Why I Resist Generative “AI” – Esther Weidauer

I don’t make a secret out of how much I despise the current hype around generative “AI”. I avoid using it as much has possible, and I’ve been vocal about the reasons for years. But often I feel the need to have a collection of all my arguments in one place instead of dozens of posts scattered around social media. This article is that collection. I’ll go through all the major reasons why I oppose generative “AI”, respond to some popular claims about it that I often hear from its proponents, and I’ll provide sources for further reading along the way.

The sections are mostly independent from each other so feel free to skip around the table of contents and read only the ones you’re interested in.

I’ll not get too deep into the technical details as those aren’t really the focus here and I’d rather look at more high-level issues to keep this from becoming way longer than it already is, and to keep it accessible for people who aren’t tech experts themselves.

I’ll also put down my own cards here: I have professionally worked in software development for about 15 years, and got about 25 years of programming experience across several languages in total. I have worked a lot on search engines and have implemented my own neural networks occasionally. Currently I’m finishing up a degree in psychology and writing my thesis in statistics. Just so you know where I’m coming from here.

It’s often not very clear what people mean when they say “AI”, especially when they are somewhat knowledgeable in that field and are aware of some of its more specific terms, so I want to put out some terminology before moving on. For the purpose of this article and I think in most colloquial use by laypeople, “AI” refers to generative systems that produce text, images, or other media based on a text prompt as well as systems derived from them, e.g. “AI agents”.

As of 2026, the time of writing this, the term artificial intelligence has been kicking around for over half a century. Various and very different technologies have been labeled as such. Systems that implemented formal logic were the big bet in the 1950s and 60s. Following the “AI winter” period after those efforts proved a dead end so called “expert systems” were the hype of the 1980s. Right now, we’re seeing a big revival of the artificial neural network, a technology that was thought to be of limited use decades ago, but that turns out to be quite powerful when combined with absurd amounts of data and computing capacity that were previously impossible, or unfeasible to bring to task.

Neural networks are quite remarkable at adapting to statistical patterns in complex data sets. A well-known entry-level example is a relatively simple network that can recognize hand-written numeric digits with surprising accuracy, when trained with sufficient amounts of training data (a large sample of written digits that are already correctly labeled). Neural networks have a wide array of applications and can model a variety of real-world problems.

Neural networks form the basis for what is now known as large language models (LLM), which is the core of the current “AI” hype, more specifically the generative pre-trained tranformer (that’s the GPT in ChatGPT). LLMs are statistical models of human language. They encode probabilities of words appearing in sequence and given an input sequence of words, called the prompt, they can infer (predict) what words are likely to appear next. A simple version of this would be the predictive typing on a modern smartphone where as you type text, a short list of likely word suggestions is shown to speed up the process. You type “appo” and it’s likely that you mean to type “appointment” so the phone gives you a shortcut to do just that. The LLMs powering popular chatbots like ChatGPT or Claude are super-charged version of that, not just inferring based on a few previous keystrokes but on quite extensive prompts. However, they still only predict a very small amount of text at a time before they need to repeat the entire process, feeding the fresh output back onto the system along with the previous prompt.

There’s a popular misconception that an LLM “remembers” your chat history and the prompt is always just the most recent message you sent, but that’s not the case. The prompt is the entire conversation. It is fed into the LLM with every new chat message. Having the chat history in the prompt helps keep the model “on topic”, otherwise these things easily drift to something else. this has it’s limits though. LLMs can only receive a prompt of a certain maximum length. This is known as the context window. Anything further back than the context window allows is no longer part of the prompt.

A few years into the development of commercial services based on LLMs, the latest and greatest right now are “AI agents”, systems that supposedly have, as the name suggests, agency, acting autonomously based on some initial task set by their user. Or at least that’s how the sales pitch goes. While a plain chat interface to an LLM might be able to respond to a user’s questions or perform things like summarizing text, an “agent” is supposed to execute more complex tasks like “book a trip to Berlin next weekend along with a hotel room for two adults”.

This sounds like a fundamentally new and different technology compared to just LLMs, but it isn’t. An agent is mostly just a program that repeatedly prompts an LLM and based on its output, executes various functions that are available to it, often called tools. Tools could be things like executing a web search, making a payment using a Paypal account, sending an email, etc. The “AI” part of it is still an LLM, mostly the same one you could also interact with directly via a chat interface.

Machine learning is a wider term that includes LLMs and other uses of neural networks, but also more generally any statistical model that can make predictions about something. Strictly speaking even a simple linear regression model, a very basic principle in statistics, can be seen as a form of machine learning as the regression model “learns” the correlation between variables and can then make predictions that fit those correlations.

Weather forecasts, image classification (the thing that finds pictures of your cat in your phone’s camera roll), street traffic modeling, power grid management, and many more applications fall under this very wide umbrella.

I believe it’s important to differentiate between the colloquially used “AI”, meaning generative systems, and other specific applications of machine learning. It’s generative “AI” that I’m going to focus on in this article and that I think needs to be resisted.

OK. So is heroin. And I wouldn’t recommend that to anyone either.

But seriously, this really isn’t an argument for or against anything. No technology fully vanishes again once it has been developed. That doesn’t mean though that it remains widely adopted, nor that it’s even a net good.

I don’t expect LLMs to disappear and I don’t think anyone who’s serious about this really believes that either. But the current boom is largely fueled by subsidized costs. End users don’t pay anywhere near the real cost of operating the big popular chatbots. The cost is covered by investor capital and that does run out eventually. And once the real cost will bear down on the consumer, a lot of people will reevaluate whether this is all really worth it. We already see large companies now restrict their employees’ use now that big LLM operators have switched to a “pay what you use” model.

Personally I think a lot of people who rely on LLMs a lot are in for a rough awakening once they have to pay the real price.

This is one of the favorite claims made by people who’d rather not think too much about how they participate in something destructive and there’s two problems with it:

An angle grinder is a tool. It’s use is entirely in your control, as long as you can provide it with electricity. but since it’s electricity requirements are well documented and not that complicated (220V AC at 50Hz for one I would get at my local hardware store), you’re not bound by anything in how you use your angle grinder. It’s also entirely agnostic whether you use it to build something or to cut a bike lock that’s not your own. And while it might break eventually, and its construction requires a complex supply chain, one it is made it’ll operate just fine for a long time.

The fire department is not a tool. This should be obvious. It’s clearly a service, provided by your municipality probably. It’s a complex web of materials, tools, processes, people, rules, and politics. It’s not something that gets built once and then just works. If you want someone to show up to your burning house and rescue you, a lot of things need to happen on a regular basis, many done by people.

For the continued operation of generative “AI” models, a complex web of materials and labor also needs to perform various tasks. It’s just less obvious because you don’t see the people making it work. You just see the input and output of the entire system. The labor and material realities are abstracted away, but that doesn’t mean they don’t exist. Moderators need to work in order to keep the LLM from spewing child abuse material at you, someone needs to replace the constantly burning out GPUs in the datacenters, and someone needs to curate the new training data to keep the model up to date, just to name a few. This is not a tool. This is a service.

Even if one is to accept the “It’s just a tool” claim, it doesn’t follow that using it widely (or at all) with little restrictions is fine. There are plenty of tools that we socially regulate, through laws or plain social conventions. If I hand you a hammer for driving some nails into a wall, I have a reasonable expectation that you won’t hit anyone in the head with it.

Tools also contain their own politics, sometimes explicitly, sometimes implicitly. Weapons for example are obviously designed to harm people, they very explicitly contain the politics of violence and war. A more implicit example could be the steam engine. It’s not just an invention that happened in a vacuum. Without the political and economical framework of the fossil fuel industry it can’t operate and would lose its entire reason for existing.

Pretending that a tool is somehow “neutral” is a very convenient tale to tell oneself, but it’s simply not true.

This one is closely related to the “It’s just a tool” claim and similar to that one, it needs to be evaluated in a wider context. Usefulness itself is not a sufficient criterion for whether something is good. Slavery and child labor are strictly speaking useful, but many societies have strict rules against those since they are also deeply exploitative. A steam locomotive is useful but we have mostly moved towards much more efficient and cleaner modes of propulsion for trains.

I don’t deny that LLMs are capable for certain use cases. A modern LLM can absolutely write a polite yet direct email to your landlord about how they should finally fix the central heating, or generate a boring but serviceable CV for a job application. LLMs can do much more than those very simple tasks even, but the more complex the task, the more likely there will be mistakes. And once you’re using it for something that you couldn’t do well yourself, you have no way of validating its output, rendering it very much not useful while still appearing as if it is.

Whether it’s climate change, space travel, curing cancer, or some other flashy promise, somehow “AI” is supposed to solve it. Yet, several years into the generative “AI” hype, there’s little to show for.

This is a claim where it’s really important to make sure we’re talking about the same thing when we say “AI”, as laid out earlier. Specialized machine learning systems can be essential to various fields of research, but LLMs are simply not capable of producing new knowledge. They can only reproduce what’s in their training data. An LLM-based system, being a statistical model of its training data and therefore effectively a form of lossy compression, can by definition never solve any problem for which the training data or information otherwise available to the system doesn’t already contain the solution.

Well yes, I am, quite proudly so I might add. But I don’t think the person who uses that term as an insult and to discredit actually knows what a Luddite is.

The original Luddites were a worker movement in the 19th century that was concerned about the economic implications of applying certain technologies in their work. They were not opposed to all new technology, as the popular misconception goes. People who are not aware of the actual history of the Luddites often think the name refers to someone who rejects all modern technology by default. This is simply not true.

Luddism is experiencing a revival today especially among people who are often technology experts themselves, who use modern technology every day, who work in research and development, but who also examine technology critically. For further reading I recommend The Mechanic And The Luddite by Jathan Sadowski.

Extraordinary claims require extraordinary evidence and this is a really extraordinary one. So far there is zero evidence towards this though. All we have are some people claiming that the text generator is somehow a real person now, usually people who worked at one of the major companies producing these systems and who benefit from painting their work as fantastical and revolutionary.

It is very easy for humans to anthropomorphize things that simulate human behavior, even when the simulation is crude. This is known as the ELIZA effect, named after a chatbot developed in 1966 that fooled several users into thinking it was intelligent.

Every first-year computer science student should know that computers that are Turing-complete can only ever do what a Turing machine can do, actually less because the theoretical Turing Machine has infinite memory which real computers obviously don’t. Making the computer bigger doesn’t magically let it surpass those restrictions. It only lets it do more of the same and do it faster.

We don’t even really know what consciousness is, other than gesturing at ourselves and going “You know, like that!”. What we do know is that the human mind is not a set of computer instructions. Neither is it a disembodied entity that’s just piloting a body. We are embodied beings and our bodies are an integral part of who we are.

Frankly to me the claim that an LLMs could even be anywhere near human consciousness is obscene and it reveals more about what the person making the claim thinks about people than it does about LLMs.

Operating a state-of-the-art LLM for a large user base requires an enormous amount of energy. According to a report from MIT Technology Review, global electricity use just for “AI” purposes could reach up to 326 TWh per year in 2028. That is more than the entire electricity consumption of the United Kingdom in 2025, not just data-centers in the UK, but everything from private homes to industry.

A lot of this electricity is coming from burning fossil fuels, often from gas generators specifically constructed to power new “AI” data-centers, leading to extremely high emissions.

I would argue that anyone who claims to care about preventing climate change getting worse than in already is, cannot ignore this ridiculous cost of LLMs. Especially when using them for tasks that could be accomplished without them, it’s like driving an SUV to a coffee shop that’s actually comfortably in walking distance.

The psychological damages caused by LLM use, sometimes even in the short time, are well documented by now, even by organisations operating or investing in these systems. In 2025 Microsoft themselves published a study about the cognitive impact.

Since then there have been more publications about the negative effects on personal relationships, about how LLM users consistently under-perform at neural, linguistic, and behavioral levels, and even deaths related to LLM use.

There’s even the concept of a “cognitive virus”, where “AI” use spreads among people who become increasingly dependent on it.

This is of course anecdotal, but I have witnessed these effects myself among students at university. Over the past two years some of them have become unable and/or unwilling to answer any question without first running it through their LLM of choice, even becoming unable to work on things like their theses when they ran out of tokens on their LLM usage plans. Their intellectual abilities have become entirely dependent on an online service that they have to pay for. It was quite horrifying to observe.

Ever since LLMs became really popular, probably around 2024, there has been a truly staggering amount of investment in the build-out of this technology. The money tends to go into buying GPUs from Nvidia, building data-centers, and the electricity to run both.

Talk of an “AI bubble” started popping up around the same time as the amount of capital poured into these systems sky-rocketed. In 2025 Deutsche Bank issued a warning about it and since then it’s become clear how the financing of the industry is partly circular, as pointed out by Michael Burry, one of the guys who correctly predicted the 2008/2009 financial crisis. Recently we saw the bankruptcy of character.ai, the first big company in this industry to topple over. It remains to be seen, if this is the first domino of many to fall but the current money-burning can’t go on forever. And when some of these bills come due, along with the fact that so much of the current economy is exposed to this, I think the shock will be severe, like 2008 all over again and worse.

Another economic aspect is the effect that LLM use has in the workplace. Shopify’s CEO Tobias Lütke who first went all-in on “AI”, forcing employees to use LLMs, now complains about “slop grenades”: generated output that hasn’t been properly reviewed and ends up creating more work for other employees. I admit that there’s some amount of schadenfreude in this one for personal reasons.

Epistemology: the philosophy of knowledge

LLMs necessarily produce mistakes. Don’t take it from me, take it from OpenAI. They fundamentally generate output that looks plausible, but isn’t necessarily correct. This is a major problem. When people use LLMs in fields that they don’t have expertise in, they lack the ability to verify the outputs. That in combination with the well-documented decline in cognitive abilities that the use of LLMs can incur, leads to an increasing risk of incorrect information that is trusted without validation.

We’re also at a point where LLMs actively destroy their own most important resource: genuine human-created training data, as OpenAI and Microsoft admit themselves. This leads to a feedback loop where future versions of these models are trained on information that was itself generated by an LLM, degrading the quality of the output even further.

I’m part of the generation that saw the internet become ubiquitous as we grew into adults. I know what it was like before, when researching anything meant spending hours and hours at a local library, waiting for weeks for a book to arrive only to then realize that it’s not the right one, and paying horrendous amounts for academic journals. The internet made so much of this so much better. I saw the greatest public collection of knowledge humanity ever had come into being, exemplified most of all by Wikipedia. When I was in my last years of school, we were all told by our teachers not to trust online sources, especially Wikipedia because “anyone can edit that”. Turns out that communal model of curating knowledge works remarkably well.

It’s ironic to see people in academia today who really should know better just trust LLM output that is proven to often contain mistakes without checking, copying hallucinated sources straight into their publications. Anything published after around 2024 that isn’t clearly labeled (and trustworthy) as being human-created will be tainted by this epistemological rot that necessarily happens when LLMs feed on their own output. If this is allowed to continue, the public collection of all human knowledge that we built over the first three decades of the Web, even with all it’s caveats of copyright and DRM schemes, is going to become a shadow of its former self.

We had the modern Library of Alexandria, and we’re burning it down again right now.

Since LLMs only reproduce information that’s contained in their training data, although sometimes in random and incorrect permutations, they necessarily reinforce existing biases and sentiments. This is strictly a consequence of how they work.

Due to their nature as statistical models, LLMs also bias towards the mainstream view on any topic. After all what makes some text more likely to appear in the output for a given prompt is how often similar text was in proximity of text resembling that prompt in the training data. However the mainstream view on something can often be wrong. Popular misconceptions are, well, popular.

Companies that operate LLMs, like OpenAI and Anthropic, apply content policies to their products. This might seem reasonable at first glance. After all, they probably don’t want a machine dispensing tutorials on how to make explosives at home, if not for safety reasons, at the very least because of liability concerns.

However, they also exclude topics that are perfectly legal and not at all morally objectionable to read or write about. Let me tell an anecdote about this:

During a psychoanalysis seminar during my psychology degree we were assigned a text by Rolf Pohl, “Vom Liebhaber zum Lustmörder” (“From Lover to Sex Killer”), that’s about sexually motivated aggressive behavior in men. A large section of the text is about sexual violence committed by soldiers against civilians during war. I was in a group of three students and one hadn’t actually read the text but only a ChatGPT summary. They had no idea about this important part of the text. The LLM had simply not included it in the summary. And of course since they hadn’t read the text themselves, they had no way of knowing about this omission.

In another seminar a similar thing happened where a discussion about the psychology of addiction and specifically alcoholism was left out of a summary, completely changing the meaning of the assigned reading and leaving the student who had only read the summary struggling to participate in the discussion.

Neither of these is a technical failure of the LLM itself. Even with the tendency to hallucinate content and produce factual errors, these mistakes are way more significant than the ones LLMs usually make. These are an effect of the policies put in place by the operators of the models. With more and more people relying on LLMs for information, this concentrates a worrying amount of power over our collective access to knowledge in the hands of a few individuals who have the means of training very big LLMs and who are not beholden to any democratic oversight.

If you ask me, this is downright dystopian.

The current strain of “AI” is a right wing project, even a fascist one.

I know this is probably a point where I lose a lot of readers. There’s a reason why I put it this far down in the article. Many people tune out when the word “fascism” is mentioned. But I’m serious about this.

Luckily other people have already done a much better job at writing about this than I ever could. So if you want to know why “AI” contains extreme right-wing tendencies, I’m going to direct you to them:

And if you’re up for a longer read:

I really do encourage you to read some of these though. The removal of human accountability, the disempowering of workers, the inherent conservatism, the centralization and control of knowledge, the fetishization of technology and “progress” for their own sake, all of those and more are facets of modern fascism and I believe it’s important to call them out as such.

Now that I probably already lost a bunch of readers during the part about fascism, I’ll close by going a little into just how weird the people are who run the current “AI” hype.

He showed me his bedroom, which was in the office; many Cluely staffers also lived there. Everything was gray, although there wasn’t much. “I’m a big believer in minimalism,” he said. “Actually, no, I’m not. Not at all. I just don’t really care about interior decoration.” He had a chest of drawers, entirely empty except for a lint roller, pens, and, in one corner, a pink vibrator. “It’s for girls, you know,” said Roy. “I used to use this one on my ex.”

This is from an astonishing piece of reporting by Sam Kriss about an “AI” company called Cluely, it’s CEO Roy, and the wider culture of “AI” boosters in the Bay Area. These are people who try to eliminate human thinking and who are at this point unable to do anything without piping it through an LLM first. They never had to grow up, so they didn’t, and now they need a robot to fill in for their mum.

Have you heard of Roko’s Basilisk? It’s a twisted version of another much older thought experiment called Pascal’s Wager which attempts to argue for following Christian doctrine because the risk of eternal damnation in case a vengeful God actually exists outweighs the benefit of an easier but sinful life. Roko’s Basilisk assumes a future all-powerful AI that retroactively punishes people who didn’t help to bring it into existence by creating simulations of them and torturing those virtual people for all eternity.

I believe this kind of pseudo philosophy says more about the people constructing these scenarios than anything else. Punishing non-believers is what they would do, if they were an all-powerful god. And since they see themselves as perfectly rational, a superhuman AI would obviously do the same. It’s ridiculous and, if these people didn’t control so much money, nobody would really entertain this nonsense.

Or what about Longtermism, a school of thought that argues that future potential beings, sometimes including artificial intelligences, will outnumber the current human population. Therefore the concerns of people alive today are irrelevant and all resources should be used to bring those vast numbers of future beings into existence, for example by colonizing other planets. That’s a very convenient argument to make when you’re e.g. running a rocket company and want to ignore things like climate protection regulation.

If you want know more about the messed up ideas, I recommend “The TESCREAL Bundle” by Timnit Gebru and Émile P. Torres. Longtermism is just the L in that acronym, there’s so much more nonsense in there and much of it is super popular in the current “AI” industry.

I want to leave you with another excellent piece of writing that examines the “AI” hype through a theological lens, as people who used to be atheists trying to create a new god in order lend divine authority to their ideas: Making God by Emily Gorcenski.

I want you to think about whether these are the people who should control our access to knowledge, determine who gets to have their voice heard in the future, and allocate the limited resources we have on our planet.

Personally I think these people are dangerous and need to be stopped.