Copyediting with ChatGPT

Pick up Grey Hat, the novel that I developed this methodology while working on at Amazon today.

Guide

When I first started, it would take me about an hour and a half to complete a chapter of between 2,500-3,200 words. However, through trial and error, I can now complete it in about 20-30 minutes. I am using the totally free version, and you will need to sign up for your own account.

A couple of words of caution: any bold, italics, or other textual elements will be lost. Also, like all Large Language Model (LLM) AI systems, ChatGPT has a tendency to do what is known as “hallucinate,” which is a fancy way of saying “make stuff up.” In my experience using it for the copy-editing process, the hallucination manifests mostly in a couple of specific ways. From the most common to the least, these are:

1) It will change characters around, especially if you have a lot of dialogue.

2) It will rerun the previous entry.

3) It will reduce an entire page of narrative, including dialogue, to a paragraph or two of straight exposition.

4) It will completely rewrite the entire submission so that it is unrecognizable. (I placed a side by side example below)

In the case of the first one, it is just a process of fixing it by hand. As for the other three, you will want to regenerate the response using the button above the prompt (don't worry; there is a button at the bottom). My personal record of regenerating it is over nine times, and more than once, I have had to either redo the prompt or, in some cases, delete the existing session and start a fresh new one.

Now onto the workflow. I copy and paste a chapter at a time from my manuscript into two separate Word documents. (This will also work with Google Docs or any other word processing solution that has the ability to compare documents.) One I label ORG, and the other I label CHAT.

I will then type in a variation of “Copy edit this,” “Copy edit the following,” or “Copy edit this text” into the ChatGPT prompt box placing chunks of about 500-600 words at a time after the prompt from the CHAT copy. I scan the output to make sure it hasn't obviously hallucinated and then copy the text back into the CHAT copy.

If the output is double-spaced or has spaces between the paragraphs on your screen, you will want to highlight the text in the browser to copy the output using the “paste and match formatting” function in Word. If it doesn't look double-spaced, there is a little clipboard icon to the top left of the response. Click this to copy the text to your clipboard, and then just paste it as unformatted text. This avoids a lot need for fixing the formatting, which is important for the next step.

With a now copy-edited version in the CHAT document, you want to make sure the formatting is the same in both. With quotation marks (which can be the biggest pain), it is best to turn on “smart quotation marks” in the settings function. But you need to do one more thing to make sure all the quotation marks in both the CHAT and ORG document match in their style. To do this, go to find replace function and put the quotation mark into both fields and replace all. Repeat this with the apostrophe. This will force Word to change all the quotation marks and apostrophes to the same style. If you don't do this, the next step gets messy. Now save the CHAT document.

Once you have completed the previous steps, use the “compare documents” function (normally in review) to compare the ORG and CHAT documents. This will generate a third document with all of the ChatGPT edits highlighted in track changes. From here, you can either go through and accept or reject changes in this document and then copy and paste the whole thing over into your manuscript, or pull up your manuscript and go over each suggestion manually. I recommend a combination of both, depending on the extent of the suggestions, and do not suggest blindly accepting the changes due to the hallucination tendency and the risk of altering the voice of your manuscript. It could remove character traits in dialogue that you intended to keep. Repeat this process for each chapter of your manuscript. A graphical representation of this workflow is right below.

A few final notes:

Change up the prompt of “copy edit this” or “copy edit the following” with each run, as repeating the same prompt may cause the system to rerun the previous entry more often. Also, keep in mind that using more than 500-600 words (slightly over one and a quarter page of Times New Roman 12-point font) may cause the system to cut off text from the bottom or you will need to tell it to continue and when you copy and paste it back you might end up with a random break. Additionally, late at night or early in the morning tends to have the quickest response times in processing.

Finally, keep in mind that there are terms of services limitations on the types of text you can input, especially with genres like horror or romance. I had one scene involving a kidnapping victim flagged, despite it not being very graphic, so the system is sensitive to certain themes.

Example of a ChatGPT Hallucination

Although ChatGPT generated an exciting scene, it would have actually broken multiple subsequent plot points within the context of the actual story. Additionally, the character Tony is reacting to an event that he had no physical or technological ability to even know about. Most of the changes suggested by ChatGPT were mundane, this particular trope filled hallucination made me laugh.

Commentary on ChatGPT Writing and Editing

Now, a quick aside: There is a lot of excitement and trepidation about ChatGPT and other generative AI systems. Personally, I believe that one of the easiest ways to protect creatives is to deny copyright protection to any original work generated by a generative AI. At that point, its value as anything other than an aid for creative endeavors is basically over since no publisher or studio will waste resources on them. Even high-volume, low-content self-production individuals will find limited potential returns.

Because of the way AI works, it needs feed material, limiting it to generally derivative generation. As this develops, well-established authors, those who populate the majority of Target and Walmart book racks, are likely to face a rash of copycats and knockoffs of their style and even characters. This will be especially true for online and international markets. It is also possible that "original content" for highly formulaic materials, especially if there is a massive body of source material for the AI to draw on, could, in whole or in part, be generated by these AI systems.

That being said, if we view creative activities such as writing on a spectrum - ranging from less creative and more informational to highly creative production - then materials generated, such as press releases (not that they can't be creative, but they are mostly about information), are more likely to be displaced by LLMs. This doesn't mean that there will be no more jobs for that type of writing, but rather that one person can suddenly produce a lot more, so the number of necessary positions will go down. You can almost think of this in the context of self-service checkout kiosks at stores. One clerk can suddenly cover 4-8 checkout stands. If your job involves generating low-creative materials, then your work is either going to get a lot easier or disappear.

Additionally, based on the output results from using ChatGPT versus copy-editing done by a person, I am amazed at both the speed and quality of the process. Given that the pro version of ChatGPT costs $20 for a full month (which you don't currently need to pay for to copyedit your book), and assuming a rate of 3,000 to 3,500 words per day, you can easily process a 100,000-word book in a month. So, professional copy-editing may face serious disruption.

Again, I don't think those jobs will entirely disappear. I suspect that developmental, copy, and line editing will face consolidation. Those who adopt/adapt may see a massive boost in productivity and market share. Given the workload agents face, they would be crazy not to adopt this model into their workflow of preparing manuscripts for submission to publishers. The same is true for companies offering services and tools to authors. Whoever gets there first will have a first-mover advantage in the market. As for publishing houses, I believe hybrid and indie houses will find real benefits in terms of the quantity and quality of their products. For the big houses, I would bet on further staff consolidation in their editorial departments.