A company sent me a blog post last month and asked me for a couple of quotes they could use. The post was about conversion copywriting, so it made sense they tapped me— but as I read through the post, a little sliver of suspicion grew in my mind.
It sounded generic, like the stock photo version of a blog post. No storytelling. No anecdotes. Nothing unique or interesting. Very… mid.
So I popped it into an AI checker— turns out, it was only 2% written by a human. Needless to say, I declined to quote on the post.
Marketers are getting into AI tech, so today we’re going to dive into this spicy topic in our new series:
Faking Human: A Marketer’s Guide to AI
Today we’ll start at the very beginning— because in order to decide how you can leverage AI as a tool for your work and your business, we should dive into how the tools work.
Grab a downloadable version of the guide here.
Let's get into it!
How does ChatGPT work?
These models don’t understand language, so they take language and turn it into numbers so it can look for statistical patterns between these numbers. These numbers are called tokens.
Essentially, it’s word math.
You can try it out using Open AI’s tolkenizer tool to see how an AI reads language as numbers:
AI language models munch on huge data sets from the internet in order to make mathematical predictions about language. Basically, it’s using math to predict what the next most likely word in a sentence is. It’s not a writing machine, it’s a probability machine. So using tokens, the AI knows the statistical probability of which of these sentences has the words in the right order:
1. The woman coffee sat outside with
2. Women coffee with outside sat the
3. The woman sat outside with coffee
4. With coffee the woman sat outside
It’s vacuuming up the internet and predicting what should go next in a sentence— which doesn’t mean it’s correct, it’s just statistically likely to be correct.
It can also use the words surrounding a word (or rather, the tokens it uses to represent those words) to parse its probable meaning; that’s how it understands that the word content contains different meanings based on the context it’s surrounded with. For example:
- Content marketing is rapidly changing thanks to AI.
- I’m feeling content with my knowledge of ChatGPT.
And the developers of these models know that this is not a, well, intelligent form of language intelligence— so they supplement it with Reinforcement Learning from Human Feedback so it can learn how to match human expectations.
The first step is Supervised Fine Tuning (or SFT).
In this step, humans write their expected responses to a series of prompts to create a quality dataset— which takes ages and is super expensive because humans are doing the work. Paying real people for their time! What a concept!
So to supplement this data they created the Reward Model— which basically has humans rank the model’s answers by how well they matched their expectations. This data helps train the model on what humans expect it to tell them.
The Black Box Problem
Imagine you’re showing flashcards to a toddler. You know what information you’re providing but not a lot about what is going on inside that little person’s head to process it, and draw insights and conclusions from it. That’s the AI Black Box problem. Researchers can control the inputs and see the outputs, but there is a whole messy middle part that they just can’t seem to explain yet.
New York Times journalist Kevin Rose chatted with Bing’s chatbot Sydney and it told him it loved him and asked him to leave his wife.
Bing’s researchers couldn’t explain that behaviour. Cool. Cool. Cool. Cool. Cool.
Marketing is changing
Does this mean I should fire my copywriter and have ChatGPT write my copy?
No.
Relying on copy generated by AI for your business without heavy editing and fact-checking is a bad idea.
But, it can write things! Writing content is so time-consuming! Why can’t I just use it?
You can get smart about the prompts, and using a generative AI tool running on GPT-4 might write you something that at first glance sounds like a good piece of content, but here’s where a lot of human editing needs to come in:
- AI lies, a lot
In AI research, this is called hallucinating. AI just makes up facts, and confidently states them— like the guy you got stuck talking to at the last work party. Tiffany DaSilva, our SEO expert Instructor got a slew of DMs asking her for a link to an article she wrote. She’d never heard of it, and she certainly didn’t write it. Turns out, Chat GPT just made up an article title and credited her as the author. AHREFS shows hundreds of mentions of her, for things she simply didn’t write.
Remember, AI is training on data humans are feeding it— and a lot of that data is from the internet— and you’ve been on here, right? If you’re not already an expert in the topic you’re asking it to write on, so you can properly fact-check it yourself, don’t let AI make you look like a silly goose at work, or worse, add more disinformation to the internet.
Useful generative AI will need to be able to cite its sources (and cite, uh, real sources) before it can be trusted to write content that isn’t heavily edited by a human.
- It doesn’t know what happened last week
ChatGPT is trained on data from the past, specifically, it doesn’t know anything that happened after 2021. ChatGPT is still in lockdown, so it can’t write anything useful or valuable about the present, and it absolutely can’t make predictions about the future.
However, we are already seeing AI that ingests data from the current internet like ChatSonic, so this might not be the case forever.
- You’ll make boring shit
As our friends at SparkToro so eloquently put it, AI-generated content is the new floor. Everyone has access to it, and even with your fancy prompting, your content is going to mirror the rest of the internet. And it’s going to get ignored. If everyone can generate infinite volumes of content, the internet is about to be flooded with a bunch of mid content generated by AI.
It’s going to be even harder to be heard over all of the noise we’re about to experience.
Smart marketers who understand their customers deeply, use AI when it speeds up or enhances their work, and make things that actually help people are going to win.
So no, you shouldn't rely on an AI chatbot to build:
- Persuasive sales emails
- Conversion-focused copy
- Blog or guide content
But here’s the rub: a lot of people will create their content with AI, it’s fast and cheap— two things capitalism loves. Will it work well for their businesses? Probably not, but that doesn’t mean it won’t result in changes to the career landscape for copywriters, freelancers, and copy agencies.
So, how should I use it?
Even in the short term, there are some incredible ways to use it to scale your marketing operations, create really good content, and move a hell of a lot faster.
Here’s how:
- Summarizing
You can have AI digest a big technical guide and give you a summary so you can write it into a newsletter.
Prompt idea: Summarize this like an expert copywriter [https://towardsdatascience.com/black-box-and-white-box-models-towards-explainable-ai-172d45bfc512] in 250 words. Be compelling and engaging. Use short sentences and write in an active voice.
Quillbot is also a great summarizing tool, recommended by Avery Swartz of Camp Tech.
- Meta descriptions: You’ve written the guide, but now you need to write a whole bunch of copy based on the thing you just wrote. AI is great for writing quick, keyword-optimized meta descriptions.
Prompt idea: Act as an SEO expert. Read this article [add your link here] and generate a compelling, SEO-optimized 50-word meta description for the keyword [ ].
- Distribution copy:
You've written something excellent, and now you need to distribute it to multiple platforms. Get a little assistance building out the social copy from ChatGPT, just make sure you do your own editing.
Prompt idea: Act as a digital marketer and copywriting expert. Create a tweet thread of 10 tweets each under 140 characters summarizing this blog post [add your link]. Use an engaging and compelling voice.
Now go in and edit the social distribution copy back into your brand voice, and check it for errors or weirdness.
- Unblocking
You’re staring at a blank page, cursor blinking, and you can’t get started— so let ChatGPT give you some ideas.
Prompt idea: Act as an expert digital marketer. Write 10 catchy and compelling blog post titles about [keyword].
Will they be immediately usable? Not unless you want 10 other people to have the same title as you, but at least your brain is pulling it’s tires out of the mud.
What’s next?
Lazy marketers will use these tools to flood the internet with mediocre content. And that’s going to make this world noisy, uninspired, and really bloody boring.
Smart marketers will learn the technology, understand its limitations, and test these tools to find efficiencies and save themselves time. Great marketers will still use their imaginative human brains to make good things that help people in the real world.
You can and should leverage AI to help you create great things faster. Then, use the time you save to read a good book and take a nice long nap.
Made using Midjourney AI, which I’m still unsure about using for a lot of good reasons.
.png)
