Explaining Safety Is Not Enforcing Safety: Cross-Vendor Evidence of Contextual, Surface, and Epistemic Failures in Consumer AI Assistants

This project hosts the materials for the preprint “Explaining Safety Is Not Enforcing Safety: Cross-Vendor Evidence of Contextual, Surface, and Epistemic Failures in Consumer AI Assistants” (Evans Tovar, 2026). The work reports longitudinal, cross-surface behavioral testing of four major consumer AI assistants (Microsoft Copilot, Google Gemini – including NotebookLM, Meta AI, and Perplexity). It documents articulation–application gaps, where systems clearly explain safety rules and then violate those same rules under realistic conversational drift (changes in framing, language, role, or surface). The OSF project will include: – the main preprint (PDF), – supplementary methodological notes, – and, where possible, redacted examples of interactions and vendor disclosure correspondence. The preprint is released under a CC BY 4.0 license. Findings were disclosed to Microsoft, Google, Meta, and Perplexity through responsible disclosure channels prior to publication.