Teaching AI to Write Like Us: The Art of Refining AI-Generated Text 

By Richard Sebaggala

 

 

While traveling on a bus, I had a rare moment of uninterrupted reading.I found myself drawn to an article discussing a fascinating study on the differences between AI-generated and human-written text. Conducted by researchers at Carnegie Mellon University, the study attempted to pinpoint stylistic differences in how large language models (LLMs) like ChatGPT and Llama construct sentences compared to human writers. What made it particularly compelling was its relevance to an ongoing shift in the conversation about generative AI.  Not long ago, discussions about AI in writing were dominated by concerns over misconceptions, academic plagiarism, and ethical dilemmas. However, the debate has evolved. Today, a pressing question is: how much content is AI-generated, and can we truly differentiate it from human writing?

 

The study findings were fascinating. The study revealed that AI-generated writing tends to rely more on structured, information-dense phrasing, with certain quirks that set it apart from human expression. For example, LLMs often overuse present participle clauses (e.g., "Bryan, leaning on his agility, dances around the ring."), favor abstract nominalizations (e.g., "implementation of the policy" instead of "the policy was implemented"), and show strong preferences for specific vocabulary choices—words like “camaraderie,” “tapestry,” and “palpable” appeared far more frequently than in human writing.

 

While these insights are valuable, they raise an important question: Are these differences a limitation of AI, or simply a result of how we use it? A closer look at the study's methodology suggests that some of its conclusions may be shaped by underlying assumptions rather than definitive proof that AI cannot adapt to human writing styles.

At first glance, the study's results seem to confirm a widespread belief that AI-generated writing lacks the nuance and adaptability of human expression. However, a critical examination reveals several methodological flaws that may have influenced its conclusions:

1. A Narrow View of Writing Contexts

n The study evaluated AI writing using specific genres—TV scripts and academic writing—but did not consider other common forms of human writing, such as personal essays, journalism, or business communication.

n Human writing is highly diverse, shaped by context, purpose, and audience. Did the researchers test AI’s adaptability across a broader spectrum of writing styles?

2. The Lack of Iterative Refinement

n One key difference between human and AI writing is the revision process. Writers rarely produce polished work in a single draft; they refine, edit, and restructure.

n The study did not test whether AI writing could become more human-like through iterative refinements, where users provide feedback and make incremental improvements.

3. Prompting Matters More Than We Think

nAI models do not generate text in isolation; they respond to prompts. Poorly designed prompts often lead to unnatural or rigid outputs.

n The study does not clarify whether the AI was prompted in a way that encouraged contextual adaptability. Could different prompt engineering techniques yield more human-like writing?

4. Comparing AI to Expert Writers vs. the Average Writer

n If the AI-generated text was compared to professional authors or seasoned academics, it would naturally fall short in style and depth.

n A fairer comparison would involve a broader range of human writing abilities, including that of everyday writers, students, and non-expert authors.

These gaps suggest that while the study highlights genuine AI tendencies, its conclusions may not fully reflect AI’s potential when used strategically.

Instead of dismissing AI-generated writing as inherently flawed, I decided to test whether ChatGPT could refine its own writing to align more closely with human standards. I asked ChatGPT o1, an advanced reasoning model, to generate a structured prompt for improving AI-assisted writing based on article concerns. The results were impressive:

"I have an AI-generated article that I want to refine to make it more natural and closer to human writing. Please follow these steps to adjust the style:

1. Adapt Sentence Structure

· Avoid overuse of present participle clauses and vary sentence beginnings.

· Mix short and long sentences to create a natural rhythm.

2. Balance Voice and Tone

· Reduce excessive nominalizations (e.g., replace "implementation of the policy" with "the policy was implemented").

· Adjust passive voice usage where necessary to improve clarity.

· Adapt the tone to match the context—formal for academic writing, conversational for blog posts.

3. Diversify Vocabulary

· Replace overly sophisticated or AI-favored words (e.g., "camaraderie," "tapestry," "palpable") with more natural alternatives.

· Use synonyms and rephrase sentences to avoid unnecessary repetition.

4. Enhance Natural Flow

· Use smooth transitions to guide the reader logically through ideas.

· Ensure paragraphs connect seamlessly rather than feeling like separate, disjointed sections.

5. Simulate Human Revision

· Edit the text progressively, focusing on clarity, conciseness, and logical progression.

· Introduce slight variations in sentence structures, as humans naturally do when revising their work."

Applying this prompt transformed the AI-generated text significantly. Instead of rigid, overly structured sentences, the writing became more dynamic and fluid. Ideas flowed naturally, and the vocabulary felt more organic. More importantly, the revision process mirrored how human writers refine their work—by iterating, editing, and fine-tuning for clarity.

 

This experiment revealed a key takeaway: AI is not inherently incapable of writing like humans—it simply needs structured guidance and refinement. Instead of seeing AI as a threat to human writing, we should embrace it as a collaborative tool. Whether in academia, professional writing, or creative storytelling, AI can enhance writing quality when used thoughtfully. However, it should not be treated as a shortcut for bypassing the critical thinking, creativity, and revision processes that make writing truly human.

 

If you use AI in your writing, try refining its output using structured prompts like the one outlined here. The best writing—whether human or AI-assisted—is always a result of careful revision. AI may not naturally write like us yet, but with the right approach, we can guide it to produce work that is technically correct, genuinely engaging, nuanced, and human-like.