Fine-Tuning Generative AI for Content Creation

As part of our mission for AI Daily, we hoped to learn how to leverage generative AI as a collaborator as we create meaningful content for the campus community. 

Initially, we found that our prompts didn’t yield content that matched the style, tone, and quality we hoped to achieve. Our editorial team spent hours editing AI-generated content, resulting in more effort than might have been required to write the article without AI.

To remedy this, we used generative AI to create the AI Daily Correspondent, a custom GPT built to draft content that adheres to our standards for style and quality. We describe our process in the following case study. 

Our Objective

We aimed to address specific challenges in AI-assisted content creation:

• Aligning first drafts with CMU’s editorial voice and standards.
• Reducing the amount of editing required and improving workflow efficiency.
• Ensuring drafts were engaging, structured, and easy to read for the web.

Refining our AI Writing Assistant

1. Analyze Differences Between AI and Human Writing

To begin, we used ChatGPT to try to better understand the differences between an AI-generated first draft and a human-edited final version. In a single prompt, we included both versions of a single article and asked ChatGPT to identify key differences in style and tone.

The AI noticed that the human-edited final draft adhered to the following guidelines:

• Clarity: Organize content into thematic sections with descriptive headers.
• Conciseness: Avoid redundancy and keep sentences focused.
• Emphasis: Highlight key points using bullets, bold text, or lists.
• Engagement: Use a conversational tone to connect with readers.
• Active Voice: Prioritize active sentences for clearer, more direct communication.

2. Creating a Custom GPT for AI-Assisted Writing

Next, we created a GPT, or customized chatbot built for specific tasks, to adhere to the style guidelines demonstrated in our polished articles. We instructed the GPT, named the AI Daily Correspondent, to follow the guidelines every time we prompt it to draft an article.

While we used a GPT, which is a feature of ChatGPT Teams and other paid OpenAI accounts, we could have also created a reusable prompt that included the style guidelines to use with each new request. 

For example, we could prompt other tools, like Microsoft Copilot, Google Gemini, or a free ChatGPT account to “write an article on [topic] and adhere to the following style guidelines:”

Results

Since creating our AI Daily Correspondent GPT, we’re able to produce higher quality first drafts, resulting in less editing time and greater efficiency. The AI’s output more consistently matches Computing Services’ editorial standard for clarity, readability, engagement, and structure.

Apply This Approach to Your Projects

This project demonstrates the potential of fine-tuned generative AI to complete specific tasks in a way that aligns more with pre-defined standards. By instructing the AI to analyze the differences between an unsatisfactory version and a polished version, we could extract rules to help in future prompts.

We encourage you to try this approach, whether for content creation or another purpose, to help use generative AI tools more effectively in your work.