Are LLMs inherently unethical?

In my view, large language models are just tools.

Just like all tools they can have interesting uses –

LLM agents; summarization, even in medical settings; named entity extraction; sentiment analysis and moderation to relieve the burden from people being traumatized by moderating huge social networks; a form of therapy for those who can't access, afford, or trust traditional therapy; grammar checking, like a better Grammarly; simple first-pass writing critique as a beta reader applying provided rubrics for judgement; text transformation, such as converting natural language to a JSON schema, a prerequisite for good human language interfaces with computers; internet research; vision; filling in parts of images; concept art; generating business memos and briefs; writing formal emails; getting the basic scaffolding of a legal document out before you check it; rubber duck debugging; brainstorming

– and really bad uses –

programming; search (without something like Perplexity); filling the internet with slop; running massive bot farms to manipulate political opinion on social media sites; creating nonconsensual deepfakes; shitty customer service; making insurance or hiring decisions; creating business plans; .

They can also be used by bad actors towards disastrous ends even when they're being used for conceivably-good proximate purposes –

as an excuse to cut jobs, make people work faster, decrease the quality of the work, deskill people, and control people –

or positive ends – make people more productive, so they can work less, and endure less tedium, to produce the same amount, help disabled people, etc –

…just like any other tool.

But that's not how people approach it. Instead, they approach it as some kind of inherently irredeemable and monstrous ultimate evil that is, and must, literally destroy everything, from human minds to education to democracy to the environment to labor rights. Anyone who has the temerity to have a nuanced view – to agree that the way capitalists are implementing and using LLM is bad, but say that maybe some of the ethical arguments against it are unjustified, or maybe it has some uses that are worth the costs – is utterly dragged through the mud.

This behavior/rhetoric couldn't, I believe, be justified if it was just in response to the environmental impact of LLM, or the way data labellers are exploited: the answer to that, like any other thing in our capitalist economy that's fine in concept but produced in an environmentally or other exploitative way, such as computers themselves, shoes, bananas, etc., would be some combination of scaling back, internalizing externalities, changing how it's implemented to something that's slower and more deliberate, all achieved through regulation or activism or collective action; not to disavow the technology altogether. (This is even assuming the environmental impact of LLM is meaningful; I don't find it to be).

In fact, all of the negative environmental pieces on LLM (two representative examples: 1 and 2) fall afoul of a consistent series of distortions that to me indicate they aren't written in good faith – that unconsciously, the reasoning is motivated by something else:

• failure to provide any context in the form of the energy usage of actually comparable activities we already do and aren't having an environmental moral panic about, such as video streaming;
• failure to take into account alternative methods of running large language models, such as local language models running on power efficient architectures like Apple Silicon;
• the unjustified assumption that energy usage will continue to hocky stick upward forever, ignoring the rise of efficiency techniques on both the hardware and software side such as mixture of experts, the Matryoshka Transformer architecture, quantization, prompt caching, speculative decoding, per-layer embedding, distillation to smaller model sizes, conditional parameter loading, and more;
• comparison to the aggregate power usage of other widespread activities like computer gaming, since its power use may seem outsized only because of how centralized it is;
• and more I can't think of right now.

It also can't be justified in response to the fact that LLM might automate many jobs. The response to that is to try fight to change who benefits from that automation, to change who controls it and how it is deployed, so it's used to make workers able to work less to produce the same amount (and get paid the same), or to allow them to produce more for the same amount of work (and thus get paid more), instead of being used to fire workers. Hell, even if that's impossible, we know how automation plays out for society in the long run: greater wealth and ease and productivity for everyone. Yes, there is an adjustment period where a lot of people lose their jobs – and you can't accuse me of being callous here, since I'm one of the people on the chopping block if this latest round of automation genuinely leads to long term job loss in my trade – and we should do everything we can to support people materially, financially, emotionally, and educationally as that happens, and it would be better if it didn't have to happen, but again, if the concern were truly about lost interim jobs during a revolution in automation, the rhetoric wouldn't look like this, would it?

Fundamentally, I think the core of the hatred for LLM, then, stems from something deeper. As this smug anti-LLM screed states very clearly, the core reason that the anti-LLM crowd views LLM the way it does – as inherently evil – is because they've bought fully into a narrow-minded, almost symbolic-capitalist, mentality. If and only if you genuinely believe that something can only be created through what you believe to be exploitation, then it would be justified and to act the way these people do.

Thus while I wish anti-LLM people's beliefs were such that discussing LLM "on the merits," and how to scale it back or make it more efficient or use it wisely, was something they could come to the table on, thier moral system is such that they are forced to believe LLM is inherently evil, because it requires mass copyright "theft" and "plagerism" – i.e., they're fully bought into IP.

<<new things>>Because yeah in theory you could make a copyright-violation free LLM, but it'd inherently be a whole lot less useful, in my opinion probably not even useful enough to break even for the time and energy it'd cost, because machine learning doesn't extrapolate from what it's learned to new things in the way human minds do. It just interpolates between things it's already learned – I like the term "imitative intelligence" for what it does – so if it doesn't have a particular reasoning pattern or type of problem or piece of common sense or whatever feature in its dataset, it can't do it or tasks like it or involving pieces of it very well. Now, it learns extremely abstract, very much semantic "linguistic laws of motion" about those things, it isn't "plagerising," but the need for a large amount of very diverse data is inherent to the system. That's why large language models only began to come to fruition once the internet matured: the collective noosphere was a prerequisite for creating intelligences that could imitate us.

So, if anti-LLM advocates view enjoying or using something they've already created, that they bear no cost for the further use of, that they publicly released, as "exploitation", simply because someone got something out of their labor and didn't pay rent to them for that public good (the classic "pay me for walking past my bakery and smelling the bread"), then like… yeah. LLM is exploitative.

Personally, it just so happens that I do not give a flying fuck about IP and never did – in fact I hate it, even when artists play the "I'm just a little guy" card. It is not necessary to make a living creating "intellectual property," and it only serves to prop up a system that furthers the invation of our rights and privacy and control over our own property, as well as the encroachment of private owners – whether individual or corporate – into the noosphere, and foster territorial, tribalistic approaches to ideas and expressions. Sell copies of your work only as physical items, or special physical editions that someone might want to buy even if they have a digital copy, to pay for your work. Or set up a Patreon, so that people who appreciate your work and want you to continue doing it can support you in that, or do commissions, where, like Pateron, payment is for the performance of labor to create something new, instead of raking in money for something you already did.

I really don't believe that if you make information or ideas available to the world, you get to dictate what someone else does with them after that point; I think the idea of closing off your work, saying "no, scientists and engineers can't make something new/interesting/useful out of the collective stuff humanity has produced, because they're not paying me for the privilege, even though it costs me nothing and requires no new labor from me", while understandable from the standpoint of fear about joblessness and lack of income under capitalism, is a fundamentally stupid and honestly kind of gross view to hold in general.

But that's what they hold to, and from that perspective, LLMs truly can't really be neutral.

Update:

1. One of the requirements for "ethical" AI that smug anti AI screed is this:

   To satisfy attribution and other prominent notice requirements in an ethical, respectful way, the model must also reference the sources it used for any particular output, and the licenses of those. You cannot claim to respect the license of your sources without respecting the attribution and prominent notice requirements. You cannot make your model practically usable without listing the sources for any given output.

   As it turns out, this is precisely what Google's AIs do:

   This field may be populated with recitation information for any text included in the content. These are passages that are "recited" from copyrighted material in the foundational LLM's training data.

   Basically, it seems that any model output that is substantially similar to something from its training data (which Google also keeps on its servers) and which is over a certain length, is automatically flagged at the API level and a citation to the original source material is added to the result from any batch request. It even is able to put license data directly in the citation object when it could be automatically retrieved from the source (such as with Github), but since it provides an original URI, anyone who's curious should be able to find out what the license of the original work is themselves. Moreover, it provides the exact portions of the output that are or may be recitations. The accuracy of the system, form my testing, also seems to indicate this is done at the API level, not just asking the model and hoping for something that isn't hallucinated — and that would make sense, since flexible textual search through gigantic databases of text is Google's specialty, and a well understood computational problem. There's no way to turn this off either. So once again (as with Gary Marcus complaining that AIs don't do search to ground their facts, when they actually do when you don't manually turn it off), this is a case of anti-AI activists being out of date on the facts, usually because they're willfully and proudly ignorant.

2. I was also possibly wrong too: there is some preliminary research that suggests that allowing web crawlers that collect training data for large language models to respect web crawling opt-outs does not significantly decrease the performance of the resulting model, only its knowledge-base of specific topics or fields, and since IMO we should not rely on models' intrinsic knowledge for any specific topic anyway, relying instead on web search/grounding/RAG/Agentic RAG/ReAct, that doesn't seem like a huge sacrifice to me. Of course the problem is that this experiment assumes that, should web crawlers start respecting these opt outs, nearly everyone wouldn't put them up, just really damaging model output. I think the better answer to the problem of bots swamping smaller websites is to have a common nonprofit foundation that collects training data for all AI companies, which then clean and refine it their own ways, that way only one set of bots needs to do the collection. They could also make their bots more respectful of resources in other ways (like cloning, instead of crawling, git repos).