AI Ethics: The Real Problems, Not the Science Fiction Ones

Most arguments about AI ethics concern robots that do not exist. The issues that matter today are duller and closer: where the training data came from, whose writing was fed into it without being asked, what happens to your own messages, and who carries the blame when a model gives harmful advice. None of these has a clean technical fix, because they are questions about power and responsibility.

Where the training data came from

Language models are trained on text that was mostly collected without asking. Web crawls, novels, code from repositories, forum threads, news archives and in some cases material scraped from behind paywalls. Authors and programmers have sued, and courts in several countries are still working through whether that collection is permitted, fair use, or infringement.

The consent problem has no tidy solution, because the training already happened and a model cannot reliably forget one specific book on request. Retraining from scratch is expensive, and opt-out schemes arrived after the fact, often requiring creators to fill in forms to exclude work that was taken without notice. Whatever you think of the law, the default was take first and ask later.

Bias in, bias out

Bias is not a mysterious property that appears from nowhere. Models learn from text written by people, and that text carries historical patterns about who is described as a doctor and who as a nurse, which names appear in crime reports, and which accents are treated as trustworthy in writing. The model averages those patterns and reproduces them at scale.

Documented failures are easy to find. Hiring systems trained on past decisions that favoured one group. Image generators that returned mostly white men for the prompt a photo of a chief executive, then overcorrected when patched. Facial recognition with much higher error rates on darker skin. Speech recognition that handles some accents far worse than others. The lesson is not that the technology is uniquely prejudiced. It inherits data, and data records a world that already has preferences.

Confident wrong answers

Hallucination is an ethics problem as much as a quality problem. A system that invents a legal citation, a drug interaction or a study finding, and states it without hesitation, can cause real harm to someone who trusts it. The ethical failure lies in how the product is sold as much as in the model, because confidence and accuracy are separate things.

Any deployment where the stakes are high, such as health, law, immigration advice or mental health support, needs a named human who checks output and a visible warning about what the tool cannot do. Where those are missing, the tool is being used to move responsibility away from the provider and onto whoever happened to trust it.

What happens to what you type

When you type into a chat assistant, that text leaves your machine. Whether it is stored, for how long, whether it trains future models, whether a human reviews it, and which country it is processed in all depend on the provider and the settings. Many workplace tools disable training on submitted data by default. Consumer accounts often do not.

Some categories should not go into a general assistant at all: medical records with names attached, other people's personal data, unreleased financial results, source code under contract, anything covered by confidentiality. If a task requires those, the answer is an approved internal deployment with a data agreement, not careful wording.

Who is accountable

Accountability is the hardest problem here. If a model drafts a rejection letter that breaks discrimination law, blame could sit with the recruiter who sent it, the company that bought the tool, or the vendor who built it. In practice liability usually stops at the person who pressed send, while the vendor's terms of service disclaim almost everything.

A workable rule is that any consequential decision must be checked and owned by a named person who can explain the reasoning in plain language. Where that is impossible, the decision should not be automated. Disclosure matters too, and a growing number of jurisdictions now require it, so people know when they are dealing with software rather than a person.

Everyday choices that matter

None of the above needs new law to change your own behaviour today.

• Assume everything you send may be stored - Treat a chat assistant like an email account that somebody else operates.
• Keep a human owner for consequential decisions - Health, legal, financial and hiring outcomes need a named person who can explain the call.
• Check names, numbers and citations - Fluency is not evidence, and an invented citation costs more than the time saved.
• Watch for skewed output - If a tool gives you one kind of person every time, that is the training data showing through.
• Notice the consent question - Prefer tools that let creators and users opt out, and say so aloud when choosing.
• Say when you used a tool - Disclosure costs almost nothing and prevents most accusations later.