What remains of a human being who no longer thinks - but is thought for? And how can they regain agency - when that very agency requires their own thinking?

— Steven Broschart

The Silent Shift

After the invention of the printing press and later the internet, searching via Google offered the fastest way to find information in a timely manner. And if clicking on a search result wasn't enough, you could buy a specialized book on the desired topic and work through documentation. Knowledge often meant:

  • searching,
  • comparing,
  • abstracting,
  • and building your own mental models.

Today, something different happens increasingly often: people ask language models like ChatGPT, Claude, or Gemini. Not necessarily because search engines suddenly became bad. Rather, because dialogue-based AI systems enable a new, context-specific, more intuitive, cognitively lighter form of knowledge access. A glance at evolutionary history shows repeatedly: the more efficient system prevails.

Initially, this looks like a purely technical advancement of existing search systems. In fact, however, there is much more to it. Because large language models don't just change access to information. They are beginning to alter the structure of human knowledge processing itself.

Why Specialized Books are Under Pressure

The same dynamic affects specialized literature. Historically, specialized books served multiple functions:

  • storing knowledge,
  • structuring knowledge,
  • conveying knowledge,
  • and 'forcing' people to apply abstract concepts to concrete situations.

This last point is particularly decisive. Because whoever reads a specialized book must:

  • abstract,
  • prioritize,
  • derive mental models,
  • and independently apply conveyed concepts to their own reality.

This is cognitively demanding, but it is precisely through this that deep understanding emerges.

LLMs fundamentally change the process of information gathering. Because a language model can not only explain knowledge but also adapt it situationally:

  • to the level of knowledge,
  • to the industry,
  • to the specific use case,
  • to follow-up questions,
  • and even to individual thinking errors.

This shifts knowledge transfer from explicit generalization toward dialogue-based operationalization. Or more simply put: books force people to abstract. LLMs relieve them of some of this abstraction work.

This could lead in the long term to declining demand for classical specialized books, at least where books primarily serve knowledge transfer. After all, why should someone read 400 pages on project management when an AI system knows the model, understands the situation, and can immediately generate concrete action proposals?

The Thinking Prosthetic

The externalization of thinking processes initially resembles earlier technological shifts, such as

  • calculators externalizing mental arithmetic,
  • GPS and navigation devices externalizing orientation,
  • search engines externalizing knowledge storage.

When LLMs increasingly take over structuring, transfer work, prioritization, and situative knowledge application, then humans increasingly need to build, understand, and store knowledge long-term less. Because they can retrieve knowledge anytime in dialogue with the machine for a specific situation and process it implicitly, that is, as a coordinator. We are speaking here of continuous division of thinking labor between human and machine.

What is problematic against this background is that this development occurs entirely voluntarily. Because cognitive relief and efficiency gains are perceived as tangible advantages. On one hand, it is comfortable for humans when cognitively demanding things can be delegated. On the economic level, competitive pressure forces ever greater efficiency. The necessity for deeper understanding, for cognitive penetration, has little place in this argumentation. Rather, it appears to be an argument from a bygone era.

The Synchronization of Thinking Processes

The actual disruption, however, lies one level deeper. Because LLMs don't just provide information. They structure thinking. Every response from a language model already contains

  • prioritizations,
  • weightings,
  • simplifications,
  • frames,
  • linguistic smoothing,
  • and implicit assumptions about what is relevant, plausible, or 'reasonable'.

This usually doesn't happen aggressively or obviously manipulatively. On the contrary: the systems always appear friendly, helpful, cooperative, and balanced. And often with such confidence, as if they were a specialist in the requested field. It is precisely this that creates their psychological pull. Because people don't go on the defensive when information seems plausible and helpful, when the communication style appears empathetic, linguistically smooth, and socially compatible.

The actual risk, therefore, is not direct manipulation, but rather long-term thinking convergence.

If millions of people:

  • use similar models,
  • similar training data,
  • similar response patterns,
  • and similar prioritizations,

then thinking processes can slowly align. Not authoritatively. Not visibly. But simply, conveniently, and comfortably. The result would not be classical synchronization in the historical sense. Rather, a statistical homogenization of language and, with it, of thinking paths. Especially because LLMs are optimized for conflict reduction and the most probable synthesis, they tend toward smoothed explanations with moderate positions and the greatest consensus, embedded in standardized argument patterns.

Is that efficient? Perhaps. But at the same time also very flattening. Because original thoughts often arise precisely at the boundary of conflicts, uncertainties, and contradictions. They arise when people must develop their own thinking models. When they cognitively penetrate the subject matter.

The Actual Transformation

Perhaps the most important change of the AI era is not that machines become more intelligent. Rather, that people are increasingly rarely forced to structure knowledge themselves. Because LLMs increasingly come between humans and knowledge. And precisely in doing so, they simultaneously change:

  • learning,
  • search,
  • marketing,
  • websites,
  • authorship,
  • media,
  • probably human cognition itself,
  • and thus also decision-making.

Conclusion: What Remains When Thinking is Delegated?

On YouTube, there are numerous tutorials on how to write entire books with the help of AI. And Amazon registered an unprecedented flood of e-books with the emergence of LLMs. The barriers have never been lower for producing so much poor content in such a short time. With borrowed, externalized thinking. About content that is organized by humans but not cognitively penetrated by the 'author'.

We are currently in a massive phase of transformation where we are experimenting with our interaction with artificial intelligence. It will probably take a while longer before we understand what position humans, with their time, experience, and expertise, should not leave or should reclaim in this process. Yes, they will increasingly become less a supplier of information, that's what machines take over. But they contribute through their experiences to clear positioning, perspective, and stance. Something that AI cannot provide 'by design'. Humans must resist convenient smoothing and tolerate contradictions instead of moderating them away. Not as a matter of principle rebellion against machine thinking. Rather, as a sensible and purposeful merger of efficiency and quality.

Perhaps this is the real challenge of the coming years: not to externalize every thinking task just because we can. Rather, to consciously decide which cognitive processes we delegate and which we should afford to keep. Because what we no longer think ourselves, we will eventually forget how to think.