From the standpoint of generative AI ethics, hyperscale generative AI can’t be justified: it is trained on data used without its creators’ consent, causes massive environmental damage and depends entirely on the commercial interests of a handful of corporations. Small, domain-specific models with transparent training data, by contrast, deliver steady value without creating these problems.
Key Takeaways
- Generative AI outsourced to US hyperscalers is built on training data whose creators, in most cases, never agreed to have their work used for model training.
- The projected rise in US electricity consumption by AI data centers, from 4.4 percent to 25 percent of total demand, roughly equals all the renewable energy capacity added over the same period.
- Statistical language models have no structural way to tell true from false, so hallucinations are not an implementation bug but a basic property of this type of model.
- Students who hand learning tasks over to generative AI don’t build the skills those tasks are meant to develop, and the first graduating classes who studied entirely with AI support already exist.
- Small, domain-specific models with transparent training data, for example for language analysis or protein structure prediction, deliver steady value but receive a fraction of the investment that goes into large commercial models.
Generative AI Ethics: The Problem Is Hyperscale Models, Not AI in General
The debate about generative AI ethics targets one specific kind of artificial intelligence, not the whole field. The criticism is aimed at generative AI outsourced to a few large providers, most of them in the US. Any sensible assessment has to start from that distinction.
The umbrella term “AI” hides it. Anyone who talks about “AI” in general can point to all sorts of good applications and change the subject. Yet many successful AI applications aren’t generative AI at all, or at least not the hyperscale kind.
Johannes Link considers today’s hyperscale generative AI, in its current form, ethically unusable. That’s a personal judgment tied to personal values, and that is exactly how ethics should work: as an individual decision about what you are willing to do and what you aren’t.
Why the Models Are Built on Taken Data
The training data comes from works whose creators never agreed to that use. Calling it theft is a moral statement here, not a legal one. Lawyers will eventually decide whether the licenses cover it, but by then it will be too late.
The problem is concrete, and it affects everyone who has ever published anything. Open source software, talks, articles: all of it ends up in training data, even though nobody ever considered that a model might be trained on it.
Data quality can’t fix this either, because the foundation is a statistical model. A statistical model has no concept of right or wrong, or of truth. Whatever comes out is still statistics.
“No matter how good the training data is, we will always get hallucinations, because that is the basis of this model: even where I have no parameter in the training data, I still get something out.”
(Johannes Link)
The Ecological Cost Swallows the Progress on Renewables
According to the projections cited, data centers’ share of total US electricity consumption will rise from 4.4 percent to 25 percent within four years. That is roughly the amount of renewable energy capacity added in the same period. In other words, the AI hype eats up the entire expansion.
Then there is water. Cooling these data centers is a hard problem and takes a lot of it. There is also electronic waste. Some of the raw materials for GPUs could be mined in the Global North, but they aren’t, so that countries there don’t poison their own soil. Mining and disposal end up where the raw materials come from.
The hardware cycle makes things worse. In this kind of operation, a GPU lasts about one and a half to two years. The data centers practically always run at full load, because they can’t be scaled up and down in small steps, only switched on or off in large blocks. After about five years the entire data center infrastructure has to be replaced, something traditional data centers don’t experience in the same way.
Generative AI Could Seriously Damage Education
The education system is calling for AI because it is underfunded. Teachers are overloaded, and so are students. That’s exactly where the promise that a machine could fill the gap is most tempting.
The problem is delegated learning. Students hand off assignments that are supposed to help them build skills and knowledge. As a result, far less learning happens, if any happens at all. There are neurological studies on what this does to the brain’s ability to solve problems.
The effect arrives quickly. A bachelor’s degree takes three years, and the first students have already done their entire degree with these tools. In a very short time there will be graduates who lack skills they are supposed to have.
The same applies to anyone who generates texts instead of writing them. The point of writing is working through the topic, the friction with your own thoughts and your research. If you can generate a text, so can everyone else. Then nobody needs it anymore.
The Open Web Pays for Someone Else’s Load
Nonprofit projects such as Wikipedia and OpenStreetMap are under heavy load from AI crawlers, which scrape every piece of input they can reach. The infrastructure for that is paid for by the very initiatives that supply the crawlers with their material, and that money is missing elsewhere.
At the same time, the data pool is poisoning itself. An estimated 50 to 60 percent of all new content is already partly generated, and you can hardly create anything new from that material. AI companies are already responding by signing deals with publishers and newsrooms to secure genuine human input.
Why Filters Can’t Stop Hallucinations
Filters placed before or after a model fail for the same reason naive text filters fail against SQL injection. If you filter with regular expressions on the way in and the way out, you won’t get the problem under control.
The cause runs deeper. In large language models, the line between data and instructions blurs, and a prompt contains both at once. If you can’t separate the two, you can’t filter reliably.
On top of that, large models have a counterintuitive property. The more data and parameters they have, the smaller the share of targeted poisoning an attacker needs to push the model in a particular direction. People who want to manipulate models know this, and it happens.
What Would Have to Change to Make Generative AI Justifiable
The models would have to be under public control. Open, in this sense, means traceable: what goes into them, whether the creators consented, and verifiable by the public. That is the only way to rule out manipulation.
Right now almost everything depends on the goodwill of a few American billionaires. In the best case, they are only after profit. In the worst case, they follow misanthropic ideologies and won’t hesitate to manipulate the models. That dependency would have to end.
The use of resources would also have to be reasonable, whatever that means in a given case. The precondition is a first tangible, qualitative benefit. So far, convenience, less friction and profit dominate. Faster Java code doesn’t count as the kind of benefit that outweighs the damage.
A technology that helps with the climate, or helps the part of humanity that lives in poverty, would make sense. Such use cases barely show up in the current calculation, even if the technology is marketed that way.
Small, Open Models Are the Useful Part
Real value comes from very small models whose training data you can actually inspect. Small means a different scale than today’s norm: a few million parameters instead of a hundred billion. A mid-sized company can train models like that on its own.
In narrow domains, these models are not just useful but reliably useful. Examples:
- Language analysis, often limited to a single language and therefore much smaller
- Information extraction from texts, a problem that couldn’t be solved before these models existed
- AlphaFold for protein folding, a generative model, but small and trained only on its own domain
Most small models available today, however, are only open-weight and distilled from large commercial models. There is a lot that could be done in this area. The billions go elsewhere.
You Don’t Need to Learn Prompting as a Skill
The fear of falling behind if you don’t use AI rests on a fallacy. People who don’t use these tools aren’t losing any significant skills. The specific prompts change with every model generation anyway.
What prompting really requires is the ability to express things clearly. Most people have that ability to some degree, and it can be brushed up quickly once you need it.
That leaves two scenarios. Either software development isn’t replaced entirely, in which case you can learn the specific skills whenever you need them. Or software development is replaced entirely, in which case something completely different is on the agenda anyway.
A Collapse Is Likely, Its Scale Is Open
The financial bubble is big, and so is the interest of governments and very large companies in keeping it going. Some kind of collapse will come, if only because of the energy shortage. When it comes and how hard it hits is impossible to predict.
Changing the system as a whole takes politics. If you are politically active, you can add this issue to the many others and make fellow activists aware of it. There is a digital Luddite movement, in Germany called the Maschinenstürmer, and it’s more progressive than its image suggests. That’s one place to get involved.
For each individual, the first step is to check what you can personally stand behind. Energy consumption, chips and raw materials seem abstract, but in the end they affect everyone, much like the climate and the social upheaval tied to it. The concrete question is: what does this mean for me, and what can I do, or stop doing, in my own environment?
Frequently Asked Questions
Does the ethical criticism of generative AI apply to the entire field of artificial intelligence?
No. The criticism is directed at hyperscale generative AI, which is predominantly outsourced to a few large providers, mainly in the U.S. The umbrella term “AI” blurs this distinction: Anyone speaking generally about AI can cite any successful application as an example and thus divert attention. Many successful AI applications are neither generative nor hyperscaled.
Can hallucinations be eliminated with better training data?
No. A statistical model does not distinguish between right and wrong; it produces an output even when no suitable parameter exists in the training data. Hallucinations are therefore not an implementation error that can be fixed with cleaner data, but rather a fundamental characteristic of this type of model. Everything that comes out of it remains statistics.
To what extent do AI data centers hinder the energy transition?
Projections indicate that data centers’ share of U.S. electricity consumption will rise from 4.4 percent to 25 percent within four years. This roughly corresponds to the total expansion of renewable energy during that period. Added to this are cooling water and electronic waste: A GPU lasts about one and a half to two years under full load, and the entire infrastructure is replaced after about five years.
What do students lose when they delegate tasks to generative AI?
They fail to acquire precisely the skills that these assignments are designed to build. Academic assignments serve to build skills and knowledge; delegated learning causes them to diminish. Since a bachelor’s degree program lasts three years, there are already graduates who have completed their entire degree using these tools. Neurological studies have been conducted on the consequences for problem-solving ability.
Why aren’t upstream and downstream filters sufficient to secure language models?
They fail in the same way that naive text filters fail against SQL injection: Anyone who checks for it before and after using regular expressions cannot get the problem under control. In large language models, the distinction between data and instructions blurs; a prompt contains both at the same time. What cannot be separated cannot be reliably filtered.
What are the consequences of AI providers’ crawling for free projects on the web?
Nonprofit projects like Wikipedia and OpenStreetMap are under heavy strain from AI crawlers and are funding the very infrastructure that supplies the crawlers with their data. This money is then lacking elsewhere. At the same time, the data set is becoming tainted: an estimated 50 to 60 percent of all new content is already partially generated, which is why AI companies are entering into contracts with publishers and editorial teams.
How do small, domain-specific models compare to large commercial models?
In narrow domains, they are not only useful but consistently reliable, for example in language analysis, information extraction from texts, or with AlphaFold in protein structure prediction. “Small” here means a few million parameters instead of a hundred billion, which even a medium-sized company can train on its own. However, most of the small models available today are only open-weight and distilled from large commercial models.
Will you fall behind if you don’t use generative AI and don’t learn prompting?
No, that’s a fallacy. The specific prompts change with every model generation anyway. What’s needed above all is the ability to formulate things clearly, and that can be refreshed as needed. As long as software development continues to exist, these skills can be relearned at any time. If it’s completely replaced, something entirely different will be on the horizon anyway.
