Language Models Embody and Amplify Human Cognitive Distortions: What Is to Be Done?

A set of recent arXiv papers (July 24, 2026) highlighting systemic risks and governance gaps around large language models (LLMs) and AI deployment. Key findings: LLMs embody and often amplify human sociocognitive biases across generations and can transmit bias to people; quantization (model compression) can increase open‑ended stereotyping even when standard safety checks pass (QuantiBias); internal model evidence and white‑box access improve some audit workflows but do not by themselves ensure valid, grounded audit reports; third‑party and regional AI auditing lags far behind deployment inthe