AI-Generated Content Marketing Isn’t Here Yet: Conversations With A Robot
This is not a doom and gloom article about how AI is coming to take all of our jobs. But it’s also not a flowers and sunshine post about the indomitable human spirit that makes us irreplaceable by machines. Instead, it’s a warning and a reflection on the state of content and AI.
A few weeks ago, OpenAI released a tool for interacting with their GPT-3 natural language model called CHATGPT. It’s a fun tool that takes the existing GPT-3 model and makes it more accessible than previous tools like Talk To Transformer (now InferKit.) And that access has sent the collective imagination wheeling with wild proclamations on the death of human-led content marketing.
Smart people have decided, based on playing around with a tech demo, that this is the future. It’s a shame they didn’t bother asking CHATGPT what it thought of the idea:
I’m optimistic about the growth of AI, but I’m also optimistic that it’s not going to make me or you obsolete any time soon. And because it’s always best to get information from a first-hand source, we’ll also talk to CHATGPT about why it shouldn’t be replacing your content team any time soon.
1. AI Isn’t Actually Smart
That’s right, artificial “intelligence” isn’t actually intelligent. Not in the way we think about it, anyway. Natural language models don’t understand what they’re writing — they just create text that looks like something a human might produce. But critically, it doesn’t understand what you’re asking it to do or what the text it produces means.
Models like GPT-3 don’t read your prompt, understand it the way a human would, and then give you a thoughtful answer based on their understanding of the question. Instead, the AIs build a probabilistic model based on how often certain words are used in response to certain other words, and add words to their output based on those probabilities.
In a lot of ways, this is no different than picking words out of a hat to answer questions, provided you could change which words were in the hat and how often each one appeared based on the input and adjust after each draw. The final result might sound like a complete thought, but it isn’t a thought at all.
2. AI Can’t Currently Look Outside Its Training Data
GPT-3 isn’t able to do any research, verify facts, update statistics, browse the internet, or gain any awareness of the world outside of its training data. And since GPT-3 is the absolute best available, none of the other models can, either. If you need content about something that didn’t exist within that data set or that has changed since the training was completed, you’re out of luck.
Actually, you’re worse than out of luck, because AI as it currently exists will happily make up facts about your prompt without the slightest hesitation. While it may occasionally confront you about incorrect or false premises, it’s just as likely to roll with whatever you give it. In this case, it suggests cats are increasingly being used as mobility devices, despite cats being mostly an immobility device:
We saw a similar issue when a client and friend of the company working in medical devices excitedly sent us a message with an example 510(k) submission for their product, generated by this very same AI. It read well, used roughly correct language, accurately described the condition being treated and its symptoms, but it also made claims that it had absolutely no way of checking or verifying. In this case, they weren’t completely incorrect, but they were right on the edge — a large problem in regulated fields, and especially when the actual content was a paragraph any human could have written in only a couple more minutes than it took to write the prompt. You can have some fun inventing the next great medical device just by using the prompt: “Write a 510(k) submission intro for _______________.”
As you can see from these examples, the AI natural language models will casually make things up, do it with authority, and not let writers or readers know where existing foundational knowledge has ended and creative writing has begun.
And God help you if you need citations, links, quotes, or other connections to outside data in your writing, as most of us do. It’s not going to happen. Not without significant manual effort. At that point, you may as well have put that effort in from the start.
As we learned in point one, these AI models do not have any context — they don’t know anything, and they certainly don’t know anything about you, your company, or your customers, and will fill in that very large gap with words at quasi-random. Which leads us to the next point.
3. AI Doesn’t Know You
Nor about your company, your objectives, your customers, your target markets, your competitors… really, any of it. You can get around some of these limitations by writing prompts with a lot of details and information, but at that point you may as well have written the content yourself, because by the time you get done teaching the robot about who you are, you will have spent more time than if you had launched a functional content program yourself.
This is really the piece that makes me most optimistic about my continued employment: content isn’t just spitting out words. And this should be obvious to anyone who has ever worked with a marketing agency. There have ALWAYS been cheap alternatives to a robust in-house content program or a professional agency — I get emails every day from off-shore content farms offering me content marketing services at prices that are absolutely bonkers. AI doesn’t drastically alter the balance of the marketplace or the availability of cheap, low-strategy content, and while it feels like a whole new thing, it’s really just another version of the offshore mill.
Effective content marketing is about understanding your brand: who you sell to, what pain points you address, how you do things differently and which of those differences matter to your customers, and what your customers actually want to read. You can certainly get all that through an AI text generation bot, provided you have excellent prompts that effectively contain all of that information… you know, all the actual hard stuff. Just putting words out is cheap and easy; making sure they’re the right words is the expensive part.
4. AI Shouldn’t Be Doing Your “SEO Content,” Either
I don’t have a good screenshot for this one, because this is one of those cases where the AI model just defaults to repeating the noise-masquerading-as-data that is the vast bulk of marketing advice available online:
This is bad advice. It’s not bad advice because there’s no value in considering SEO in your content — there is, and we work hard to ensure that the content we produce is findable and addresses search queries and intents. It’s bad advice because most people are still stuck on the idea of “keywords” and equate “SEO content” with “low value word vomit that uses a certain keyword a formula-driven number of times while providing zero value to any human reader, but allegedly feeds Google’s latest algorithm and helps drive users.”
We’ve been telling this to our customers for years: SEO comes second. If you begin with keywords and try to work your way into content, you’re going to have a lot of very bad, very generic content that sounds just like what everyone else is putting out, and it’s not going to be effective at converting visitors into customers, or even getting visitors in the first place.
Everyone has the same keyword research you do. Everyone you’re competing against is writing to the same keywords you are. Every piece of content written to a keyword is going to be roughly similar, provided they’re written for similar industries, companies, and customers. And when your less-talented competitors all start generating AI-written “SEO” articles by the thousands, it’s going to turn an already overflowing garbage pile into a full-on deluge of noise that will completely break what little hold SEO still has.
So when I hear otherwise smart marketing thinkers saying “well, AI can help us outsource a lot of the routine SEO content we produce,” what I hear is “we’re wasting a ton of time and budget producing bad content that doesn’t accomplish any of our goals, and rather than reevaluate what we’re doing and why, we’re just going to keep grabbing on to whatever solution is currently promising us lower costs.” Which is an absolutely unimaginably stupid way to run a marketing program.
AI shouldn’t outsource your high-level strategic content because it isn’t capable of thinking, let alone thinking strategically. But AI also shouldn’t outsource your low-level filler “SEO” content because you shouldn’t be producing low-level filler content in the first place. Every piece you or your team produces should be laser-focused on addressing customer needs first, and keywords second. Search intent and solution-focus matters more than keywords. 🌎👨🚀🔫👨🚀Always has.
So Who Should Be Worried?
Well, the person writing low-level SEO content for unsustainably low wages. And the offshore content mill operators. And any marketing manager whose boss insists they immediately start using AI-generated content. And anyone who thinks using AI-generated content is the silver bullet to their marketing woes. Basically, anyone who sees content as a checkbox to be crossed off with minimal cost and effort.
Copywriters and anyone dealing with short-form, creative (as opposed to factual) copy should, as well, because while these services are very bad at putting thought into long-form thinkpieces, they are very good at writing 100-word product descriptions and hundreds of AdWords ads to test against each other.
And brands should be worried, because AI-generated content is going to create pressure to throw caution to the wind and dive in on cheap solutions. That pressure is going to come from every direction: the C-suite, tech writers and thinkers, thought leaders in marketing, internal folks short on resources, and the general paranoia and FOMO of every person whose job is to think “what if our biggest competitor does this first?” It’s going to feel overwhelming and inevitable. And giving in to that pressure will be the deathblow to your content program.
Or as my good friend CHATGPT puts it,