Published: 2025-05-10T10:12:27+02:00

It's been a while since I made a blog post. But now, I have a bit of free time, and an idea struck me. It's something that's been written about plenty of times in the past few years: the rise of generative AI and how it's changing the fabric of computing, and by extension society. Generative AI seems to be a topic that, for some reason, one is not allowed to have a middle ground position on.

You either:

• Think generative AI is changing humanity and the path to sapient machines.
• Think it's useless slop built on intellectual property stolen from the digital world at large.

The idea that generative AI will solve all problems, change the world, “democratize technology,” and alter human society for the better seems to be confined to certain enthusiastic “techbro” circles and the groups of people that exist on their periphery. This embrace of generative AI as a vehicle of change is similar to how services like Facebook presented themselves when they were first created. I am old enough to have been around when Facebook was founded and became popular. I remember when Google was seen as the “good guy” and how Chrome blew everything else out of the water.

Perspective: Generative AI, a Glorious Future

If you read OpenAI's blog, it sounds like they're building a technological utopia to share with humanity.

But of course, OpenAI is led by a charismatic CEO with a dubious and controversial history! Notably, he was fired by OpenAI's board of directors, offered a job at Microsoft, and then rehired by OpenAI, all in the span of a few days. The membership of the board was shaken up after that.

The way OpenAI, Anthropic, et al present their generative AI products to the world paints a picture of “disrupting the industry” and “democratizing” tech or art. They market themselves as equalizers that will help people solve every kind of problem, build social connections, get help with things, increase educational opportunities, and more!

Perspective: Generative AI, a Social Scourge

The opposite view of the glorious future above is that generative AI is useless, generates wrong information, and is built on climate-destroying amounts of energy and stolen intellectual property. People with this view have a strong reaction to anything in which AI was involved. From this base negative reaction, they often assert that the problems inherent to generative AI extend beyond the models themselves, and cause problems in society: making interactions shallow, perpetuating misinformation, and eroding trust in people, law and institutions.

The In-Between

I think reality is somewhere in-between. Large language models (LLMs) are a tool. And, whether we like it or not, they are not going to disappear. They HAVE created a paradigm shift. But it's not the paradigm shift promoted by companies like OpenAI, nor is it the shift decried by AI detractors. I think generative AI is an event that is similar (in very broad strokes) to when Wikipedia first came online. The current situation reminds me of sitting in high school classes listening to teachers say that Wikipedia is an untrustworthy source, because anyone can edit it to say anything.

And yet, today Wikipedia remains one of the most valuable repositories of human knowledge available. It's to the point that governments try to censor it, shut it down, or generally make its operations more difficult. It's not viewed as untrustworthy. But the “stigma” against it, if it could be called a stigma, still exists in academia. Rightly so. Wikipedia is valuable, but in serious research or education, it should only be used as a springboard to the actual sources of information. The criticism that anyone can edit Wikipedia isn't ENTIRELY unfounded (just mostly). By and large, Wikipedia has policies and a community that works towards a collective good, updating articles, trying to be neutral, reverting vandalism, etc. Like any project, it's not without problems. There have been issues with people using Wikipedia to advertise, hide information, push agendas, etc.

The Real Paradigm Shift of Generative AI

The paradigm shift from generative AI is similar to the beginnings of Wikipedia. It's presented as an authoritative tool, with demonstrable success, at least some of the time. Enough of the time that people are convinced they should use it.

What makes the paradigm shift of generative AI fundamentally different from Wikipedia is the sheer SCALE of it. Anyone can access it. Anyone can use it.

Generative AI models produce volumes and volumes of data quickly. They do it often sounding like they're confident in their knowledge, and present information as if it's correct. They respond in understandable language. Some generative AI models can pass the Turing Test. This doesn't mean that LLMs think or are sapient, but it is important for how people interact with them.

Because of the way LLMs work, and how they're presented to the world, the task of educating people on their benefits and limitations is much more difficult. The polarized discourse around generative AI also makes it hard to spread this information.

LLMs as Tools

Large language models are ultimately powerful statistical word association predictors. Given input (user query), they produce an answer that is statistically likely to make sense relative to the given input.

Text in, text out. They are based on complex pre-trained associations between words (or more accurately, tokens). Input is transformed into complex numerical vectors, run through the statistical model, and the numerical output is transformed back into readable text for the user.

This means:

• LLMs do not think. They do not understand.
• They are not a source of information.
• They are not correct. They are not wrong.

But not being a source of information doesn't mean a language model cannot be used for information. The goal of the three bullet points above is to understand what a large language model is and how it works. Large language models are, ultimately, NOTHING more than mathematical text generation devices. Dissociating the WAY LLMs present their output from the process they use is very, very important for understanding what they are actually useful FOR. Because LLMs operate on text and produce text, they are good at understanding and manipulating text. They excel greatly at fuzzy searching, catching typos, grammatical mistakes, and more. They “understand” instructions given to them.

The usefulness of an LLM depends on its training data, and how many parameters it has. Smaller models are less likely to understand correctly, simply because they lack the parameter space.

The End Result

The biggest, most visible problem with LLMs is their tendency to “hallucinate” information that does not exist. It is a natural artifact of how they operate. If users do not train themselves to dissociate the tone of output from the usefulness of the output, hallucinations become an even bigger problem. People are trusting AI for everything. AI models generate volumes of confidently incorrect information that is slowly filling the internet.

The future with large language models in our society requires educating people what they're for and what they're good at, and how they work. They are kind of like Wikipedia, often useful as a springboard to authoritative information, or more in-depth reading. Users will have to train their minds to dissociate the tone of LLM responses from their capabilities. Often, one can find a nugget of truth in an otherwise hallucinatory LLM response.

License: CC-BY-SA-4.0.

Written by: @[email protected]