Induced from Bharanitharan Rajendran’s human-risk lecture at the KENX AI in GxP East 2026 conference (2026-09-15-1024-gorani-ai-governance-human-risk-gxp), 2026-09-15.
Five properties of defensible AI control
A documented intended use that states what the AI cannot do; risk-proportionate validation run as a continuous lifecycle because models drift; ALCOA+ audit trails extended to capture user, model, and version, inputs, outputs, justification, and confidence score; enforced approval by a named accountable human; and continuous lifecycle monitoring. The intended-use property is The intended use statement is the master control lever for GxP AI; the continuous-lifecycle property is Validation is data, not an event; the named-human property is Human-in-the-loop must be a named hard gate, not rubber-stamp oversight. The memory property needs Autonomous agents need their own auditable identities in regulated systems, or the trail has no attributable actor. If an unauthorized change alters a validated result, that is also A security failure that changes a validated outcome is a quality failure.
Why the phrasing lands
“AI acts, humans decide, and the system remembers everything” separates capability from authority. He reframes human-in-the-loop as human-centric design because nobody genuinely reviews a hundred AI-generated documents; the control has to make the review real, and the memory has to make it provable. Human verification of AI output must be explicit because document creation no longer carries implicit verification is how the review becomes real; An LLM judge is a GxP system and needs its own qualification pack can triage but does not decide. AI is not a faster human. His three-step start: inventory every AI touching GxP including shadow AI, enforce the human gate, treat validation as continuous. Human error, not AI failure, drives GxP warning letters is the same speaker’s citation record, and Assuming humans will follow SOPs is hope, not control is what you have if the gate is only hoped for.
Related
- Governance Risk and Human Factors MOC
- Human verification of AI output must be explicit because document creation no longer carries implicit verification
- An LLM judge is a GxP system and needs its own qualification pack
- The intended use statement is the master control lever for GxP AI
- Validation is data, not an event
- Human-in-the-loop must be a named hard gate, not rubber-stamp oversight
- Human error, not AI failure, drives GxP warning letters
- Assuming humans will follow SOPs is hope, not control
- Autonomous agents need their own auditable identities in regulated systems
- A security failure that changes a validated outcome is a quality failure