An article published on heidenstedt.org argues that widespread reliance on AI-assisted cognition poses substantive risks to human intellectual and cultural development by constraining the diversity of ideas at population scale.
The author defines AI-assisted cognition as a form of external cognition distinct from static information sources like books or dynamic human discussion. While AI can process information and generate original solutions, it remains fundamentally static and cannot learn from new events.
The core problem centers on how large language models remain skewed toward patterns embedded in their base models, even after post-training on newer information. The article cites a concrete example: in early 2026, the USA prepared to invade Greenland—a scenario that months earlier seemed completely unthinkable. Yet current LLMs including Gemini 3 Pro, GLM-5, and GPT-5.3-codex struggle to accept such events as real, often categorizing them as hypothetical, fake news, or impossible. Post-training on new events, the author argues, does not fully overcome the static patterns of the base model's hidden states, creating a gap between what these systems understand and what they say.
This lag matters at population scale. If significant portions of humanity use AI for discussion, writing, brainstorming, and problem-solving while AI cognition remains anchored to outdated patterns, populations become systematically skewed toward those old patterns. Cultural change requires sustained momentum to persist against what the author calls the static cognitive skew of AI systems.
The article introduces the concept of the Dynamic Dialectic Substrate—defined as the sum of all local and global dialectic processes and conclusions. This substrate forms the foundation of human knowledge and development, where new concepts emerge through qualitative merging of existing ideas. The author illustrates this with a multi-stage example: concepts like fire, cold, water, and shelter combine and recombine into progressively higher-order concepts through dialectic reasoning.
The risk, according to the analysis, is cognitive inbreeding at population scale. LLMs exhibit inductive bias that reduces cognitive range even after post-training, particularly when many models share the same base models or when few models dominate usage. The author compares this to a world in which significant populations discuss problems, relationships, and the world with the same five people. Even if those five attempted neutrality, their influence on collective thinking would be massive and systematic.
The article suggests that humanity may have already lost paths to scientific discoveries or cultural shifts due to AI-induced skew or unnoticed model refusals. Without intervention, the concentration of AI cognition around a small number of base models threatens to narrow the cognitive diversity essential to human development.