Copywriting in the age of generative AI

It’s probably fair to say that there isn’t a creative industry, medium or discipline that hasn’t been profoundly affected by the sudden-seeming arrival of generative AI. But it is the humble copywriter, in particular, who has perhaps felt this seachange more keenly than most.

Unsurprisingly, the use of generative AI in copywriting has gained traction within the digital landscape practically overnight — a momentous shift that has presented brands with fresh creative avenues to explore, as well as new obstacles to overcome. At Bernadette, we’ve always been invested in taking an strategic approach to written content that services brand identity in a way that’s relatable, inspiring and, above all, human. That being said, the creative potential offered by this new generation of machine tools has been impossible for us to ignore. 

As a copywriter (and as a human), I wanted to get an early understanding as to where the possibilities and limitations of the technology currently lie, and to find out just how close we might be to handing over technical copywriting duties to AI on a wholesale basis.

As most people are discovering, Generative AI is good at syntax and sentence construction. It’s able to assess grammatical execution better than most humans in a fraction of a second. And, it’s able to reliably produce consistent, functional copy off the bat which is mostly serviceable (albeit rather formulaic).

Coupled with that, Large Language Models including ChatGPT and Google’s Bard — on the face of it at least — appear to ‘know’ a great deal. Having been trained on increasingly massive data sets, they’re armed with an encyclopaedic memory and are able to transpose this information into coherent, convincingly phrased content in real-time.

But how well informed is ChatGPT really? And does AI really have such a firm grasp on something as nuanced as human language?

Ok. While it’s not exactly Shakespeare, a cursory glance at these two verses might suggest that ChatGPT has neatly answered this prompt down to the letter. That is, until we check for a definitive component of the classic sonnet form. I’m talking of course about — say it with me — iambic pentameter. 

Iambic pentameter is a poetic device that uses alternating stressed and unstressed syllables to give the lines a standard signature rhythm — something that ChatGPT neglected to include in this instance. It’s worth mentioning that I picked the AI up on this. ChatGPT over-apologised for the ‘oversight’, revealed that it did in fact understand the rules of iambic pentameter, and promptly resupplied the poem with corrections.

And, when pressed to prove it understood what it was replying with, it even provided a breakdown of the stressed and unstressed syllables in the first quatrain:

In REALMS where WORDS in HAR-mo-NY doth DWELL a CRAFT that WIELDS the POW-er TO ig-NITE Where LET-ters DANCE, their RHYTH-ms CAST a SPELL a SON-net OF the SOUL, the COPY-wri-ter's LIGHT

At this point I was feeling pretty sorry for it, so I didn’t bring up that last line (which is still a bit off if you ask me). But, the point this exercise helps illustrate is that it’s very easy to assume AI is always capable of writing copy that’s somehow grammatically infallible — only for a trained eye to spot inherent weaknesses in its execution.

Nevertheless, these simple interactions also show that there’s a way to push the creative limits of AI language models. With the help of a professionally trained writer at the helm — My name’s Alex, and I’ll be your Captain for today — we’re able to tweak prompts to yield increasingly better results in content generation. And, when it comes to marketing communication, we can begin to steer AI closer to a more recognisable brand tone of voice. This is a practice that’s become all the more essential when we observe how the use of ChatGPT in digital copywriting can impact more practical business scenarios.

For example, when it comes to more technical content like, say, a Tweet, casual assumptions about how AI handles even the simplest prompt can be shown to come with some degree of risk. And, for unwitting brands who aren’t quite on top of their copy optimisation practices, this could mean the difference between getting a product the exposure it needs to survive in the marketplace, or not. 

Look at this prompt:

So firstly, I have to admit, I could really chug on a ‘Volt right now. But, effective sales pitch aside, in this instance what we’re really looking for is a short piece of content optimised in format to return a measurable reach on Twitter.

ChatGPT’s solution was to create a Tweet that was as close to Twitter’s standard 280 character limit as possible. But, a trained content strategist will be aware that it’s way too long to qualify as an optimised Tweet. Research has shown that Tweets shorter than 100 characters in length get a 17% higher rate of engagement than longer Tweets, and that retweet engagement is higher when character length is between 71 and 100. In addition, the optimum number of hashtags for Tweets is generally considered to be one or two, rather than three.

So, in this case, the intervention of a Content Manager — followed by some heavy manual editing OR more informed ‘few shot prompting’ — to correct the character length and hashtag count (not to mention the US spelling and the explicit mention of ‘Gen Z’) would have been essential for our Tweet to have had the desired impact. So, is the same true of other comms mediums? 

Let’s find out! ...With something like a marketing email.

I asked ChatGTP to produce some email copy based on the following Lord of the Rings themed prompt.

Prompt: 

Give me some reasons why this copy might not be optimised for open rate and conversion. Response:

While the previous copy captures the essence of Mordor's dark allure, it may not be optimized for open rate and conversion due to a few reasons:

— Lack of Personalization: The email does not address the recipient individually, which can make it feel impersonal and decrease the chances of engagement. 

— Length and Intensity: The paragraph is quite long and densely packed with dramatic language, which might overwhelm recipients and deter them from reading the entire email.

— Negative Tone: The copy relies heavily on sinister and foreboding language, which might not resonate with a wider audience. It could be perceived as overly intense or off-putting, leading to a lower open rate and conversion.

— Unclear Value Proposition: While it evokes the atmosphere of Mordor, the email does not clearly communicate the specific benefits or value that recipients can gain from engaging with the content or taking the desired action.

It’s not wrong — at least on the first two points. To add to that, the pacing is off, the messaging rambles somewhat and I’d have expected it to land a few more knowing references to the Lord of the Rings franchise.

This is part of what we're calling iTest - an ongoing content series by real people in Bernadette, documenting experiments and discoveries in the world of AI - focusing on how humans and machines can collaborate, co-create and co-exist in harmony.

Bernadette is proud cohorts with faith - the ai creative agency from VCCP. We have faith that AI, used responsibly, will be an unparalleled accelerator of human creativity.