4 years as a content writer working with AI

My profession is one of the main targets of the AI job doom. I've been working as a professional content writer for 4 years now, even longer if you consider the fact that all I ever did in college was write essays and thesis papers.

When Generative AI first arrived in 2020, everyone was talking about how it could make "good content" in orders of magnitude less time and money than what you'd have spent hiring a writer like me.

And yet, I'm still here. Like a cockroach that won't die, no matter what kind of bug spray humans come up with.

I don't attribute my survival in this complicated job market solely to my merits. I believe I'm a good writer, but being "good" isn't enough these days. I had the advantage of getting an early start on creating AI content. When OpenAI first introduced its breakthrough model GPT 3.5, I was lucky enough to be working in an SEO agency with a founder who prioritized learning about AI, and how it could be used in the marketing space.

Ever since then, I've had continuous exposure to how AI is used in content marketing. I've had a lot of clients, ranging from people who are extremely against AI to others who would give me access to their model and tell me to just generate everything with AI and focus on polishing it. It's this exposure to the development of AI in the content marketing space that led me to the opinions I'm sharing with you today.

I have a lot of reasons why I think AI won't (and can't) replace human workers. I've condensed them into 3 points, all based on my experience as a writer.

1. Good Prompting Is NOT Good Enough

I've heard one of my past clients say these exact words: "All you need is good prompting." Which I, in my humblest, most biased opinion, nope, it doesn't.

Unless you're the type of person to settle for just "good enough", then sure, go ahead. But your customers aren't stupid enough to fall for "good enough" content.

I also find it amusing how most of the people who say this are not writers themselves. There is more to making content than just explaining stuff to the reader, you know. If good prompting is all it takes, then why do marketing companies still hire writers? Simple, because even they know that good prompting is not good enough. The people yearn for more authenticity, something only human writers can give.

We are experiencing what is called AI fatigue. People are tired of seeing AI-generated content. Your readers can tell when you're just stating what everyone else has already said, and they stop reading.

The only instances I see where AI can make high-quality content are when it is used by competent writers or experts who have enough time to properly review and refine the content. They cover where the AI fails.

I remember working as a ghostwriter for a client who had this mindset that AI can do everything. He was an SEO specialist, and he needed some content done for a fashion company he'd contracted with. He told me to use AI and deliver the content immediately. I followed what he said and did everything I could to improve the content within the time I had. I submitted it to my client, and he approved it.

He came back to me a few days later with feedback from the fashion company. They said the content "lost" their brand voice. It was too AI-sounding. They weren't satisfied. I asked my client if we should rework it.

I genuinely felt bad. But my client insisted it was okay. We didn't have to make any adjustments or revisions. "They got what they paid for," he said. I'm not really sure where to place the blame there. Is it me, the writer who technically made the content? Some of my colleagues said it was understandable with the timeline given to me, but I still felt bad. Was it the fashion company? Based on what my client said, it seems they paid a low rate. Still, I have only his words. Is it my client who insisted that the timeline was feasible because of AI? He was pretty open about his amazement at AI, and there were times when I felt he trivialized my work as a content writer, but if he was working with a lower-than-usual budget, then I understand.

This is also where I realized what my main issue with the "good prompting is all it takes" mindset is. It minimizes the skill, time, and attention to detail needed to write great content that resonates with readers. There's also this insistence that if the quality of the content is bad or lacking in some way, all we have to do is "fix the prompt".

Why do people listen to podcasts? You can make AI generate a podcast script, but it lacks soul. AI-generated videos have also failed in capturing an audience. Why?

There are certain ways you can make content engaging. One is writing like you're talking to someone, like a friend. You have your own quirks, your own ways of expressing stuff. And this is what piques human interest. Making good content requires you to address your readers in a personal way. That's how they'll know they're actually engaging with a real person with actual, lived experiences, not a corporate bot who just wants something from them.

Here's what I believe: AI could never write the same way humans would, no matter how good your prompt is. It's been 4 years since the AI boom, and yet people can still see signs of AI usage. You're probably reading my article because you were curious what a writer like me would say about AI, right? Making good content requires the human experience.

1. With Every New Model Come New Problems

When I used to work at a digital marketing agency that was big on AI, I remember one of the most common topics during our weekly meetings was this:

"What sort of errors has the AI given you this week?"

I was already working as a writer when ChatGPT 3.5 was first released. The main problem we had with it were hallucinations. Terrible hallucinations.

At this time, I wrote content for a highly regulated industry. I spent many hours reviewing laws, because I couldn't trust the AI to give me accurate drafts. Sometimes, it would even make statements that looked credible at first, but were completely misleading. I was less experienced back then, so I was prone to missing these details. I thank God for my editor.

ChatGPT 4.0 came out, and it hallucinated less. However, the quality of the answers got worse the more complicated your prompting was. The longer you pressed on with your queries, the more it seemed to "panic". Sometimes, it would generate a paragraph that looked good at first glance. But when you read it fully, you realized that it's basically just saying the same idea again and again.

A colleague of mine told me that the way to solve this is to "give the AI more material to work with." So I did. Did it produce content good enough for me? Nope. In fact, the more I fed it with quality data I found, the more it misled me. The longer my guardrails were, the higher the chance it would hallucinate again. I've tried different ways in my own prompting, and to this day, this still ends up being a common occurrence for me.

I do see the improvement of the recent ChatGPT models, but as usual, I've encountered bugs not present with the earlier models. One time, it started talking in a different language. I'd be asking it to write me a draft, and it would write parts of it in Russian or Arabic. When asked why, it would just say that it's an "error" on their end. Has anyone ever experienced that? I think it comes from training data where people switch languages mid-way when answering a question. That's just a theory, though.

With Claude, I have a whole other problem. Dear god, it tries so hard to be politically correct.

One time, I needed to make product descriptions and blog content for a toy maker. I had an outline, and asked Claude to help me write a draft for a makeup toy for girls. And what did it do? It began lecturing me about how toys are not exclusive to gender. All because one of the keywords I added was "toys for girls." THAT IS LITERALLY AN E-COMMERCE KEYWORD.

I am not making a political statement, I am optimizing for search! People write multiple keywords in e-commerce websites, such as "toys for kids", and when people want to specifically find toys for their daughters, they are more likely to type "toys for girls."

And yet what I got was a lecture from an AI model. I tried to "reason" with it, but it wouldn't budge. It did generate a draft, but every time the keyword "toys for girls" popped up, it had an accompanying "(or boys)" with it. And on top of it a huge paragraph after the draft that reminded me how bad it is of me to think that there are toys exclusive to gender. If I wanted to be lectured on political stuff, I'd just open Twitter.

I've used these major models extensively in making content, and in my experience, none of them has ever given me content with no issues.

1. AI Cannibalization

Half of the articles published online are written by AI. Coming from the content marketing side of business, that's not really surprising. Most digital marketing agencies nowadays use AI to speed up production. They pair it with writers and a few editors who can polish and refine, and you can produce twice as much content as you did before AI.

However, with this increased use comes a problem: AI cannibalization.

Ever noticed how the difference in quality between new ChatGPT model generations seems to be shrinking? The "trust me bro" benchmarks AI companies put out suggest otherwise, but in my opinion, the gains in content quality since o1 have been quite incremental.

I believe this is because the amount of authentic human-produced content in the training data has decreased. Because more and more people are making AI-generated content, this ends up being the bulk of what newer AI models "consume" for training. As a result, what we get is a model that reiterates the same ideas and biases.

Back in 2023, Reddit was my main competitor for Google rankings. Reddit posts were dominating search with every keyword I was aiming for, especially those that focused on asking questions. Why? Because it was the site with the most human-written answers and nuanced takes. As more and more blogs and business sites adopted AI, the insights they offered became much more similar to each other.

AI doesn't reward ideas that aren't already shared by many. It's a tool based on probability. So, if your main goal is to make content that "offers a unique perspective" to your readers, using AI as your only source of information is insufficient.

AI cannibalization is dangerous because it reduces the chances of finding varying perspectives. What if we rely on AI content so much that all that's left on the internet is a bunch of content that presents the same idea, just in different packaging?

It's precisely this problem that made marketing companies hire writers again. This time, the job revolves around using AI as a tool to speed up production, focusing on editing and refining it enough to make it valuable.

I'm not going to be a hypocrite here and say that I don't use AI in my workflow. ChatGPT is very good at helping me continue my sentences when I get stuck. If I have an idea, but I don't know how to express it, I give GPT a barebones description of it, and it generates paragraphs on how I could possibly say it.

Newer AI models have also been much more reliable in making drafts, though I find myself still rewriting most of them. I usually see what it generates, take a few parts, and enrich it with my own words. Using it to check for grammar is another good use case. For research, Perplexity is a great tool. If I remember reading something somewhere and don't know where I got it, I ask Perplexity to find it for me.

However, and this is a big disclaimer, my use of AI comes with careful consideration. I've learned from experience that AI, despite its strengths, is incredibly unreliable. That's why I make sure that whenever I write about a topic I don't know, I take the time to research it.

[truncated for AI cost control]