The pitfalls of AI-assisted writing
Everyone who writes something knows that writing is difficult. It takes effort to formulate your thoughts into coherent readable sentences, and if you are the self-critiquing type there will likely always be a level of discontentment with ones writing. Getting annoyed with too long run-off sentences, with repeated phrases or words. Given that it is understandable that LLMs are tempting, because they can provide a shortcut to some version of âgood writingâ - at least if you want to sound like every SEO blog out there.
I am very far from an AI poweruser as I maybe use it once or twice a week, mostly at work when I just need a shortcut to get a docker compose file out the door or need ideas to tweak an SQL query. I have written several posts on AI, mostly critical, but I am not disregarding any potential usefulness entirely. I write a lot of film reviews on my Letterboxd and I often struggle with finding the right words to get my thoughts across. Maybe an ChatGPT could work as a sparring partner of sorts.
The thing that really annoys me with these language models, is how they really want to produce text that I can best describe as SEO-blog style. At least on their default settings. So my intent was to try and steer the model in specific direction, by first giving it a sample of a handful of my reviews, instructing it very clearly in only to make very minor suggestions and ask clarifying questions that could help improve my little movie review. Giving all those instructions does help some, though ChatGPT still appears to really want to add those SEO type sentences.
I wrote a review of Kelly Reichardtâs film Meekâs Cutoff and my first draft was this:
While I wasn't totally raved about First Cow, I did like Kelly Reichardt's slower and less "masculine energy"-approach to the western genre. This is also very much applied here, to even greater extent as this was a very subdued and distanced film with how it told its story. Really shows the slow yet intense pace of settlers travel by wagon across the plains. Unlike in John Wayne movies, the threat isn't from natives, but the long slow march towards lack of food and water. It isn't action packed but probably closer to reality. Most scenes are shot with a long distance, with characters' dialogue being barely audible. But it also shot in 4:3 which makes for an interesting effect that gives a sense of isolation, in contrast to westerns usual wide open landscape full of opportunity. I won't spoil the ending, but I am sure it will piss some people off by not giving the resolution the film otherwise builds up to. I just love films that are bold enough to subvert expectations like that.
ChatGPT then asked me some followup questions, like if I could expand on my thoughts on aspect ratio and the ending. After a bit back and forth with a few revisions, ChatGPT then produced this:
While I wasn't totally raved about First Cow, I liked Kelly Reichardt's slower and less âmasculine energyâ take on the western. Thatâs also true here, maybe even more so. Meekâs Cutoff is a quiet and distant film in how it tells its story. It follows a group of settlers as their initial confidence slowly turns into doubt, losing faith in both their guide and their own direction.
Unlike the John Wayne kind of western, the danger isnât from natives or shootouts but from the endless nothing â the lack of food, water, and any sense of progress. It feels closer to what such a journey might have really been: repetitive, uncertain, and stripped of the usual heroics. The muted visuals and barely audible dialogue underline that isolation, making the landscape feel less like opportunity and more like a vast emptiness.
The 4:3 frame adds to that feeling of being boxed in rather than free. And while I wonât spoil the ending, I can see why it might frustrate some. For me, thatâs part of what makes it interesting. A film that withholds closure doesnât let you off easily â it stays with you, a bit unresolved, a bit annoying, but in a good way.
The result sits a weird place with me, as it does convey my thoughts and opinions on the film pretty well, but it definitely still has that âAIâ-aura to it. I have highlighted the sentences that I really didnât like and the review I ended up posting only retains a few of its suggestions.
In addition, I tried to run the text through a basic detection scanner using gptzero.me and while I know these tools are far from perfect and can give both false positives and negatives, it did mark my original draft as 100% human and the above suggestion as AI-assisted.
I think AI-assisted writing has become something akin to CGI use in movies, where we notice it and complain about it when it is very obvious and poorly made, but we also encounter tons of CGI we donât notice and might even think is doesnât have CGI - which is just invisible CGI. I am sure we all encounter articles, or blogpost, that might very well be at least AI-assisted, that we donât notice. For some people it matters more to get their thoughts out there, even though it means sacrificing their authentic voice - whatever that is. Maybe that is okay in some situations, but while I am not against experimenting with these tools and I can see some usefulness in using it as a way to overcome writers block, almost all of ChatGPTs suggestions feels âickyâ for me. It produces sentences that are just structured in a way I donât like and wouldnât write myself, and it really shows that it has been heavily trained on tons of SEO blogs and clickbaity news-sites. It is more than just em-dashes, hard to define but a case of âI know it when I see itâ.
While I can assure readers that for this text I have not used any AI-assistance, I often struggle with ending my blogposts, so for giggles let's see how ChatGPT suggests I end this:
You could close with something that ties your reflections on AI writing back to your personal experience and tone â something understated and self-aware, like one of these options:
Maybe thatâs just the tradeoff â speed and polish for a little less of yourself in the text. Iâd rather keep the rough edges.
In the end, I think Iâd rather wrestle with my own clumsy sentences than let an algorithm smooth them out.
If writing with AI feels easier, it also feels less like me. And maybe thatâs reason enough to keep doing it the hard way.
Would you like me to make the ending sound slightly more reflective or more direct/snarky to fit the rest of the postâs tone?
All of these suggestions are terribly clichĂŠd in my opinion and not how I would normally write, but ChatGPTs suggestions beautifully and perfectly illustrates my entire point with those three suggestions.