Using AI in writing

[HumiliationInc]

Hope it's alright if I add to this discussion, because some posters wanted the topic to be "done."

I work in a sector where AI has the potential to do a lot of good and it has equal potential to do a lot of societal harm. Unfortunately, things seem to being leaning toward the later. I'm not wholesale anti-AI. I supervise a data scientist (though their projects and workflow really don't use ML/AI), and there are a few tasks at work I find it's useful for. But I do think we as a society really need have a serious discussion about its implications. IME, the people who overhype AI are either techbros looking to cash in on it or people who just don't understand it all (i.e., don't use it and don't have training in ML or data science) but read about it in Harvard Business Review articles and think it can literally do anything.

When it comes to writing, one thing I noticed right away when LLMs hit the mainstream a few years ago is that they can write very technically correct prose—its output is very Struck and White—but it really struggles with creating anything creative or original sounding. AI in its current form isn't going to write the next Infinite Jest or Northanger Abbey or Ulysses or American Psycho. It can't write anything with multiple layers of meaning or that relies on deep and complex semiotics. That's because it doesn't actually "understand" what it's writing; it can only mimic what it's trained on. And I'm skeptical that it ever will be able to write something truly fresh and creative, regardless what the futurists say about AGI.

However, at a micro level, it can be effective. I find that it can be useful if I'm struggling with sentence structuring. But I only use it maybe once every couple thousand words. It is EXCELLENT as a spellchecker though, much better than the onboard spellcheckers in MS Word or OpenOffice Writer. Though for the stuff we write on this forum, most mainstream LLMs are going to throw the filters up and wave the cyber naughty finger at us! A year or so ago, I did find an LLM that didn't have a lot of the naughty filters. I wasn't huge on the faux-libertarian/alt. right coded copy on its website, but I gave it a try and it did produce output that other LLMs refused to. But it only gave something like 10 free prompts and I wasn't about to put my real name into a credit card form to create porny stuff.

I do like to play around with AI images (and ironically enjoy some of their butchered outputs). I mainly do that because I'm a visual person and an evocative picture can inspire me or if I want to give a visual representation to a story. I'd never use it to actually replace an artist, though. I've commissioned some art and comics throughout my adult life and I wouldn't ever try to get AI to take the place of something I want to last as legitimate art. Also, I find that after producing so many pictures, it really starts entering slop territory by creating pictures that are only slightly different from one another. I mainly use Bing for AI images as I like the slight uncanny valley look it gives people and its pretty easy to create something useful. I haven't gotten into the more advanced programs that use loRAs and what not. I don't have time for that.

[Vela Nanashi]

I enjoy using AIs mostly for hearing the echo of the shadow of mankind, but you should never just trust the AI output, but it can be helpful sometimes to give an overview and places to go to research more, and I like it for that, I also enjoy some image creation things, and have had not boring role plays with the AI, luckily that one allowed editing the output so that the AI got better at writing like you want it to. Still most people are more interesting than that AI was.

I believe personally that the path we are on is mostly a toy however, if we truly want them to know and learn and understand we need to get closer to how real brains work, while also using what computers can do better than us. Ideally if I were to decide for AI stuff I would want the ingested material to be given royalties that are part of the sum the AI companies make after they pay for all the expenses, a large part of it. That way people would potentially willingly give the AI more material to work with, and to facilitate such royalties I want the original work to be in a huge archive, linked to the ripples it created in the matrix, I also want concepts to have links to other concepts, synonyms, antonyms, this is a part of that, this group of things have this property, etc, I know AI scientists have seen some trace of such a clustering inside the multi dimensional search space that is the AI, but it would be better if we could more directly encode some relationships. Also I would like for it to keep track of sources for pieces of information, and then for users to be able to say "I trust this person, I don't trust that person, this person has good knowledge on food, but don't ever ask them about cars" etc, and AI really needs to know that just because someone strongly believes something to be true and is telling you that thing in good faith, genuinely trying to help you, does not mean that is true, they may be harming you without knowing and be really upset with themselves over that. But just eating text and images the way it currently is, no we really won't get AGI anytime soon, it needs more.

Also there is a huge efficiency problem in AI right now, so it would be good if it was given non AI tools to use for many tasks, like math for example, and also if some parts of the neural net could be cut off into smaller areas where it is a true expert on that, and I hear there are some experiments with changing the link type from a floating point to either there is a link or not, and if there is a link it is -1 or 1, that allows turning multiplications into additions instead and would make things a lot cheaper to do.

But I guess I derailed the conversation here. For me when I tried to use AI for actual writing and I saw it in the role play too, it is clear that the AI can't remember what objects are, or does so very poorly, and even when talking with it about non imaginary things, but things in the real world, it treats everything said about the subject as true it feels like. Unless you are extremely specific, and even then it can fail.

You will notice the issue most if you use the AI in a field or area you are good at, that is when you will see how much crap it is failing at, while if you do it in an area you don't know, you will believe its confident answers that are utterly incorrect at times, it is really quite frustrating, but it is not really unique to AIs, they are just forced to answer you, while humans are not and can ignore you, so they don't have to invent something to appease you.

[RapeU]

Also there is a huge efficiency problem in AI right now, so it would be good if it was given non AI tools to use for many tasks, like math for example... (snip)

You will notice the issue most if you use the AI in a field or area you are good at, that is when you will see how much crap it is failing at, while if you do it in an area you don't know, you will believe its confident answers that are utterly incorrect at times.

Yup. I used AI tools during graduate level math courses and I noticed right away "wait that's not how you're supposed to be doing it." I had to work at getting the AI to understand exactly how to do higher level math problems. There's literally tons upon tons of job openings now where a math person can work on feeding AI math problems and correcting it. I'd do it, but they don't offer health insurance. Bastards.

A student in my Precalculus class one day used AI to do calculus level work and said "yeah, sir I'm working!" I shook my head and said "you just used snapchat AI cause I haven't even taught you integration yet." For those math brained like me, the student answered a trigonometric identity problem by integrating it. For the non math brained, it's like teaching someone to crawl and they all of a sudden start walking.

I'm not well versed in English classes, I've written papers for different college classes that were just pure garbage. However, even I can tell that prolonged use of AI while writing a story makes the story as a whole clunky. I've said it before and I'll say it again, the longer a story is written by AI, the more syntax errors you'll have. You'll also have lack of character development, continuity errors, and then there will be things that don't make logical sense. Best example of that is a story here that had a parking garage lined up in front of a runway. Massive safety issue for both the parking garage and the planes.

Best way to use AI for writing is to write a few sentences, ask it ideas to enhance your sentence, then edit the AI response unless you feel like that is how you would say it in your own words. Worst way is to have the AI write most or all of it, fail to put it in your own words, and try to pass it off as a story. If you depend on AI to write a story, you don't have a story. Instead you have words arranged in a certain order by a machine that isn't capable of complex emotional thought.

[Vela Nanashi]

After seeing the trouble it has with editing existing text, I don't think it is there yet for writing at least in my style and I happen to like my style

Still I am sure we will one day have actually thinking machines, if we want that, one thing those would need though is to be able to think on their own without being poked by a prompt, but I suggest we also gift them what we don't have, an "I am bored wake me when someone wants me" button