A four-part series examining why general-purpose language models used as therapeutic tools lack the engineering rigour applied to every other safety-critical domain. Grounded in 150+ peer-reviewed sources.

Continuum Pulse has published the first article in “Building for Harm’s Way,” a four-part series on the engineering gap in therapeutic AI systems.

The series draws on 150+ peer-reviewed sources and regulatory documents to examine a specific problem: general-purpose language models are being used as therapeutic tools by vulnerable populations, but the systems they depend on were not built with the safety infrastructure that other high-stakes domains require. Per-user data isolation, language safety filtering, session persistence, observability, and failure recovery are standard in medical devices and aviation software. They are largely absent in the AI systems people turn to during mental health crises.

The work is informed by Continuum Pulse’s consulting work with Novii, a therapeutic AI platform. That project applies the architectural principles the series describes: building safety into the system layer rather than relying on prompt-level guardrails.

Article 1, “Therapeutic AI Is a Safety-Critical System. The Standards Exist. They Are Not Being Applied.,” is available now. The remaining three articles will follow through June 2026.