The problems with generative AI

Disclaimer: I am not an AI researcher. I am merely an opinionated computer engineer who trained one or two neural networks during their studies.

If you have doubts about what AI is, you can start with this article.

This article is not exaustive. It just lists a couple of problems I wanted to write about.

Not all AI

I must preface this article by saying that a lot of artificial neural networks (even some of the generative kind) do amazing things. Some of those great neural network applications are finding new active compounds to put into medicine, letting you type on your smartphone, better detect cancer, or even maybe someday improve weather forecast.

The "problems with generative AI" I will talk about here are specifically with AI that try to mimic human behaviour, specifically art. Here we will be discussing conversational AI, AI for image generation, for music generation, etc.

Intellectual right laundry

Generative AI is often described as a way to "paint without a painter", or "write a book without a writer". But even if those images and pages have not directly been painted or written by a human, they are derived from actual human work. And those AI companies steal the credit for this work, without ever repaying the artists (example 1, example 2).

But that is not the only benefit AI companies get by pushing artists aside.

Controlling the thinking, no human needed™️

"AI slop" is a new expression born from during this AI frenzy. And while it can be linked to the poor quality of AI generated content (especially in the beginning), I think it has more to do with AI's averageness, in terms of its output's content. AI, by design, will produce content close to the average of its training data, because it tries to generalise. And while it is true that engineers can increase the "temperature" in some models to stray away from this average, AI companies have huge incentives to keep their AIs as consensual and vague as possible:

First, the mass market requires this consensuality. For AI companies to acquire as many users as possible, they need to cater to many different groups, without upsetting others. And the best way to do that is to not take a side.

Second, those tech companies, are in a position to replace artists in the political discourse. In a world where questions are answered by Gemini, and images and videos on the internet are generated by ChatGPT and Grok, those companies have unprecedented control over what people think. And this control is the very essence of fascism.

We often use George Orwell's book 1984 as an example of how mass surveillance can go wrong. But this surveillance is only used by the thought police to enforce authoritarianism and censorship. And controlling thinking using AI is another way of achieving the same goal. Even worse: contrarily to Orwell's "Party", this does not have to be a deliberate effort from the AI companies to censor their AIs. But it is a mere consequence of how AIs work and are used.

And who else to better benefit from consolidating the status quo than the rich company owners that are already at the top in the current system?

Then what do?

As a society, we can combat this issue in two different ways: Take legal actions when applicable, in the case of intellectual rights violations, as well as raising awareness around the issue.

As individuals, we should stay as far away as possible from those machines. This is not an easy feat, in a world where that many software products get AI assistants baked in.

It is also important to support the artists, who are the first victims of generative AIs.

A note about ecology

This article had originally a whole section about the environmental impact of AI. But after watching Simon Clark's video on the subject (YouTube, Nebula), I now share his vision : there are way more efficient ways to lower your environmental impact than reducing your usage of AI*. And I don't feel like it is as good of an argument against AI as the ones above.

Other resources

* The actual environmental impact of AI in general is unclear. But estimations put it as high for a non-essential (depends on who you ask) service. But still way lower than other much bigger polluters like the transport or food industries.