I avoid using LLMs as a publisher and writer
lifehacky.net
lifehacky.net
That's AI.
In my case, I find the value with LLMs with respect to writing is consolidation. Use it to make outlines, not writing. One example is I record voice memos when driving or jogging and turn them into documents that can be the basis for all sorts of things. End of the day it saves me alot of time and arguably makes me more effective.
AI goes bad because it’s not smart, and it will pretend that it is. Figure out the things it does well for your scenario and exploit it.
It is not a reason to accept stupidity or incompetence. We should reject these things and demand better.
now imagine he can be scaled indefinitely
you thought software was bad today?
imagine Microsoft Teams in 5 years time
When I use Claude for code, for example, I am not asking it to write my code. I'm asking it to review what I have written and either suggest improvements or ways to troubleshoot a problem I am having. I also don't always follow its advice, either, but that depends on how much I understand the reply. Sometimes it outputs something that makes sense based on my current skill level, sometimes it proposes things that I know nothing about, in which case I ask it to break it down further so I can go search the Internet for more info and see if I can learn more, which pushes the limits of my skill level.
It works well, since my goal is to improve what I bring to the table and I have learned a lot, both about coding and about prompt engineering.
When I talk to other people, they accuse me of having the AI do all the work for me because that's how they approach their use of it. They want the AI to produce the whole project, as opposed to just using it as a second brain to offload some mental chunking. That's where Gen AI fails and the user spends all their time correcting convoluted mistakes caused by confabulation, unless they're making a simple monolithic program or script, but even then there's often hiccups.
Point is, Gen AI is a great tool, if you approach it with the right mindset. The hammer does not build the whole house, but it can certainly help.
It works but it simply not what most people want. If you love to code then you just abstracted away the most fun parts and have to only do the boring parts now. If you love to manage, well managing actual humans and seeing them grow and become independent is much more fulfilling.
On a side note, I feel like prompting and context management is something that is easier for me personally as a person with ADHD as I am already used to working with forms of intelligence that are different to my own. I am used to having to explicitly state my needs. My neurotypical co-workers get frustrated that the LLM can't read their minds and always tell me that it should know what they want. When it nudge them to give it more context and explain better what they need they often resist and say they shouldn't have to. Of course I am stereotyping a bit here but still an interesting observation.
Prompting is indeed a skill. Though I believe the skill ceiling will lower once tools get better so I wouldn't bank too much on it. What is going to be valuable for a long time is probably general software architecture skills.
I set up SSH certificates in my homelab last night with Claude Code. It was a somewhat aggravating process - I had to remind it a couple times of some syntax issues, and I'm not sure that it actually took less time than I would've taken to do it myself. And it also locked me out of my cluster when it YOLO'ed some changes it should not have. On the whole, one of the worst AI experiences I've had recently.
But I'm thrilled with it, TBH, because it got done, it works, I didn't have to beat my head against the wall for each little increment of progress, and while Claude Code was beating its own head against the wall, I was able to relax and 1) practice my French, and 2) read my book (Steven Levy's _Artificial Life_, which I recently saw excerpted on HN).
The general state of things is probably still pretty terrible. I know there're no end of irritations that I have with Claude Code, and everything else I've looked at is even less pleasant. But I feel like this might be going in a good direction.
*EDIT*: It should go without saying though that I'd much rather be mentoring a junior person, though, as you said.
People keep writing this sentence as if they aren't talking to the most tool-ed up group of humans in history.
I have no problems learning tools, from chorded key shortcuts to awk/sed/grep to configuring all three of my text editors (vim, sublime, and my IDE) to work for their various tasks.
Hell, I have preferred ligature fonts for different languages.
Sometimes tools aren't great and make your life harder, and it's not because folks aren't willing to learn the tool.
We have intelligent people using ai and claiming it’s useful.
And we have other intelligent people who’s saying it’s not useful.
I’m inclined to believe the former. You can’t be deluded about positives usefulness. But you can be about the negative simply by using the LLM in a half assed way and picking the most convenient conclusion without nuance.
If you honestly believe that, I've got a bridge to sell you.
But you can miss the positive results if you haven’t used LLMs recently or used agentic ai like cursor. it’s easy to miss the positives
If that were true, then we would not have the Dunning-Kruger effect. Regardless of your intelligence, all of us are susceptible to a cognitive bias that makes us think that we are better than we actually are at some things.
The classical case used to demonstrate the Dunning-Kruger effect is self-assessment. That is, how well you think you can do a task. Rating the performance of a task - which is precisely what is happening here!
People are shit indicators of their own performance. With a great new placebo tool, people are incredibly likely to say it improved their life. Even though it did nothing at all.
Being deluded about positive usefulness is normal.
I mean as much as we complain about LLMs hallucinating, here's an example of a human making shit up out of thin air. What's going on here is NOT self assessment. It's obviously assessment of an LLM.
Additionally the Dunning-Kruger effect like all of psychology stands on shaky ground.
DKE has been confirmed more than anything else. It was one of the few things not hit by the replication crisis.
You're assessing how well you do, when aided or not, by a tool. That's still self assessment, I'm afraid.
And that self-assessment is flawed. [0]
Aid? I assess the output of the tool. Whether that tool aids me is another topic all together. I am not using the tool to augment my existing abilities.
The tool is literally doing the task for me and I am evaluating the results afterwords. This is not some wrench that augments my existing strength. This is more of an assistant than a tool but wording can be manipulated so that assistants can also be thought of as tools. Let's not manipulate the wording to be in our favor and go for intent. The intent here is that clearly the LLM is different from a wrench. When you evaluate a wrench you also evaluate yourself because you are operating the wrench. When you evaluate an LLM, you are not operating it. You gave it a prompt and it went off on it's own to do something.
https://www.scientificamerican.com/article/the-dunning-kruge...
This is what I mean by dunning kruger. Either way, whether it's legit or not my points still stand.
>So... Attack first? That what you're going with?
What is this? Attack first? Who is getting attacked here?
> You just made that up... I mean as much as we complain about LLMs hallucinating, here's an example of a human making shit up out of thin air
You called me a liar, delusional, and accused me of making up shit.
As you appear to ignore all things contrary to your current opinion - as the article you linked points out it effects everyone; and have acted like a dick, this conversation is now over.
But that's simply not true.
Not only have I used these tools and found them to be unhelpful to me, I have good reasons why I don't think they are helpful. I can even give two modalities in which I find them actively unhelpful:
- for creative work, they don't allow me to chew over the details which I find important to struggle with as I express my thoughts and how to communicate them
- for rote lookup or facts, I either understand the underlying material such that my code completion or templating tools are faster and clearer for me or I probably need to struggle with the underlying complexities until I can generalize the problem myself.
You simply assume that I'm not, like, a 47 year old with an annoying theory of mind and learning and who has conceptual models for how I learn things based on almost 3 decades of teaching hundreds of students, coaching dozens of my cohort, and learning many skills across several domains.
Which is fine. I am old enough that "you're holding it wrong" is something I've seen several times in my life.
But at the end of the day, all you have is the usual "you're holding it wrong" objection that most folks have to technology that doesn't actually fit well.
I will give you some free advice, totally worth what you're paying for it.
It is indeed entirely possible that humans are quite often "deluded about positives [sic] usefulness" of different tools. That delusion can often be a difficult or painful lesson. I've got a lot of tendon issues from rock climbing and bad scalar patterns in my clarinet playing to prove that well enough for myself.
I suggest that if you really believe that anything which helps you in some short term kind of way won't hamper you in your future endeavors, you might want to question that belief.
If you can't think of any examples (cocaine being one easy example) then I suggest that you don't know enough about the world to be conjecturing about it as you have been doing here.
In any case, good luck. Clearly all the people disagreeing with you here are wrong.
Doesn't prove your case. Plenty of instances where everyone is wrong and one person is right. Lead for example was once thought by everyone to be healthy. Very few people considered it toxic.
>I will give you some free advice, totally worth what you're paying for it.
Could be completely useless advice and totally worthless. You declaring it worth it does not suddenly make the advice valuable. In fact I'm anticipating negative value.
>It is indeed entirely possible that humans are quite often "deluded about positives [sic] usefulness" of different tools. That delusion can often be a difficult or painful lesson. I've got a lot of tendon issues from rock climbing and bad scalar patterns in my clarinet playing to prove that well enough for myself.
Of course it's possible. It's just more rare. I put values into a calculator. The calculator does a calculation faster than me. Was that delusion? There clear example. Can you give me a clear example of the alternative? Where you use a tool it only feels useful but isn't. Your rock climbing examples feel like a bit of a stretch. In fact they feel like counter examples, you eventually noted that they aren't useful.
>If you can't think of any examples (cocaine being one easy example) then I suggest that you don't know enough about the world to be conjecturing about it as you have been doing here.
I suggest that you actually don't know enough about the world compared to me given my 60+ years being alive. Your attitude is rude and condescending. But you know I often wonder what would trigger someone to be like this? Like why can't you be impartial and just give counter evidence? Why did you have to approach this whole thing with this attitude of "Let me give you a fucking tip".. Is it because I hit a nerve? Because one aspect of what I'm talking about is right and it's hard to face the truth? I don't know. I can only speculate.
Cocaine was at one point in time not known to be addictive. You could be right here with that analogy. But we can't fully prove it can we? The answers given by an LLM are too varied to form a definitive answer. Cocaine EVENTUALLY outputs a definitive symptom of addiction and other bad outcomes that are statistically significant. So even though at one point in time we didn't know... over time cocaine yielded definitive answers. but LLMs used for programming? What are we even measuring? We don't even know. So it's hard to see some definitive answer revealing itself over time. All I see are endless debates where I'm right, and I can't convince a kid like you that you're wrong.
1) people who reject it completely for whatever reason. 2) people who use it lazily and produce a lot of garbage (lets be honest, this is probably going to happen a lot which is why maybe group #1 hates this future. reminds me of the outsourcing era) 3) people who selectively use it to their advantage.
no point in groups 1 and 3 trying to convince each other of anything.
I think your explanation for group 1 is true to a degree but have two other additional explanations: (1) Some element of group 1 is ideologically opposed. It might be copyright, or Luddism, or some other concern for our fellow humans. (2) Some are deluded into thinking there are only two groups and that group 3 people are all delusional.
Although it is probably an uphill battle I do think both groups 1 and 3 have things to learn from each other.
A guy I went to highschool with complains endlessly about AI generated art and graphics (he's an artist) and like you, just wants to bury his head in the sand.
Consumers don't care if art is generated by AI or humans and in a short period of time, you won't be able to tell the difference.
With the money being poured into AI by all major tech companies, you will be unemployed if you don't keep up with AI.
Maybe not yet. The real "art" consumers were always very sensitive and asking for originality (thus scarcity). It is an essential principle of the art that it is a result of thousands/millions of deliberate choices. If you use machine for creation, you less choices. You delegate most of your talented/crazy/hard choices to the model (which is based on such choices of already talented but combines them in a random way). The result is thin, diluted even it seems like deliberate. In my opinion the most art lovers will continue to seek for the dense art made by human, asking for some kind of proof. :) The real art will be even more appreciated. I guess.
Anyone who spent time learning the AI tools over that period of time has basically wasted their time. Working with agents is nothing like prompt engineering. I imagine whatever comes after will be nothing like agents etc. Sounds like those who try to keep up with AI will be equally unemployed.
I suppose I shouldn't care too much. Less competition for people like me that have embraced the change.
I mean you're not wrong the serious people drop into assembly when they need too. Even if you work in a context where you can't or don't drop down into assembly being able to make your own compilers is incredibly useful.
My point was that comparing the rise of AI tooling to the rise of HLL compilers is a much better comparison than comparing it to crypto.
HLL compilers were originally seen as crutches and inferior tools and that "real" programmers used assembly. Compiler-generated code was derided as inefficient and ugly.
And it was! In the early days, a good programmer who knew the machine could outdo the compiler. But that didn't stop a huge expansion of new programmers who could write COBOL and FORTRAN but never learned assembly. And the compilers got better over time. These days it's a rare wizard who can outdo a compiler's optimizations, and it takes multiple orders of magnitude longer for those rare humans to achieve it.
LLM tooling isn't going away. Even in these very early days, it enables non-programmers to construct basic applications that work, using English requests! And the tools have gotten better on almost a monthly basis.
You can like them or not like them, just like the early programmers could like or not like compilers. But dismissing them as analogous to empty crypto hype is a bad comparison.
Then again, maybe I'm too old now and being left behind if I remember the old hype like this....
The entirety of the tech field is constantly hyping the current technology out of FOMO. Whether or not it works out in the future it's always the same damn argument.
If LLMs eventually become useful to me, I’ll adopt LLMs, I suppose. Until then, we’ll, fool me once…
AI is like the absolute worst outsourced devs I've ever worked with - enthusiastically saying "yes I can do that" to everything and then delivering absolute garbage that takes me longer to fix/convince them to do right than it would have taken for me to just do it myself.
It works best if you keep close to mainstream styles and if you keep it easy and straight-forward.
I personally think an LLM could help with some of this, and this is something I've been thinking about the past few days. But I'd have to build a pipeline and figure out a way to make it amplify what I like about my voice rather than have me speak through its voice.
I used to have a sort of puritanical view of art. And I think a younger version of myself would have been low key horrified at the amount of work in great art that was delegated to assistants. E.g. a sculptor (say Michelangelo) would typically make a miniature to get approval from patrons and the final sculpture would be scaled up. Hopefully for major works, the master was closely involved in the scaling up. But I would bet that for minor works (or maybe even the typical work) assistants did a lot of the final piece.
The same happens (and has always happened) with successful authors. Having assistants do bits here or there. Maybe some research, maybe some corrections, maybe some drafts. Possibly relying on them increasingly as you get later in your career or if you're commercially successful enough to need to produce at greater scale.
I think LLMs will obviously fit into these existing processes. They'll also be used to generate content that is never checked by a human before shipping. I think the right balance is yet to be seen, and there will always be people who insist on more deliberate and slower practices over mass production.
[0] Aside from internet comments of course, which are mostly stream of consciousness.
https://en.wikipedia.org/wiki/David_(Michelangelo)#Process
Maybe later he got lazier. I haven't really heard of famous authors using assistants for drafts instead of research (I don't mean commercial authors like Stephen King).
Even research many authors simply could not afford.
I don't see where the article says he worked alone on David. It does seem that he used a miniature (bozzetto) and then scaled up with a pointing machine. One possibility is he made the miniature and had assistants rough out the upscaled copy before doing the fine work himself. Essentially, using the assistants to do the work you'd do on a band saw if you were carving out of wood.
> I haven't really heard of famous authors using assistants for drafts instead of research (I don't mean commercial authors like Stephen King).
Restricting to non-commercial authors would narrow it down since hiring assistants to write drafts probably only makes financial sense if the cost of the assistant is less than the cost of your time it would take drafting.
Alexander Dumas is maybe a bit higher brow than Stephen King
> He founded a production studio, staffed with writers who turned out hundreds of stories, all subject to his personal direction, editing, and additions. From 1839 to 1841, Dumas, with the assistance of several friends, compiled Celebrated Crimes, an eight-volume collection of essays on famous criminals and crimes from European history. https://en.wikipedia.org/wiki/Alexandre_Dumas
But in general I agree, drafts are often the heart of the work and it's where I'd expect masters to spend a lot of their time. Similarly with the statue miniatures.
I like the perspective of "choices" during creation. It is an essential principle of the real art that it is a result of thousands/millions of deliberate choices. This is what we admire on the art. If you use mostly machine (or other kind of ways that decide instead and for you) for creation, you as an creator simply do less choices.
In this case, you delegate many of your experienced/crazy/hard decisions to the model (which is based on such decision made already by other artists but combines them in a random way). It is like decompressing JPG – some things are just hallucinated by machine.
From the perspective of pure human creativity, the result is thin, diluted. Even it seems like deliberate. In my opinion art lovers will seek for the dense art made by human, maybe asking even more for some kind of "proof" of the human-based process. What do you think?
When I was in high school I really leaned on friends for edits. Not just because of the changes they would make (though they often did make great suggestions), but for the changes I would make to their changes after. That’s what would inevitably turn my papers from a B into an A. It’s basically the same thing in principle. I need to see something written in a way I would not write it or I start talking in circles/get too wordy. And yes this comment is an example of that haha
Some technology is just capital trying to find growth in new markets and doesn't represent a fundamental value add.
VR/AR is tech that is nowhere near ready so its premature to make judgements there. I am bullish on XR.
In terms of value-add, many times technology needs to be invented first before we decide whether its worthwhile. Nobody asked for computers in every home, smartphones in every pocket etc.
I expect this will be around the time that websites are no longer a thing and we see companies directly pumping information into AI agents which are then postured as the only mechanism for receiving certain information.
As an example, imagine Fandango becoming such a powerful movie agent that theaters no longer need websites. You don't ask it questions. Instead, it notifies YOU based on what it knows about your schedule, your preferences, your income, etc. Right around 5pm it says "Hey did you know F1 is showing down the street from you at Regal Cinema in IMAX tonight at 7:30? That will give you time to finish your 30 minute commute and pickup your girlfriend! Want me to send her a notification that you want to do this?"
People install a litany of agents on their smartphones, and they train their agents based on their personal preferences etc, and the agents then become the advertisers directly feeding relevant and timely information to you that maximizes your spend.
MCP will probably kill the web as we know it.
It will be controlling and disempowering - manipulative personality-profiled "suggestions" with a much higher click rate than anything we have today.
And the richer you are, the more freedom you'll have to opt out and manage your own decisions.
This. I need access to banking , maps and 2FA. If I could use a dumb phone, with just a camera, GPS and whatsapp, I would use it.
The ability to transfer funds is something I'm now fine doing via a dedicated device with a dedicated password manager account, and I'm fine uninstalling banks' apps from my phone and dis-enrolling cell phone numbers.
Given the wanton collection and sale of my data by many entities I hadn't expected (naivety on my part), I've restricted access to critical services by device and or web browser only. It's had the added bonus of making me more purposeful in what I'm doing, albeit at the expense of a convenience. Ultimately, I'm not saying my approach is right for everyone, but for me it's felt great to take stock of historical behavior and act accordingly.
Not to take any words it gives but read what it says and decide if those things are true, if so, make edits. I am not saying it is a great editor but it is better than any other resource he has access to as a teenager. Yeah better than me or his mom
- Writing groups. They often have sessions that provide feedback and also help writers find/build a sense of community. Your son would also get to listen to other writers talk about their work, problems they’ve run into and overcome, and other aspects of their craft.
- School (sometimes library) writing workshops. This helps students develop bonds with their peers and helps both students: the ones giving feedback are learning to be better editors.
Both of these offer a lot of value in terms of community building and also getting feedback from people vested in the the craft of writing.
For one, the world doesn't need to be that way, I.e. We don't need to "leave behind" anyone who doesn't immediately adopt every single piece of new technology. That's simple callousness and doesn't need to be ruthlessly obeyed.
And for two, it's provably false. What is "the future?" VR? The metaverse? Blockchain? NFTs? Hydrogen cells? Driverless cars? There has been exactly ZERO penalty for not embracing any of these, all sold to us by hucksters as "the future".
We're going to have to keep using a classic piece of technology for a while now, the Mark 1 Human Brain, to properly evaluate new technology and what its place in our society is, and we oughn't be reliant on profound-seeming but overly-simplistic quotes as that.
Be a little more discerning, and think for yourself before you lose the ability to.
Do you have kids? Outside of discipline, and even there, I want to have a positive relationship with my sons.
My oldest knows that I am not a writer, there are a ton areas that I can give legit good advice. I can actually have a fun conversation about his stories, but I have no qualifications to tell him what he might want to change. I can say what I like but my likes/dislikes are not what an editor does. I actually stay away from dislikes on his writing because who cares what I don’t like.
I would rather encourage him to write, write more, and get some level of feedback even if I don’t think my feedback is valuable.
LLMs have been trained on likely all published books, it IS more qualified than me.
If he continues to write and gets good enough should he seek a human editor sure.
But I never want me to be a reason he backs away from something because my feedback was wrong. It is easier for people to take critical feedback from a computer than their parents. Kids want to please and I don’t want him writing stuff because he think it will be up my alley.
You think you shouldn’t give advice because your feedback is not valuable and may even cause your son to give up writing, but you have so far given no reason why AI wouldn’t. From the entire ChatGPT “glazing” accident I can also argue that the AI can also give bad feedback. Heck most mainstream models are fine tuned to sounds like a secretary that never says no.
Sorry if this sounds rude, but it feels like the real reason you ask your son to get AI feedback is to avoid being personally responsible for mistakes. You are not using AI as a tool, you are using it as an scapegoat in case anything goes wrong.
It has also be trained on worthless comments on the internet, so that’s not a great indicator.
Do you want an LLM to be the reason? You can explain that your Feedback is opinionated or biased. And you know him better than any machine ever will
I formed the habit of exporting entire chats to Markdown and found them useless. Whatever I found of useful from a given response either sparked a superseding thought of my own or was just a reiteration of my own intuitive thoughts.
I've moved from ChatGPT to Claude. The results are practically the same as far as I can tell (although my gut tells me I get better code from Claude) but the I think Anthropic have a better feel for response readability. Sometimes processing a ChatGPT response is like reading a white paper.
Other than that, LLMs get predictable to me after a while and I get why people suspect that they're starting to plateau.
But I still don't like that the same model struggles w/ my projects...
We are a publisher which succeeded due to the highest-quality translations. Our readers appreciated it and ask for it. Czech language is very rich and these machines are not able to make the most of it. The non-fiction sphere needs a lot of fact-checking e.g. in local and field terminology too. So even we can imagine the process of translation could be technically shortened by machine translation, it would probably ruin our reputation in a long term.
At least for now...
Now consider someone further down the scale - someone at the 75th, 50th or 25th percentile. The output of an LLM very quickly goes from "much worse than what I could produce" to "as good as anything I could produce" to "immeasurably better than anything I could hope to ever produce".
Perhaps LLMs can move someone's results from the 25th percentile to the 50th for a single task. (Although there's probably a much more nuanced discussion to be had about that: people with poor writing skills can still have unique, valuable, and interesting perspectives that get destroyed in the median-ization of current LLM output.) But after a couple years of using LLMs regularly, I fear that whatever actual talent they have will atrophy below their starting point.
However good writing is a skill you can get good at with enough practice. Read a lot, write a lot of garbage, consult more experienced writers and eventually you will write readable articles soon. Do 10-100x more of that and you will be pretty great. The rest is some kind skill and experience in many other fields than writing which will inform how to write even better. Some of it is intelligence, luck, great mentors and perhaps something we call talent even. As with most things you can get far just by working diligently a lot.
That does, to my mind, explain all the vengeful "haw haw, you're all going to get left behind" comments from some LLM proponents. They actually do get benefit from LLMs, unlike the highest part of the scale who are overrepresented on HN, without realizing what that implies and they think they can overtake the highest part of the scale by using them. Well, we'll see.
Writing is entirely different, and for some reason, generic writing even when polished (ChatGPT-esque tone) is so much more intolerable than say AI-generated imagery. Images can blend in the background, reading takes active processing so we're much more sensitive. And for the end user of a product, they care 0 or next to 0 about AI code.
Very interesting point!
Coding, robotics, navigation of constrained data spaces such as translation, tagging, indexing, logging, parsing, data transformations… those are all strong target candidates for transformer architecture automation.
Creative thought is not.
To be fair, I also don't like using Copilot when working on code. In many cases it turns into a weird experience when the agent generates the next line(s) and I basically become a discriminator judging if the thing really understands my problem and solution. To be honest, it's boring even if eventually it might make me turn in code faster.
With that said, I cannot ignore that LLMs are happening, and this is the future. The models keep improving but more importantly, the ecosystem keeps improving with things like MCP and better defined context for LLM tools.
We might be looking at a somewhat grim prospect. But like it or not, this is the future. Adapt and survive.
For me survival means: - continuing to do my best at the language level – even if more people would start be gradually satisfied with less - I just believe that education, critical thinking and evidence-based principles are at core of humanity progress and one day it will make comeback - I am ok with smaller income and not wishing to exchange it for creating bullshit
The adaptation for me means: - generally: stay open-minded - I have to understand and somehow accept that the prospect is a bit grim but not to fall into some extreme and doom thinking - I have to explore new ways how to augment human-oriented creativity (with or without these tools)
What do you think?
Well, it's a trap. You see a snippet is right, you accept it. Next time you do it faster, and faster. And then you get one that seems right but it's not. If you're lucky, it will cause an error.
It's either the tech is advancing so quickly that many people can't keep up, or simply the cost of adapting outweighs the potential profit from their remaining careers, even when taking the new tech into account.
Well, it has a problem with my use of the Oxford comma, for one. Because a huge amount of the corpus is American English, and mine ain't. So it fails on grammar repeatedly.
And if you introduce any words of your own, it will sometimes randomly correct them to something else, and randomly correct other words to the made up ones. And it can't always tell when it's made such a change. And sometimes it does that even if you're just mixing existing languages like French or English. So you can make it useless for spellcheck by touching more than one language.
I do keep trying, despite the fact my stuff has been stolen and is in the training data, because of all the proselytising, but right now... No AI is useful for my writing. Not even just for grammar and spelling.