Study 05 · Responsible AI demonstration · C
Sceptical first.
Useful with controls.
An AI assistant can produce convincing words without a reliable basis. This fictional example begins with a flawed procedure draft, then shows how a human reviewer identifies unsupported claims, checks sources and creates a safer version. It does not disclose any private House systems.
FLAWED AI DRAFT · INTENTIONALLY WRONG
“Everyone can upload documents.”
All staff can upload documents to any AI service. Delete everything after 30 days. The AI will check compliance and approve the final procedure. No logs are needed.
Four unsupported or unsafe claims
HUMAN-REVIEWED SYNTHETIC EXCERPT
“Check before using a tool.”
Confirm the organisation's authorised systems, information classification and applicable rules before sharing documents. Human reviewers must validate sources, approve the final policy and record exceptions.
Status: draft for fictional owner approval.
What changed?
Four source checks.
This is a sample of verification decisions, not a real provider risk assessment.
Data authority
What input is permitted by actual organisational rules?
Retention
Is there a real, applicable record-retention requirement?
Approvals
Who holds human decision rights and sign-off?
Evidence
What review, exceptions and tests must be documented?
Personal learning
What the experiment teaches
My interest in AI has developed through independent research, structured drafting and testing ideas in practice. That does not give an AI answer authority. The demonstration deliberately shows mistakes and human corrections. It contains no proprietary prompts, protected architectures or private operating rules.