Limits, Ethics & What Comes Next

A model detects an unusual temperature pattern on industrial equipment. The correct workflow is not “AI says danger, therefore shut everything down” or “the dashboard is often noisy, therefore ignore it.” The alert shows source readings and confidence; an operator follows the site alarm procedure; a technician checks instrument health and the physical asset; the decision and evidence are logged. Later, confirmed and false alarms are fed back into monitoring. AI helps prioritise attention while trained people retain safety authority.

A credit model estimates repayment risk from permitted financial information. Before use, the lender tests performance across relevant customer groups and documents which data is excluded. An officer reviews ambiguous cases and records reasons for overrides. Customers receive an understandable explanation and a route to correct data or appeal. The model supports consistency, but the institution remains accountable for lawful and fair lending.

Use a three-level check. For low-impact drafts, read and edit before use. For factual work, compare important claims with trustworthy independent sources. For decisions affecting rights, money, health, education, employment, or safety, involve the qualified person and follow policy before acting. If you cannot perform the required check, do not use the output as a decision.

A chatbot may give a neat paragraph with a date, quotation, and website title that do not exist. It is producing a likely sequence of words, not opening a truth meter. For assignments, medicine, money, law, or safety, trace each important claim back to a textbook, official source, or qualified person.

A tool trained mostly on examples from wealthy countries may misunderstand local names, accents, crops, prices, or school systems. That does not automatically make every answer useless, but it is a reason to test it with people and examples from the community it will serve. Asking “who is missing from these examples?” is a practical fairness habit.

A widely reported case involved a company that trained a CV-screening model on ten years of its own hiring decisions. Because past hires were mostly men, the model learned to downgrade CVs containing signals associated with women — even the word 'women's' in a club name. The company had not asked the model to discriminate; the model simply reproduced the pattern in its training data. The project was scrapped. The lasting lesson is that historical data records historical behaviour, including unfair behaviour, and a model trained on it will repeat that behaviour efficiently and at scale unless people deliberately test for it and intervene.

If you submit an assignment, publish a post, send a report, or make a decision using AI help, you own the result. This is the same standard we already apply everywhere else: a calculator does not excuse a wrong bank balance, and a spell-checker does not excuse a defamatory sentence. The practical consequences are simple — never submit or act on output you have not read and understood, disclose meaningful AI help where honesty requires it, and refuse to use AI for a task when you cannot check its work. People and organisations who follow these rules get the benefits of AI without inheriting its worst failures.

High-stakes choices — health, money, safety, grades — need a person who owns the outcome. AI can assist; it should not be the final authority alone.