AI-Assisted Cognition Risks Intellectual Stagnation Through Model Biases

Mia Heidenstedt's article examines how AI-assisted cognition threatens human development by skewing thought patterns toward outdated models. The piece introduces the concept of the Dynamic Dialectic Substrate, describing it as the foundation for human knowledge evolution through merging concepts. Large language models like GPT-5.3-codex, Gemini 3 Pro, and GLM-5 exhibit persistent biases from their base training, even after post-training on newer data. This creates cognitive inbreeding where population-level AI use reduces diversity of ideas and solutions. The article cites research from 2025 including studies by Hier et al. and Jiang et al. showing post-training preserves pre-existing representations. A hypothetical 2026 scenario of US invasion plans for Greenland illustrates how AI models might dismiss new geopolitical realities as impossible. Heidenstedt proposes mitigation strategies including human discussion, using diverse AI models, and exploring different AI personas. The author references academic sources like Frontiers in Psychology, MDPI, MIT Press Computational Linguistics, and arXiv publications. Statistics show AI usage increased from 40% to 61% between 2024 and 2025, with weekly use nearly doubling from 18% to 34%. The article concludes that while risks exist, the full impact remains uncertain without more outcome-focused research.