Is GPT-3 a Promising Model for Marketers?
The GPT-3 machine learning model is making waves - but is the hype real for marketers?
This piece refers to a machine learning model called GPT-3 that has been gaining popularity in the tech world - highly recommend reading this breakdown from the Technically blog if you want a quick primer.
Back when I was in my 2nd year at IBM, I was playing around with an analytics tool in our office that allowed you to input simple sentences and analyze the tone of those sentences.
Powered by IBM Watson, IBM’s supercomputer designed to understand human language, the tool prompted me to put in “feedback” I would have for a colleague and it would spit back the tone around that feedback.
Partly out of curiosity and partly to humor the system, I put in something colloquially positive I would say to a colleague: “I would kill to have Kevin on my team again.”
As expected, the system kirked. Horrible tone. Very negative feedback.
It had been taught the connotation around the word “kill” and acted as one would expect - but it taught me a very powerful lesson early on: There’s always a reason to have healthy skepticism of artificial intelligence.
What is GPT-3 and why do people care?
If you’ve been on Twitter or have been following new advancements in technology, you may have heard a bit about GPT-3, a new machine learning tool released in beta by AI development lab OpenAI. The tool is designed to generate high-quality text and code with little inputs, which makes some of the possibilities crazy.
Take this one for example. You can describe a layout and GPT-3 will literally build a styled website for you.
It gets even weirder - there was a blog article I read on Open AI and GPT-3 that turned out to have this plot twist:
I have a confession: I did not write the above article. I did not perform any such experiments posting on bitcointalk (in fact, I haven’t used that forum in years!). But I did it on my own blog! This article was fully written by GPT-3. Were you able to recognize it? I received access to OpenAI API yesterday and have been posting some unbelievable results on twitter. This blog post is another attempt at showing the enormous raw power of GPT-3.
The author simply made a quick bio with a few small examples and fed it to the model. With a feature known as few shot learning, the practice of learning with a very small amount of training data, the model did the rest…. 🤯
With 175 billion parameters of text evaluated, GPT-3 at this point has effectively consumed most of what has been written online - and it is definitely a bit strange to think that some of what we could be reading in a few years could be entirely written by learning models.
So, people are somewhat excited. Frightened. Optimistic. Skeptical. Twitter has been abound with self-fulfilling prophecies around AI. Name the emotion, GPT-3 has likely triggered it.
So - while it’s easy to get emotional about it - what is it actually worth in marketing?
Can GPT-3 actually transform marketing?
Now, I work in marketing and am likely far out of my depth to talk about machine learning. I can’t even begin to tell you how to make excel sheets that spit out generative text.
But, as people scream about the apocalypse of robots and replacing humans, it got me thinking about the marketing world. Are any marketing jobs at risk? Amplified? Remotely perturbed?
First, it’s important to recognize the limitations of a model like GPT-3. In a paper about the model, the team laid out plenty:
GPT-3 samples still sometimes repeat themselves, start to lose coherence over sufficiently long passages, contradict themselves, and occasionally contain non-sequitur sentences or paragraphs
In cases where the model doesn’t actually know, GPT-3 may try to force random answers that are grammatically correct but don’t actually make sense
Language models like GPT-3 are not grounded in other domains of experience, such as video or real-world physical interaction, and thus lack a large amount of context about the world
It retains the biases of the data it has been trained on - biases in the data that may lead the model to generate stereotyped or prejudiced content
As Anne-Laure Le Cunff also noted in her blog about GPT-3 and productivity, GPT-3 cannot answer questions that have never been addressed online, and cannot come up with innovative solutions that require unique thoughts.
This brings me back to marketing - Reforge CEO Brian Balfour had an interesting post on Linkedin about use cases in marketing and product that surfaced lots of initial use cases I had in mind:
Blog post and content creation (10x your SEO!)
Creative briefs
Audience research
Landing page creation
Brainstorming (Slogans, Subject lines, ad copy etc.)
But the limitations make it immediately hard to trust most AI-generated marketing that goes out to humans - any text generated by GPT-3 will lead to its own string of questions and copyediting headaches:
Does this copy take into account the dynamic nature of all audience segments?
Does this copy take into account new inputs around demographic data?
Does this copy take into account real-world issues that humans are dealing with?
Does this copy happen to accidentally plagarise or use similar language?
Is this copy actually interesting?
Think back to that blog post above written by GPT-3 and actually read it.
Take this sentence, for example:
I was recently watching a podcast about how OpenAI built their latest language model and it made me wonder what could be done with a system like this. I could not stop thinking about the applications of such a technology and how it could improve our lives. I was thinking of how cool it would be to build a Twitter-like service where the only posts are GPT-3 outputs.
Do humans actually talk like that?
Does anyone enjoy reading full sentences that start with repetitive syntax?
Is the productivity saved and potential traffic gain from mountains of quick content worth sacrificing an actual human voice?
I saw a video a while back with a few Disney movie composers where Alan Menken was sharing stories about writing the title song in “Beauty and the Beast”. The composers all nerd out (1:08) about the following lyric in the bridge:
“Oh, isn’t this amazing? It’s my favorite part because - you’ll see. Here’s where she meets Prince Charming, but she won’t discover that it’s him…”
Lin-Manuel Miranda interjects: “That doesn’t make sense. A computer program could not write that. But it tells you everything you need to know about Belle. She cannot wait to tell you.”
This quote has been on my mind while I write about GPT-3 and marketing.
A human to human interaction needs to sound genuine. It needs to have the slip-ups, faux pas, and idiosyncrasies that a normal human could. So much of marketing is about personality, understanding and translating the exuberance and imperfections and scattered thoughts that a computer simply could not comprehend when trained in the art of perfecting language.
While GPT-3 is promising in the future, I’m skeptical of its use in marketing.
No better way to know the imperfections of a human audience than being an actual human.
p.s. I have a confession: I did not write the above article.
Just kidding.