strategies I personally rely on. As we'll explore together, these techniques offer tangible benefits in steering the collaboration, yet it's equally important, I believe, to remain transparent about their inherent limitations—this is an ongoing dialogue with the technology, after all.
Now that I’ve made my cuts–removing any hedging, excessive formality, and conversational filler–I’ll work around the remaining text and simply fill in the blanks, while making additional cuts and rewrites as necessary in order to match my concise, direct, active, and purpose-driven style:
Regarding the process I've described earlier, my interaction with the LLM involves specific techniques to guide the model's output closer to my own authorial voice. Here I will demonstrate a few key prompting and revision strategies, showing how methods like using distinct rounds for information and then for voice help move towards outputs that resonate with our unique style. Though these techniques offer benefits in steering the collaboration—this is an ongoing dialogue with the technology.
Even though I’ve abandoned the initial structure the LLM suggested for me, I found within its response something I could work with and make my own. By finding just that one introductory paragraph draft, that had the information I needed, I was able to work through exactly what I was thinking about this process, and use an LLM to help me along the way. Though this is a long way from one- or even two-turn prompting, I have landed on a piece of writing that was only possible through creative collaboration.
Engaging deeply with the draft provided to me by the LLM—analyzing, prompting further, and ultimately applying significant editing—allowed me not only to clarify my own thinking, but push the boundaries of my own capabilities. This iterative workflow, far from simple prompting, led me to something that feels–and sounds–distinctly mine.
Why Human Writing Matters More Now
At a time when we are saturated with AI-generated content, the ability to write with a voice that is at once distinctive, engaging, and unmistakably human is not just a symbolic aesthetic choice but a competitive necessity. My aim throughout this piece has been to demonstrate how an LLM can support my writing process, maintaining my own authentic voice even as I used content directly taken from LLM outputs. When we write, we should remember that readers crave connection and want to witness our journeys of understanding, writing not just what we know, but what we are actively figuring out. In the LLM era, there are always new fears that our value as humans will decrease. Remember the distinct possibility that the opposite, in fact, is true.
Sources:
-
The New Generation of A.I. Apps Could Make Writers and Artists Obsolete
-
Over half of American adults have used an AI chatbot, survey finds
-
The Widespread Adoption of Large Language Model-Assisted Writing Across Society
-
Detecting LLM-Generated Text in Computing Education: A Comparative Study for ChatGPT Cases
-
The effect of source disclosure on evaluation of AI-generated messages
-
LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?
-
"It was 80% me, 20% AI": Seeking Authenticity in Co-Writing with Large Language Models
-
Over ½ of Long Posts on LinkedIn are Likely AI-Generated Since ChatGPT Launched
-
Do LLMs write like humans? Variation in grammatical and rhetorical styles
-
Beware of Words: Evaluating the Lexical Diversity of Conversational LLMs using ChatGPT as Case Study
-
Contrasting Linguistic Patterns in Human and LLM-Generated News Text