To be fair, I was warned.
See? Right here, “The feedback from 5 is a bit savage”.
That was Jeromai, offering a heads-up after sharing his incredibly interesting findings on the differences in emotional intelligence between GPT-4o and the latest new-fangled GPT-5 — and their respective abilities to meaningfully contribute to creative writing.
This time around… The test subject? One of my own posts.
Jeromai then turned the scalpel on the feedback itself, comparing the outputs of the two models.
GPT-5 accused me of rambling.
Which… okay, fair. But also: rude.
The feedback was laser-focused on tightening things up — cutting sentences down, trimming the fat, eliminating repetition. And look, I get it. But I like a little ramble. I cherish a good wandering tangent. Grammarly’s been trying to rap my knuckles free of this for years, but I still hit ‘Ignore’ nine times out of ten.
GPT-5, bringing the bedside manner of a brick, described the post as an “unfiltered thought-dump.”
*blink*
GPT-5, you say that like it’s a bad thing.
But honestly, thank you for recognising my process.
I feel seen.
That is how I write. Not as some grand act of defiance — it just feels right. I’ll throw in an “actually” or a “really” where it sounds like me — even if Grammarly lights up like a Christmas tree and tuts disapprovingly in the background. Sometimes I’ll look at the suggestions. Sometimes I’ll even take them. But a lot of the time? Naaaaaah.
Because yeah, if I stripped it all down, it might read ‘better’ by some technical standard — more concise, more precise. But then you’d just be reading some machine-approved version of me, and I don’t think anyone wants that.
Well, that’s what GPT-5 had to say. Brutal. Blunt. No holds barred. But what about GPT-4o — the nice AI? The one with the “emotional intelligence”? Surely it understood me.
“…could be tightened for more impact. It does meander a bit…”
Goddamnit.
*voice echoing from down the hall*
I’m still not changing anything!
Okay, okay — slightly more seriously for a moment.
Not all the notes were bad.
GPT-4o, for instance, suggested I consider adding one-sentence caps at the end of each game section to offer a thematic takeaway. Something like:
“ToME reminds me why I fell for this genre in the first place.”
“Qud teaches me that normal is relative.”
“Elin? Elin is a fever dream I’m still decoding.”
And you know what? That’s… actually not bad advice. I’d be up for considering that.
Just don’t expect me to stop rambling while I do it.
This was posted during Blaugust 2025, the annual blogging event normally hosted by Belghast, with a twist this year of being hosted by Krikket. Blaugust is an event aiming to welcome new blogger blood into the fold and revitalise those who’ve been at it a little longer.
The Blaugust Discord is still available to join in, year-round!
My understanding is the GPT-5 is experiencing some growing pains. Instead of giving us a choice of what model to use, they’ve “simplified” it so it now picks the appropriate model and that seems to be failing an awful lot.
But that’s just what Open AI says so who knows?
Yeah, I was under the same impression as you initially — but playing around with it a bit, it seems that it only really selects between GPT-5 Quick and GPT-5 ‘Thinking’. With a premium plan you can still use 4o, if you enable legacy models in settings, though.
Off the back of Jeromai’s first post, I queried GPT5 on what it thought the best model for this type of work was, and it ended up generating a flowchart of when to switch in each model. >.<
I’m sure you’ve seen my comment on Jeromai’s post so I won’t rehash that but I’ll say here that I completely agree with your reasons for disregarding CGPTs “advice”. It has no clue what it’s saying – literally – so why would anyone pay it any notice?
Independent of tha AI issue, I was taken by your mention of “really” and “actually”. I don’t just use those a lot – along with kinda and sort of and even kinda-sorta, which I really like – I use them intentionally and purposefully. Were you crazy enough to go back through my posts and analyze them for style, you’d hit a point where rarely used supposedly empty qualifiers like those. I used to take them out if they found their way in.
But then it occured to me that I like to see and hear those kinds of expressions in the work of others. They feel inclusive and warm and foster a sense of belonging. Someone talking to you with that degree of informality is inviting you into their world and sharing it with you, not speaking outwards to you from inside it. And I already use more than enough complexity in my sentence structure and syntax without making it even more thorny and off-putting by excluding the softening words.
And the thing is… I know that. Those words aren’t there by accident. They haven’t not been considered (Neither was that double negative. Very little I ever put to print is unconsidered. ) I’m not going to fall into the old intentional fallacy trap where content is concerned but when it comes to aesthetics, a recognizeable personal prose style is crucial. Grammarians, tutors, academics and especially AIs are never in any position to tell anyone what their personal prose style ought to be.
Writing workshops talk a lot about “finding your voice” and this is what it looks like. If the AIs don’t appreciate it, screw ’em. Their loss.
I think there is a common issue across all writing assistants, and it seems true for AI as well to a lesser but still true degree, where they optimise for a more formal setting by default. Grammarly is never going to change its ways, but I still won’t entirely get rid of it because it does occasionally catch the odd ‘true’ whoopsie.
AI I wonder if we might be able to train it to better suit our intended sense of style and voice with time and effort, as Jeromai clearly has with 4o and the work he does with it. I’m also not sure that I’m (yet) that invested to try and do the same.
As much as I think (well, let’s be honest, ‘hope’) that AI has a bright future — even after Wilhelm’s recent post, and after watching some videos going over the ‘AI 2027’ paper — right now, I view it as more of a toy in most use cases. I’m giving it a shot as a programming/project assistant though, which I thought this post might also go into, but I’ll save that for another day.
Asides from the whole AI thing though, and more on ‘Finding your voice’, I do sometimes wonder whether my ‘voice’ is just a poor one, and I’m just too lazy to ‘fix’ re: tighter wording for clarity and removal of the softer (but ultimately meaningless) words.
The truth probably lies somewhere in the middle — I’d never suggest there is never room for improvement, but whether it lies in that direction or somewhere else… shrug
GPT-5 is an idiot. That’s pretty much all I gotta say on the matter. For now, anyway, until and if OpenAI chooses to refine it, or not. It can deliver things that you want in smooth flowing sentences, but you pretty much have to sit there and include every last instruction you want it to mirror in the prompt itself. That one prompt. Because it sure isn’t capable of pulling sufficient context from elsewhere to think properly.
GPT-4o did recognize the ramble as deliberate style and called it delightful. So… :P
But thank you indeed for being a great sport about it!
Feel free to do the same if desired to anything of mine. I may enjoy 4o’s writing style, but I can very much differentiate that they’re artificial intelligence LLMs and whatever they say needs to be taken with a grain of salt. A whole saltshaker, sometimes. (Might die of high blood pressure working too long with GPT-5 too, come to think of it.)
Hah, it was fun. And to be very clear — nothing the AI models said was found to be offensive (or even surprising).
I thought I might just have a little fun with it, since I saw your post come through literally as I was about to sit down and work out what my topic would be. Then I thought the post would half be on this, and half a new project kicked off with AI more around a coding project, but just ran with it long enough that made it its own post, and the rest can wait. :)
I may well take you up on the offer to run something of yours through, both as a fun cross-blogger activity, and also because I’m interested in what happens if I follow GPT-5’s own advice of starting with 4o for new writing or heavy editing with style considerations, but then to run it through GPT-5 with an instruction not to edit for style, but to review only transitions, grammar, logic, etc.
Given that Chat GPT is generally only capable of producing the most bland and predicable prose you can imagine, something about rocks and delicate transparent structures comes to mind.
It’s actually is pretty good for the “I haven’t got the first f-ing clue what a good version of one of these looks like” use case, but even there you have to take the output and pull and twist and kneed it (or at least ad some salt) until it says something compelling. Polished and compelling are not the same thing at all, and ChatGPT doesn’t have the foggiest idea what the difference is.
Put differently, if your reader falls asleep halfway through a post, it’s doesn’t much matter if the prose is so tight that it gleams.
I like that distinction between compelling and polished.
I wish that I knew what they looked like a little better!
I think with both, they’re somewhat easy to tell when you see them, but creating them from scratch? Oof, so much more difficult.
There are some writing-specific AI Agents being developed, so I’m kinda curious whether they’ll be able to do a better job with, I would assume, a much more refined training data set.