AI in Writing

When ChatGPT first hit the public in late 2022, one of the first fields that were worried about losing their jobs to our impending AI overlords was writing. After all, the entire “point” of this new technology was to generate text natively and naturally that was capable of fooling humans into thinking it had been written by a living, thinking being. Even I was terrified for a time and outright refused to use it or support it at all out of fear for my own livelihood.

I was wrong.

While it’s true that AI chat tools like ChatGPT are game-changers for the writing and content creation fields, they alone aren’t about to take (too many) jobs. Over the last few months, I’ve begun using ChatGPT more and more for my own writing and I’ve picked up some tricks, tips, and hurdles to keep in mind while you start using AI to assist in your writing.

Recently, I was hired to write a blog post titled “Everything You Need to Know about the 2024 Bitcoin Halving”. I get A LOT of this kind of stuff. While I don’t really buy into the whole crypto world, it keeps the lights on so I sat down to force out the words.

This is when I remembered that I also had to write this blog post about using AI in writing. Well, what better way to kill two birds with one stone than by writing about my writing? So, the first thing I did was head over to ChatGPT and see how lucky I could get. I just straight-up asked it to write the article for me. It’s worth noting that for this article, I was using the free public GPT 3.5 version.

“Write an 800-word article titled “Everything You Need to Know about the 2024 Bitcoin Halving”.”

Almost immediately, the AI began explaining, in very clear terms, exactly what the Bitcoin halving is and what might happen as a result. Everything that ChatGPT wrote, in this case, was pretty much correct and the article flowed neatly.

However, this simply wouldn’t cut it as is. For one, I had been given some fairly detailed instructions to follow for the article, and this just didn’t hit those metrics. Additionally, I asked for 800 words and I only got about 650. This is around the upper limit of what ChatGPT can reliably put out.

OK, those are problems I can solve. I figured I would again try and solve both of these problems at once and headed over to Bing to try out the new GPT 4 that had recently been opened to the public. After having to install Microsoft’s Edge browser (ugh) I was face-to-face with the latest in AI chatbot technology.

I simply copy/pasted the relevant portion of the brief I had been given and asked it to write the article.

The first thing I noticed was that the response it gave me was not any attempt to create an article. It simply cited the most relevant search results in a paragraph to answer my question. This is a great palace to start for research, but I already know this stuff. I just don’t want to write about it.

Let’s try being a little more specific.

“Give me an outline for a 1,000-word article on the topic”

Alright, this looks promising. Let’s go further.

“Write the article for me”

Well, that’s less than helpful. Maybe GPT 4 is smarter, but it sure does seem harder to work with. At least in this interface. That’s fine, I’ll go back to the tried-and-true ChatGPT.

At this point, it had become pretty clear that I was going to have to do at least some work if I wanted this done. Well, OK, I can do that. I spent half an hour gathering links and writing up an outline and putting together a fairly detailed prompt for ChatGPT.

Now that’s a detailed outline. I started off by telling it exactly what were doing. This is an article, not a script or a pamphlet or whatever. That helps set the tone and flow that it will use. Next up, I fed it a few links to some similar articles that I had written in Google Docs. In the past, the Ai had been able to access and summarize open documents that I had given it, however, that seems to no longer be the case.

While I can’t find anything official, it does seem like the ability to access Google Docs may have been restricted by Google. Whether this is for privacy or to push features of Google’s own upcoming Bard AI, I can only speculate.

However, ChatGPT can and will write in given styles. If you have some published stuff, it’s easy to feed it those links instead.

Next up is the real meat of the prompt. The detailed outline that I had written with links to sources for the AI to use. This is something that ChatGPT does pretty well (for the most part, I’ll talk more about this in a bit) and is a great “best practice” for getting what you want. Finally, I asked it to start with the intro.

This last bit is important. As I mentioned earlier, ChatGPT has an upper limit of around 650 words so by asking it to start with just a 100-word intro, I can get it started clearly and concisely.

Finally! While this isn’t great writing, it’s clear, accurate, and to the point. Also worth noting is that this doesn’t really sound like me. And I know for a fact that most of my clients will put anything they get these days through an AI detector. So, from here it was a quick rewrite to get something that I could actually use.

Next up, it was as simple as asking it to continue.

“Continue the article from "What is the bitcoin halving" in 150-200 words”

Again, the exact same situation. This was going well. I continued to do this for most of the article right up until the end. In the “expert Opinions” section, I tried something a little different.

“Continue the article. Use specific quotes from the Forbes article linked above. 200 words.”

At first, this looks like exactly what I asked it to do but there’s a major problem here. I read that article, I don’t remember those names or quotes appearing in it. After going back to confirm I made a few interesting discoveries.

First, Oluwaseun Ayandeyanju is not the CEO of Block Stale. He’s the author of the Forbes article. Looking up Ajeet Khurana, he does seem to be the CEO of Zebpay, but, as far as I can tell, he’s never said those words. This was an issue.

Thinking there had been some miscommunication, I slightly refined my prompt.

“Rewrite using links from this article: https://www.forbes.com/sites/oluwaseunadeyanju/2020/05/11/bitcoin-halving-unlikely-to-drive-btc-price-higher-in-the-short-term/”

The same thing. At this point, I wasn’t about to try and track down the source of these hallucinations. I knew it was wrong and it was easy enough to find some quotes from the article that actually did exist.

Finishing up the article was easy enough and in the end, I was able to turn in something that was officially “100% human-written” and met all of my client’s needs.

In the end, this project has just reinforced everything I’ve seen in any AI project. This is a powerful weapon, but it’s not a smart bomb. AI isn’t coming for writers. Writers who use AI are coming for those who don’t.

A colleague of mine was dropped by a client a few months ago in favor of ChatGPT. Within 2 months, they were back asking for help because the chatbot just wasn’t able to consistently deliver what was needed. Writing has always been collaborative, now we just have another voice in the room.