Patterns / Safety & Error Prevention
When a change of consequence (deleting files or making breaking code changes) is about to happen, the AI system should seek confirmation first. The confirmation step should clearly communicate what is about to happen.
Model documentation should be accessible to everyone in your organisation, not just the technical team. It is important that everyone understands the capabilities and limits of the underlying product or system models.
AI products and systems should flag when an input is outside the AI’s training distribution or ability to reliably give a high quality output rather than fail in silence.
Design clear, accessible deactivation mechanisms for AI features in systems or products that are fast, simple and do not punish people for using them. The mechanism should be thought of as core features not emergency switches.
During development of AI products and Systems it is critical to be honest and transparent about what populations and contexts training data represents and where it falls short.
During development implement guardrails that reduce the likelihood of confident sounding false information. Examples of such guardrails are things like Retrieval-augmented generation (RAG), source grounding, linking to citations and forcing models to structured outputs that can be systematically checked.
When designing iterative AI workflows that require many versions of a document or asset, retain version history and prevent models to silently overwrite prior versions. This de-risks changes of consequence and vastly improves the user experience.