Chapter 14. What Is Needed Today—and What Is Not

14.1 What people and organisations need

AI literacy: understand basic capabilities, limits, uncertainty, and the difference between fluent output and verified evidence.
Clear purpose: define the problem before selecting a tool; do not adopt AI because it is fashionable.
Data discipline: classify information, minimise sharing, and use approved tools for confidential material.
Verification: confirm claims, citations, calculations, and decisions at a level appropriate to their consequences.
Human accountability: name the person who owns the decision and can intervene.
Security controls: least privilege, strong authentication, updates, logging, backups, and incident response.
Fairness and accessibility: test relevant groups and languages, provide accommodations, and offer appeals.
Transparency: disclose material AI use where people reasonably need to know or rules require it.
Training and change management: prepare staff, students, and affected communities before deployment.
Monitoring and review: measure outcomes, incidents, drift, complaints, costs, and environmental impact.
Exit strategy: know how to export data, change providers, stop the system, and continue essential operations.
Independent scrutiny: obtain specialist review when the stakes or complexity exceed internal competence.

14.2 What is not needed

Uncritical adoption simply to appear innovative or modern.
Claims that AI is always accurate, objective, conscious, or superior to professional judgement.
Uploading every document or personal detail into a public tool.
Using the largest or most expensive model when a simpler method meets the need.
Automating a broken process before understanding why it is broken.
Replacing human contact in sensitive services solely to reduce cost.
Treating an AI detector as definitive proof of authorship or misconduct.
Allowing AI to make consequential decisions without documented review and appeal.
Accepting vendor promises without evidence, testing, contract terms, and exit rights.
Assuming compliance, a watermark, or a policy document eliminates all risk.
Chasing every new product before evaluating whether it solves a real problem.
Confusing a generated answer with a source, a forecast with a fact, or a score with a full human judgement.

14.3 A simple decision filter

Before using AI, ask five questions: (1) Is there a clear, legitimate purpose? (2) Is AI better than a simpler alternative? (3) What data and people could be exposed? (4) What is the worst credible failure and how would we detect it? (5) Who is accountable and how can the result be challenged? If these questions cannot be answered, pause or run a limited, low-risk pilot rather than deploying broadly.