Claus Aranha @ University of Tsukuba

Preamble

In this post I'd like to clarify my position regarding Large Language Models (LLMs, or what many people call "AI" today): I want nothing to do with LLMs at the moment, be it in my personal usage, my work usage, or as a research topic.

In a classroom context, I consider LLM usage by students to be counter-productive, at best, and plagiarism at worse. In a wider sense, I think society is ill served by the use of this technology, and I think that the companies pushing its wide use are nocive institutions, emburdening society with heavy externalities purely for private gain.

Please take this into account when considering joining my laboratory, requesting reviews, proposing research topics, etc. Hopefully, this will invite you to rethink your own positions about LLMs.

For the rest of the post, I'll give a general outline of my reasons for the position above. This will be a short summary of the issues I see, but I'll not elaborate or link too much today. For each issue there are ample resources available with more information, and I invite the reader to look for them. In time, I hope to write more detailed and well sourced blog posts, and link them here.

Why Write about this now?

Because LLMs are everywhere, and I'm tired about having to explain my position again and again to people, or having to fume about it quietly. Now I have something I can point to.

Today, it feels like every online service is pushing the use of AI. Governments, Universities, Scientific Conferences are jumping on the bandwagon as well. Almost always the proposed use is unwarranted, often little more than advertisements for the companies behind the LLM boom.

On the other hand, recently I've been thinking about the social responsibility of computer scientists, and talking about it in my lectures. I want to lean into this, talking more publicly about how CS technologies affect our society, and what we should do about it.

Finally, someone recently told me they read and appreciated the previous posts on this blog, which motivated me to write more. Funny how that works.

The costs of LLMs

Here are the main issues that make me reticent about the usage of LLMs, roughly grouped by concepts (with lots of overlap). This is no way an exhaustive list, just a list of things that come up in my mind first.

Environmental Costs: "Privatize the revenues, socialize the costs". Companies benefiting from the LLM hype are externalizing these costs to society at large.

• Energy: The big one, extensively documented. LLMs require an unsustainable amount of energy, accelerating issues related to climate change.
• Water: In addition to energy costs, the cooling necessary for data centers uses water resources that directly hurt local communities.
• Mineral: Data centers also require specialized hardware that are a strain on rare earth resources. This has direct implications in global conflicts, and also raises the global costs of computer hardware.

Social Costs: More direct harm to people.

• Harmful Training: The large datasets used for LLM training require extensive human labeling. This is a often grueling work that is extracted from developing countries, without providing the necessary social support for the workers.
• Noise-shaped Signal: The increasing amount of LLM generated media (slop) makes it difficult to find reliable data, both on the internet and increasingly off as well. Separating truth from invention is becoming increasingly more difficult.
• Cognitive Dependency: "Check the results for accuracy", except that in practice no one does (otherwise they wouldn't be using LLMs in the first place). Increasingly, research is coming out about the cognitive ills of personal LLM use.
• Isolation: A point that I think often about, and I think might be underappreciated, is how isolating these tools are. People using LLMs ask less questions of specialists and peers, engage less with online communities. Every new use of this technology seems aimed at reducing the need for social interaction.

Research Costs: More niche concerns that I have about the use of LLMs in CS research.

• Opacity: LLMs are essentially blackboxes. In the majority of the cases, you don't know, or control, the exact data, algorithms or parameters. So you can't tell how these influence or are influenced by the software you build around the system or the experiment setup. If you change a word in your prompt, or if [Company X] changes the version of the model, how does that affect the result of the research? No one ever seem to know or care about these questions.
• Non-reproducibility: A direct collorary of the above, research based on LLM is essentially irreproducible. Irreproducible CS research is in the majority of cases useless.
• Burden on reviewers / literature: The large amount of LLM-slop in scientific papers is causing a burden on reviewers, and decreasing the signal-noise ratio in the scientific literature.

Economical Costs: Harmful ways that the use of LLM is changing how society is organized.

• Work Precarization / de-skilling: A lot of work replacement related with LLMs is focused on entry level placements and freelance work. It seems short-sighted to cut avenues for growth.

• Burden on small information providers: One issue that particularly interest me is how independent information providers (small websites, etc), are being constantly 'attacked' by misbehaved crawlers that gather data for LLM training. This discourage people from sharing information on the internet (in a compounding effect with the proliferation of LLM slop in the information landscape).

• Security and Technical Debt: The extensive use of code generating tools is causing new and exciting security issues in all sorts of software. The large amount of code that "no one wrote" and "no one can read" feels like a ticking time bomb.

• Centralization: Because of their high overhead, LLMs encourage centralization: A small number of huge corporations tasked with managing and offering large models to everyone. Maybe as a reflection of my scientific background, I tend to think this kind of social/economical organization to be brittle, and undesirable for many reasons.

Note that I didn't get to the discussion of whether LLMs can be useful or not. While I'm personally sceptic about their usefulness, any possible use of LLM should be first weighted againts the harm they cause.

This post will hopefully be updated with more information and links in the future.