Safety: examine the harm
A system can follow a request and still cause harm. A wrong appointment reminder may inconvenience someone; a wrong instruction in a medical setting can have much more serious consequences. The stakes determine how much evidence, review and restriction the task needs.
Alignment: examine the objective
Here, alignment means checking whether the system’s behaviour fits the intended purpose and the constraints of the people affected. A support tool told to close tickets quickly might produce impressive numbers while leaving customers without answers. Define success as a resolved problem, then inspect the cases the metric overlooks.
This is a practical introduction to alignment, not a claim that a checklist solves the wider research problem of aligning advanced AI systems.
Five checks before release
- Purpose: describe the outcome and who benefits. Identify people who could lose out.
- Boundaries: restrict data access and actions to what the task needs. Treat instructions inside retrieved documents as untrusted content.
- Evidence: test ordinary cases, exceptions and plausible failures against approved sources.
- Authority: require appropriate approval before external promises, payments or other consequential actions.
- Recovery: name an owner, preserve a manual route and specify when the system must stop.
Listen to the people doing the work
Job anxiety is a legitimate organisational concern. Ask what work is changing, who will carry the review burden and what support people need. A preference game cannot tell you whose fear is “correct”, and a quiz score cannot establish someone’s employability.
Use the daily games for low-stakes practice and reflection. Their scores are not professional certification, psychological assessment or evidence that a system is safe to deploy.
Go deeper
NIST AI Risk Management Framework offers a structured approach to assessing and managing AI risks. Our pilot guide shows how to measure the human work around a system. The four Ds help make responsibility explicit.