LLMs are bad for you

I do not use LLMs. I do not call them AI. I often speak up against LLMs.

This post is my current stance on LLMs. I am primarily exploring the topic from the perspective of someone that writes code, although a lot of the arguments apply in many other scenarios or fields.

There are over 20 references and sources cited below, simply because I wanted to keep this post short without going into a lot of details on every single point. To me, all of these points are obvious and important factors contributing to why I don't use LLMs.

I oppose LLMs for four main reasons:

• ethical,
• socio-political,
• psychological safety, and
• environmental.

Ethical

LLMs are trained with stolen work from the public (and private) domain. Books, websites, posts, open source code, everything people wrote for the benefit of other people is now fed as training data into LLMs, regardless of whether it's ethical or legal to do so. ¹

Whether you believe patents or copyright should be opposed or not, it's still blatantly obvious the same rules are not currently applied to companies training LLMs. A lot of specialized LLM outputs are often direct copies of existing work. ²

Even worse, the content generated using LLMs is growing in size, and it's getting harder to find legitimate, real content on the internet (even for people who deliberately avoid using LLM-enabled search engines, summaries, etc.). ³

LLMs are also not trained passively, there's tons of underpaid labor (often from the global south) that is basically used to oversee the training of these models. Similarly to content moderation teams on large social media platforms, these jobs are awful and often involve going through disturbing and traumatic content. ^{4 5}

LLMs are extensively used by the police, ICE, the IOF, and military forces in operations that casually infringe the rights of citizens, international law, and actively participate in citizen surveillance, deportation, genocide, and (primary imperialist) wars. ^{6 7 8 9}

Social/Political

Enshittification is the natural result of the profit motive being the driver for all goods and services. Racing to the bottom, extracting as much value as possible, and cutting costs have always been key pillars in companies (especially as they grow and distance themselves from any humanity and towards spreadsheet/capital-centric entities). This did not start with LLMs, but LLMs have definitely been supercharging and accelerating this trend.

On top of all of this, LLMs themselves are subject to enshittification. Right now they are still in the early phase of gaining widespread use, and even so they've already started showing ads ¹⁰. Next steps are likely to include cutting costs as much as possible, increasing the subscription price, selling your data/conversations, etc. ¹¹

Taylorism (ridiculously rebranded as scientific management) is the framework which, among other things, is about turning workers into replaceable cogs that focus on monotonous or singular tasks that they are expected to do as fast and cheap as possible. It's a significant factor in work alienation, and made possible by various advances in workplace automation.

LLMs bring the prospective of introducing taylorism to white collar (and artistic-adjacent jobs). Smart workers that have unique talents are terrible for business: they are costly, have opinions, care about quality, some may even have morals.

LLMs are also a tool for further displacing labour and extracting more value from workers, at everyone else's expense (which is somewhat a consequence of Taylorism). It means fewer jobs, worse quality, worse everything for everyone except the people owning the LLMs and companies using them: they get even more profits/capital. ¹²

Psychological

LLMs are deliberately built to mimic human connection, friendships, and maximize how much time people spend using them. Companies encourage people to use more LLMs because they can get workers to do work faster, while keeping all responsibility on the side of the worker (and mask that behind declaring the output is still your responsibility, while also increasing your load, and hide behind pseudo-science terms like prompt engineering).

All of this also significantly erodes human connections at work and throughout our society and deteriorates mental health and happiness.^{13 14 15}

An LLM is a powerful statistical model. And it will use the way you address it as input to craft familiar phrases, and over time it will craft and fine-tune its responses to you in ways that feel trustworthy, familiar, making you want to interact with it more. It's very hard for our brains to tell the difference between it and a real human; we can't evolve as fast as our technology.

I've seen a lot of people respect go into this kind of LLM psychosis, and I don't think I am smarter or better than them. I am honestly terrified to use LLMs. I can definitely see myself going into such a psychosis if I started using them (and were able to ignore all the other issues I have with them). ^{16 17}

Environmental

I don't think I need to go too deep into this one. They use a shit-ton of water and energy, they don't care about keeping said water clean, and they're loud as fuck (disturbing the environment around them as well). ¹⁸

They are also at the forefront of the resurgence of non-green energy throughout the world through its immense power demand.

Bonus round: they're not even competent

LLMs produce average code at best, and I want to keep learning and do better than that. I enjoy what I'm doing. If I were a musician or painter, I would be so because I enjoy creating music or paintings, not because I want to listen to my own music or look at my paintings.

Personally, using LLMs would take the fun and pleasure out of writing software. Even if it's boring boilerplate code, I welcome that. I want to feel the annoyance, and have that drive my refactoring later. I want to get a feel of the code, such that I can think of ways to improve it.

There's plenty of examples of LLMs getting things wrong, breaking production, deleting databases or home directories, hallucinating answers, etc. ^{19 20 21}

I would even go one further: there's no concept of understanding and of answers. It's just an incredibly expensive auto-complete feature. It has no mental context, no understanding. Even the AGENTS.md file is just a placebo ²².

Conclusions

I won't presume to tell anyone what to do. I will just share how I am dealing with this.

Language: I do not use the term AI. I deliberately always use the term LLM.

Personal use: I do not touch LLMs. I write and create the things I want to create all on my own. I do not ask LLMs questions, I use search engines (with LLM overviews turned off), and I try to avoid LLM-generated content (to the best of my knowledge).

I also actively try to avoid software tools and projects that use LLMs in any capacity, and try to gravitate towards alternatives that have a strong stance against LLM use. This is unfortunately not always possible (e.g. there's no viable alternative to the Linux kernel).

Personal relations: I try to advocate against LLM use and try to help the people around me both see my perspective (this is partly why I'm writing this blog post) and reduce their LLM use. I try, to the best of my abilities, not to be aggressive or personally attack people, especially if they're using LLMs exclusively at work in order to make a living (unless their job involves doing propaganda that advocates for LLM usage).

Professional use: I would ideally love to work for companies that have a clear anti-LLM stance. This is very hard at the moment, so for now I am settling for at least don't force me to use these things.