Raw LLM Responses

I use AI in a very specific way for my writing. I give very clear guidelines to deepseek to only look for technical English issues, not to give me an updated draft, and to list every instance of contention it finds. I then go through the feedback, look at the explanation, look at the context, and 95% of the time, I reject the suggestion. What I need it to do is find typos, missing words, homophones, and wrong word usage (as French is technically my first language and as such, I tend to get tricked by similar words with different meanings). What it often ends up doing is hallucinating punctuation issues and getting lost in the tenses, convinced that everything must be the same tense and anything that isn't is an error. It can do 270 points lists with 250 "tense shift" errors without seeing that maybe, given how many it's finding, it might be wrong in its analysis. Every now and then, out of curiosity, I ask it to analyse what a chapter does well and what it does wrong. When I do that, it takes the liberty of making edit suggestions. Anything pertaining to creativity, it fumbles royally and systematically. Because it can't memorize information, it sucks at writing anything coherent for more than 3 paragraphs. Its prose is formatted and repetitive. Oftentimes, it does word vomit with fantasy words, names, and organisations, and it loves to throw useless baited-hooks all over the place regardless of whether or not they make sense for your plot, theme, setting, or anything, really—the cheapest bait for a reaction at every occasion. It also loves purple prose, adverbs, and said-bookisms, which is confusing because you would think that they trained it on the best novels, but clearly, there is no barrier between its social media posts training and its literary works training. All of this is clearly a consequence of how LLMs function. They aren't actually intelligent. They're barely more intelligent than the next word prediction on your phone, which often takes a lot of charactersto get the next word right. What they do is yap, and it doesn't matter if the yapping is coherent as long as it is grammatically correct. As such, the best use in quality creative writing is as a specialized tool for specific applications. I use it for the proofreading, I don't let it edit or apply anything. I pretty much never use any of its change suggestions outside of fixing real errors. And even then, it still misses way too many things, and google docs corrector ends up finding them for me. Sometimes, I go over a draft with both, several times, submit to test readers, and they still find issues. Suffice to say I am not worried about AI taking creative writing jobs anytime soon, unless the industry heads have a big lapse in judgement and decide to pump out quantity over quality and ruin their consumer trust with terrible products. Which is unlikely, and would be answered by competition finally getting a chance to pull their pin out of the hat (I know that's a French expression, but I can't remember the English one about pulling something from a game).