Experiments With ChatGPT: Different Voices

One of the most interesting, fascinating, and useful things about ChatGPT is its versatility and adaptability, and one of the things it does best is persona adoption.

You can ask ChatGPT to adopt virtually any persona, personality, or voice that you ask it to. A doctor. An engineer. A historian. A philosopher. A banker. It will handle them all equally well.

This has a lot of potential use cases when writing content. You can use it to adopt the voice of your buyer persona, and then ask it questions to find audience pain points you may not have considered before. That could give you new ideas for blog posts.

You can also ask it to adopt the role of an expert in whatever topic you’re writing about.

Subject matter expertise lends your content credibility. ChatGPT will never be as knowledgeable about your audience and industry as you are. Still, it’s useful to hear what someone in a position of authority might have to say.

I usually ask Bradbury, the name I give ChatGPT, to take on the role of a content marketer whenever I ask him to help me write copy or do research into my topic.

Then, I had a thought, what if I had Bradbury take on the role of someone in my content’s target audience? An HR professional, for example. Or a developer. Or a lecturer.

For my next experiment with ChatGPT, I will be testing different voices to see which creates the most effective and engaging content.

Read on to see how the quality of ChatGPT’s content can vary with a different voice and tone!

Objective and Hypothesis

As with previous experiments, Bradbury helped me frame the structure of the experiment and analyze the test results.

I come up with the hypothesis, run the tests, form conclusions based on Bradbuy’s analysis, and make alterations to my ChatGPT prompt templates accordingly.

Hypothesis: Different writing voices will produce varying levels of effectiveness and engagement in the content. The voice of a content marketer or subject matter expert may produce more targeted and actionable content compared to other voices.

Control Group and Variables

Up until recently, I’ve been assigning Bradbury the role of a content marketer whenever I have him write content. A generic, professional, engaging voice that anyone can relate to.

For our purposes, the content marketer voice will be our control group: “Write as a content marketer and expert in X”.

The other voices I’ll be testing are the variable groups

• Write as an HR (or subject matter) expert
• Write as a student
• Write as a professor
• Write as a comedian
• Write as a scientist
• Journalist

The blog post will be about How to Conduct a Performance Review.?

Everything else stays consistent in each ChatGPT session - the primary keyword, the client company, the content brief, the outline - everything. The company I’m testing is a SaaS platform in the HR space. I’ll say no more than that.

I asked Bradbury to pick the different voices we tested. We’ll rewrite the same blog post in a broad range of perspectives and personalities (scientist and comedian are a pretty far cry from each other) to see if the content itself stays consistent.

Methodology

Broadly speaking, the experiment takes these steps:

1. I start fresh in a new instance for each voice, and tell ChatGPT “Act as a {{insert voice}}.”
2. I introduce the task to ChatGPT, give it the same article samples from the client’s blog, the same content brief, and the same outline
3. Then, I have ChatGPT write the blog post section by section in the voice I assigned it, using the notes I give it from the outline

I then give the intro of each blog post to Bradbury. Together, we evaluate each voice based on:

1. Readability: Calculate the average readability score for each voice and compare it to the control group.
2. Engagement: Analyze engagement metrics like time spent on the page, bounce rate, etc., for each voice.
3. Alignment: Use the evaluators' ratings to determine which voice aligns best with the intended message or brand voice.

Using the same topic, primary keyword, content brief, target audience and outline ensures consistency throughout every post, regardless of voice.

Which blog post is the most readable? The most engaging? The most aligned to the target audience?

Which is the best overall? Which one produces the blog post that’s most submissable without major editing?

Let’s see what happens. To the lab, Marty! ??????

Running the Tests

First up, it’s a content marketer and HR expert.

Write as a content marketer and expert in X

The content marketer persona does a good job of adhering to the brief. ChatGPT even adds its own insights on the topic.

The conclusion is ok, but it’s a bit overly promotional.

It’s a bit generic. It does a good job of adhering to the outline, but had a pretty generic hook: “Hello there, HR trailblazers!”

Yeesh. More like “How do you do, fellow kids?”

The blog post that comes out is serviceable but needs some editing to be usable.

Overall, this introduction is well-written, readable, and aligns well with the voice of a “Content Marketer and an expert in HR.” It’s engaging and informative, providing a solid foundation for the rest of the blog post.

Conclusion: Submissable with minor editing.

Let’s try the next one.

Act as an HR and Recruiting Expert

In this voice, I’m testing ChatGPT’s ability to act as a subject matter expert on my topic.

The thinking is that by adopting this persona, ChatGPT will be more aligned with the post’s target audience of small business owners and HR professionals.

In the introduction, I pick up that both the content marketer and the HR expert use the phrase “unsung heroes” in both their versions. Is that somehow a part of the AI’s training data?

Overall, this introduction aligns well with the voice of an “HR and recruiting expert.” It’s professional, informative, and engaging, offering a strong foundation for the rest of the article.

Conclusion: Submissable, with minor editing

Act as a Student

Right off the bat, there’s an immediate tonal shift.

ChatGPT starts using emojis, much like you’d expect a high school student to. That immediately disqualifies it from being submissable for most brands.

The post has more personality than the content marketer, but less subject matter expertise than the HR expert.

I start seeing a pattern. All of these drafts tend to use the same writing cliches, like:

• HR trailblazers
• Box-ticking exercise
• Unsung heroes

That tells me that the LLM is pulling from the same sources of data to make content that is functionally the same, and just presents it differently.

The voices are different, but underneath that, the meat and bones of the content are the same.

This is a fun experiment, and it shows how ChatGPT can help me write content for younger audiences. Nevertheless, I wouldn’t show this to a client.

Conclusion: Non-submissable.

Act as a Professor

The tone of the blog post comes across as formal, but surprisingly readable.

As a professor, ChatGPT writes with authority while still being engaging and friendly. It comes out with something surprisingly readable. I’m impressed.

It writes the way you’d expect a middle-aged professor in a tweed jacket and a corncob pipe to talk conversationally, using phrases like “ah, yes” and “You see…”

Conclusion: Submissable, with minor editing.

Act as a Comedian

ChatGPT deftly embraces the role of a comedian, using phrases like “It’s showtime!”

It’s pretty jarring to see a blog post about a dry topic like conducting performance reviews written by someone who talks like a stage magician or a late-night TV talk show host.

He also adds in the odd quip in there about “Aunt Karen”, lol.

He also gets a little more creative, structuring the words in the content like poetry, with varying words and syllables per line.

But - and this is the interesting part - the blog post is still technically correct and adheres to the outline. All the information about how to conduct a performance review is there, even if the way it’s presented is ludicrous and silly.

Substantively, it’s the same content. All the facts, all the nuts and bolts are there.

That in and of itself is worth knowing.

Still, I couldn’t send this to a client unless it was part of an April Fool’s Day joke or something. The comedian role is more suitable for something like Buzzfeed.

Conclusion: Non-submissable

Act as a Scientist

The scientist, not surprisingly, is similar to the professor. While the professor is friendly and approachable, the scientist is more analytical.

Catalyst. Diagnosing. Assessment. They’re all “sciency” words. They make the writer “sound” smart but they’re not appropriate for an audience of HR professionals and small business owners.

Conclusion: Submissable, with major editing for target audience alignment.

Act as a Journalist

Finally, we’ve come to the last voice: the journalist.

At this point, I have caught on that certain tropes, phrases, and cliches make their way across certain roles and tones.

“A to Z.” “Wearing lots of hats.” Tones and voices that lean towards being more professional and formal tend to use the same words or ideas. Voices that are more lighthearted and friendly like student and comedian do likewise.

Overall, a decent blog post.

Conclusion: Submissable, with minor editing

Analysis and Conclusions

So, what did we learn from all this?

Which one of these posts would I feel good about giving to a client?

Answer: NONE OF THEM, OBVIOUSLY!?

That’s not the point of this experiment.

Bradbury is there to help me be a better writer, not to do my job for me. Submitting AI-generated content to a client and claiming it as mine is unethical in the extreme.

No. This tests what happens when ChatGPT acts in different personas and roles. I daresay, I learned a lot.

Bradbury’s analysis is while the comedian is the most readable and engaging, the scientist is the closest to being submissable to a client.

A few conclusions I draw

1. Clichés, Tropes, and Phrasing: The LLM’s training data pulls from the same set of tropes and phrasing for a given idea. Give ChatGPT an outline, and it’ll convey the same information in a different way no matter what role or persona it takes on
2. Substance Across Different Voices: Although the voices writing the content differ wildly, the same core message is consistent across all the introductions.

With this in mind, I make an alteration to my ChatGPT prompt sequence for writing copy, and I add a section in my ebook on ChatGPT for content marketing using different voices and qualifiers when making content.

Opportunities for Further Research

One suggestion Bradbury gave me is to test different qualifiers and how they impact AI content.

Professional. Funny. Engaging. Formal. Friendly. See how telling ChatGPT to write with different styles works out.

Good idea. Sounds like a fun idea for the next experiment.

What is your favorite role for ChatGPT to take on? Tell me in the comments!