AI – Risks and Drawbacks

Aug 31, 2026 - 5 min read

As a programmer, there’s hardly any topic that’s as prominent to me as AI.

I am aware that there are many meaningful and useful applications for AI, for example in science, education, and technology. I, too, use AI regularly.

Nevertheless, I have decided to draw a line: I do not use AI-generated code, and I do not publish AI-generated content. The reasons for this are listed below.

This is not a balanced, nuanced discussion of the pros and cons of AI. The following list is entirely one-sided, as it represents only my position. The list is intentionally brief, which means that some points are greatly simplified. It is not intended to convince you; it merely explains the reasoning behind my decision mentioned above.

A) As a programmer

AI leads to a loss of competence: Anyone who constantly uses coding agents will become less skilled at writing or debugging code on their own.
Programmers do not have as deep an understanding of AI-generated code as they do of code they have written themselves. As a result, they lack the deep expertise that comes from grappling with problems and unraveling complexity.
The use of coding agents compromises code quality.
- AI-generated code is often repetitive and unnecessarily complex—statistics show that, on average, it is 20% to 30% longer than equivalent human-written code.
- AI typically considers only individual files and functions without taking the overall architectural concept into account. It does not refactor code on its own initiative when it is actually necessary.
- AI-generated comments usually describe only what the code does, but almost never why.
The use of coding agents compromises the correctness and security of code: AI-generated code statistically contains more bugs and more security vulnerabilities.
In most cases, coding agents do not give sufficient consideration to non-functional requirements such as accessibility and performance.
During code reviews, errors in AI-generated code are often overlooked.
- AI can write code faster than we can review it. This creates time pressure.
- We’re not very good at finding errors in code written by others because we don’t understand it as well as code we’ve written ourselves.
- During code reviews, concentration tends to wane after just a short time, which leads to problems being overlooked.
- AI-generated code looks visually clean. Since its flawless appearance gives the impression of high quality, we subconsciously examine it less closely.
When refactoring someone else’s code, you can ask the original author about their goal and thought process. With AI-generated code, this thought process can no longer be traced later on, which makes maintenance more difficult.

B) In general

AI strips users of their autonomy. With autonomy, I mean the ability and determination to use one’s own intellect without guidance from others. Those who use AI without restraint eventually lose this ability. The intellect is like a muscle, it needs to be exercised.
AI lowers our tolerance for frustration. Genuine learning and problem-solving require the ability to endure friction, as it is the struggle with a problem that leads to profound insight. By using AI, we lose the capacity for “deep work”—a process of deeply engaging with a subject that demands sustained, distraction-free concentration but anchors knowledge firmly in long-term memory.
- Querying an AI creates an illusion of competence: one understands the answer superficially but doesn’t internalize the knowledge. Shortly after, the information is no longer retrievable.
Generative AI is characterized by mediocrity. Since it always produces the statistically most probable token (setting fine-tuning aside), it never generates anything rare or extraordinary; it possesses no genuine originality.
If we constantly settle for AI—that is, for the average—we will soon lose the incentive to create anything extraordinary.
AI reinforces biases because it consistently reproduces the majority opinion. It tends toward cultural homogenization, causing minority perspectives to increasingly disappear.
Model collapse (the use of AI-generated content to train newer models) drastically amplifies these effects, leading to the loss of nuanced information.
LLMs hallucinate. This is not a bug that can be fixed; it is a fundamental aspect of how they function. Consequently, they give us a false sense of security regarding the accuracy of their statements.
AI-generated works lack artistic value because they involve no physical, intellectual, or emotional labor.
- AI-generated works may look beautiful. However, just as printing money leads to a loss of its value, beauty loses its worth through the mass production of beautiful images.
The flood of AI-generated information contaminates our collective knowledge. Since any piece of content could potentially be AI-generated, we are forced to constantly question its authenticity. In particular, photographs and videos lose their former status as reliable evidence.
AI creates dependency. Those who outsource processes to proprietary models find themselves at the mercy of a handful of tech corporations whose pricing and terms of use can change arbitrarily at any time.
LLMs are trained on open-source data, often without regard for the associated licenses. This practice is ethically and legally questionable, as LLMs sometimes reproduce training data word for word.
The massive consumption of energy and fresh water required for inference and training can hardly be justified by increased efficiency if the process destroys the very foundation of our existence.

This page was machine-translated from German.