Who should decide?

Laws draw the outer lines. Inside them, most of the choices are yours, and they don’t call for a lawyer. They call for four questions, asked before you hand a task to AI. These are the book’s rules of thumb, not a legal test:

• Who is affected? You alone, or a customer, a student, an employee, a patient?
• What would a mistake cost them? An awkward email, or a job, a grade, a diagnosis?
• Could you explain the decision to them without pointing at the tool?
• Would you tell them AI was involved? If not, ask yourself why not.

The answers sort most tasks into three groups, as Figure 14.1 shows. If only you are affected and a mistake is cheap and easy to spot, let AI draft and review what it gives you. If the task is a judgment about a person, AI may help you prepare, but a person decides and can say why. And if the material itself must not leave a protected place, such as a patient’s chart, keep it out of general AI tools altogether; Chapter 10 explains why.

Here is the judgment worked through, in an example written for this book. A hiring manager has 200 applications for one job, and an AI tool that will rank them for an interview shortlist. Who is affected? Two hundred people, each hoping for work. What would a mistake cost them? A fair chance at the job, and the cost falls on them, not on the manager. Could she explain the decision? Only if she has read the applications herself. “The tool ranked you low” is not a reason anyone can answer. Would she tell them AI was involved? If she would rather not, that is a warning sign.

Try it: pick a task, then pick how AI should be involved.

The verdict: the middle column. The European Union lists AI used for hiring as high-risk, and people can over-trust a machine’s ranking as they do any automated aid.¹² So the tool may help her prepare, for instance by summarizing each application against the job’s stated requirements. She reads the applications, makes the shortlist herself, and can give a reason for each choice. Knowing about automation bias, she also reads a few of the applications the tool ranked lowest, to see what it might have missed.

The same questions sort the rest of the figure. A practice schedule for a youth soccer team affects a few families, and a mistake costs a rearranged Saturday, so AI can draft it. A student’s final grade is a judgment about a person, and the AI Act lists AI that evaluates learning outcomes as high-risk;¹ AI might suggest comments, but the teacher grades. Telling an employee that their job is ending is both a decision and a conversation. AI might help plan what to say. A person decides, and a person says it. A summary of your own budget-meeting notes is a good use, in an account your organization approves (Chapter 10); check the numbers before you send it.

A ghostwriter is a good picture of the middle column. A ghostwriter can draft a speech, but the person at the podium answers for every word. And some words, like hard news for one person, the speaker has to find alone. The picture has a limit: a ghostwriter is a person you can question and hold to account. An AI tool can write a fluent explanation of its ranking, but, as Chapter 12 showed, fluent is not the same as checked.

Your turn

Try it. List five tasks from your week where AI could help, at work, at school or at home. For each, answer the four questions and place it in one of the three columns of Figure 14.1. Then pick one task from the middle column and write, in two sentences, the explanation you would give the person affected, without mentioning the tool. If you can’t write it, the tool was deciding. With any AI chat tool you already use, try one more thing: paste in a paragraph you wrote yourself, years ago, and ask whether AI wrote it. Whatever it answers, ask what it could possibly be basing that on.

Think about it. Think of a decision about a person that you make at work or at home: a hire, a grade, a schedule, a care plan. If an AI tool got it wrong, would you notice? And who would that person come to for an explanation?

So when should a decision stay with a person? When it is about someone’s job, grade, health or rights; when a mistake would fall on them; and when someone has to answer for the outcome and explain it. AI can still help prepare, as long as the person who decides has the time, the information and the authority to overrule it. That leaves the questions people ask most about AI: whether it is biased, whether detectors work, what it costs in energy, and what it means for jobs. The last chapter takes them one at a time.