Note: I am taking a much needed day off to get some outside things done, which means starting before the cock crows. No time to write this morning, so this is a green door post, gifted to you on Good Friday. Happy Easter everyone.
For the last two weeks, I have been allowing AI to edit my posts. It is not a single AI but a team of AI. The editors are ChatGPT, OpenAI, Grok, and GabAI. Every post is fed into the maw for editing grammar and spelling. Sometimes I will feed a post into one, then take its output and feed it into another, and so on. I have also continued to use Word, which seems to be going insane.
One of the first things that jumps out is the AI tools will not simply fix grammar and spelling, no matter how you instruct them. Instead, it is a full rewrite that attempts to make the text like what was used to teach them. My guess is these tools attempt to consume everything on the internet as the baseline, but the starting place was probably many publicly available texts in every language.
Curiously, all the AI tools have an obsession with hyphens. If they are given a sentence like, “He jumped the fence, as he was a track athlete,” AI will try to rewrite the sentence with a hyphen between the two clauses. The results are often ridiculous, so I started adding a rule to avoid all hyphens. That means they change “twenty-two” to “twenty two,” but I can fix that more easily than the other option.
In truth, I could start with a longer list of rules and get a result like what you would get from your teacher in primary school. The output would be the original text with suggestions and corrections noted in the text. That requires far more work than simply handing it to a human and saying, “Proofread this for me.” The point here is to make the test apples to apples or as close as possible.
Another bit of weirdness is AI loves contractions. Every occurrence of “it is” will be changed to “it’s” unless you demand otherwise. This is a curious thing, as contractions are generally frowned upon. The style guides I have all say to avoid contractions unless they are in quoted text. Word will flag all contractions. For some reason, the AI editors have gone the opposite direction.
Here is where the basis for the AI knowledge bases comes into play. It starts with formal text and then continues to learn using what is online and fed to it. Casual writing will be littered with contractions, and since that is the bulk of what is online, the robots assume contractions are clearer and more concise. If everyone is jumping off the bridge, the AI editors will jump off the bridge too.
Probably the most amusing bit is none of the AI editors agree. I will feed a post into one and then feed the output into another, and so on. Every output is different from the others, and when you get back to the starting AI and input the last output, it spits out a different version from the first go. Like real editors, there is a desire with AI to change the text when no changes are warranted.
Another amusing bit is that the output from AI pasted into Word causes the Word spell and grammar check to have a stroke. Word has become almost unusable at this point, but it does a few things well. For example, it will change “have to” to “must,” which is better in terms of efficiency. Otherwise, Word often hates what comes out of the AI editors like it is an angry old schoolmarm.
For basic spelling and grammar, it is a useful tool as long as you do not mind it rewriting the text or you are willing to supply many limiting instructions. Is it better than having old school Word flag spelling and punctuation? It depends. If you are like me and have confidence in your style, it is not better. If you do not have confidence in your writing, then it provides a sense of security, which is not the worst result.
That is the point of sites like Grammarly. They are for people who probably should not have been taught to read but who have it in their head that they need to tell the world their opinions. These users can paste their text into the site, and the result is obviously better, and it comes with the approval of an authority. In a permission society like ours, most people need that pat on their head.
The banality of the output is something else I have tested. Instead of writing my post and then submitting it for editing, I have had the AI team write the post based on the points I supply and then asking it to use my site as a guide. This takes far more work than you would expect for some reason. I found I needed to think about what I was planning to write far more than I do when I do the writing myself.
Maybe it is just me, but when I write an essay, I am not entirely sure what I will be writing when I start off. I get going, and after a few minutes, I have a few paragraphs and a few ideas for what to do with it. This happens in a few cycles until I have about what I want for a daily post. Then I think about how to put a bow on it. This approach cannot work when using AI to write a post.
What I did instead is write a post and then use it as the basis of the prompts for AI to write an original post on the topic. With some tinkering, I can get a result that is pretty close to what I would write, but it takes much longer than writing it myself. Otherwise, the resulting text reads like a technical manual. There is a flatness to the writing that fails to engage the reader. Reading the result feels like work.
What this suggests is that the ceiling for AI may very well be the absolute middle of human creativity when it comes to communication. It will quickly write text that mimics a mediocrity at National Review. It can quickly produce audio that lacks the sort of variability that makes hearing one another enjoyable. For many things and most people, this is more than enough to do the job.
If you like my work and wish to donate, you can buy me a beer. You can sign up for a SubscribeStar or a Substack subscription and get some extra content. You can donate via PayPal. My crypto addresses are here for those who prefer that option. You can send gold bars through the postal service to: Z Media LLC P.O. Box 1047 Berkeley Springs, WV 25411-3047. Thank you for your support!