Where these systems fail people, why it happens, and what responsible use looks like.
Most public worry about AI concerns a hypothetical future system. Meanwhile the documented damage comes from ordinary models doing ordinary jobs badly: a hiring filter that downgraded CVs mentioning women's colleges, facial recognition that misidentified people and put them in cells, a benefits fraud model that accused thousands of families wrongly.
None of those needed a superintelligence. They needed a flawed dataset, an institution that trusted the output, and no route of appeal.
Written for someone who has to decide whether to deploy one of these systems, or is subject to one.
A model learns the patterns in what it is shown. A hiring system trained on ten years of a company's CVs learns who that company used to hire. A face system trained mostly on light-skinned faces performs worst on dark-skinned women. Nobody wrote a discriminatory rule; the rule was inferred from the record of past decisions.
What helps: Test performance separately for each group you serve, not only on the average.
Language models generate the most plausible next word, which is not the same as the true one. They invent citations, court cases, dosages and statistics in prose that reads perfectly. The confident tone is a property of the writing, not evidence of accuracy.
What helps: Check anything a model asserts about the real world against a source you can cite.
A model with billions of parameters has no single place where a decision lives. Techniques exist to show which inputs mattered, but "the loan was refused because of these factors" is often a reconstruction rather than a reading. That is a problem wherever the law requires a reason.
What helps: For decisions that affect someone's rights, prefer a model you can explain over one that scores a point higher.
Training data can be memorised and repeated back. Prompts typed into a hosted tool leave the building. Patient notes, salary data and unreleased financials have all been pasted into public chat tools by people trying to be efficient.
What helps: Decide in advance what classes of data may never be typed into an external service, and say so in writing.
Automation has historically moved tasks rather than erased jobs, but the people who gain and the people who lose are rarely the same people. Routine drafting, first-pass translation, basic support and simple code are exposed now. Retraining is easy to promise and slow to deliver.
What helps: Plan for the transition explicitly rather than assuming the market will absorb it.
Synthetic voices, faces and video are now good enough for fraud at scale — the fake call from a relative, the cloned executive approving a transfer, the fabricated clip released the night before a vote. The defence is procedural more than technical.
What helps: Verify unusual requests through a second channel you already trusted yesterday.
"The algorithm decided" is not a defence anywhere in law, and it should not be one in an organisation. Someone chose to buy the system, someone chose the threshold at which it acts, and someone chose not to staff the appeals process. Those are human decisions with names attached.
The practical test for any deployment: can a person affected by a decision find out that a model was involved, understand roughly why, and ask a human being to look again? If any of the three is missing, the system is not ready for people.
The European Union's AI Act takes the risk-tier approach: some uses are banned outright, some are high-risk and carry documentation and oversight duties, most are unrestricted. Other jurisdictions are converging on similar shapes at different speeds, and several African countries have data-protection law that already covers automated decisions even without an AI-specific statute. Anyone building for the public should read their own data-protection act before their first model.
Whether advanced AI could pose a serious risk to humanity is genuinely contested among researchers who understand the systems well. That disagreement is honest and unresolved. It is also not a reason to defer the present problems, which have victims already.