The King Midas Test for Family AI Prompts

King Midas received exactly what he asked for.

Everything he touched became gold.

The wish succeeded as a sentence and failed as a life.

This old story offers a useful way to talk with children about AI. A system can follow the visible instruction while missing the values, limits and consequences that remained unspoken.

A prompt usually contains a visible goal and hidden expectations

Suppose a child asks an AI:

“Write the most convincing speech so I definitely win.”

The visible goal is winning.

The child may also expect the speech to be truthful, respectful, age-appropriate and based on their own ideas. Those expectations are not guaranteed merely because they feel obvious to the person asking.

A better request makes important limits explicit.

Ask what could be sacrificed

Every narrow goal can create a trade-off.

• Maximise speed and accuracy may fall.
• Maximise clicks and quality may fall.
• Make the answer simple and important nuance may disappear.
• Finish the homework and the child’s own thinking may disappear.

Before using an AI result, ask:

“What might this answer have ignored in order to satisfy the request?”

Do not let one metric become the whole goal

Children often meet systems optimised for visible numbers: views, streaks, speed, marks and completion.

Numbers can provide useful feedback. They can also replace the larger purpose.

For example, a reading app may reward pages completed. The family may care more about whether the child understood, questioned and enjoyed the text.

Ask children to name both:

• The measured goal: What number or result is visible?
• The human goal: What do we actually care about?

Treat AI instructions as drafts

A prompt does not need to be perfect before the first response. It should remain open to correction.

1. Ask for a first version.
2. Inspect what the AI prioritised.
3. Identify what was missing or distorted.
4. Add limits, values or examples.
5. Compare the new version with the original goal.

This teaches children that using AI is not a one-command transaction. It is a process of specifying, checking and revising.

Keep human judgment for difficult decisions

Some tasks involve values that cannot be reduced to a simple instruction.

An AI can suggest ways to resolve a disagreement. It does not know the full history of the relationship. It can help compare options. It cannot carry responsibility for the family’s choice.

A useful boundary is:

Use AI to widen the options. Keep people responsible for the decision.

The five-question Midas test

Before accepting an important AI answer, ask:

1. What did we ask the AI to optimise?
2. What did we assume without saying?
3. What might be sacrificed to reach the visible goal?
4. How will we recognise an unacceptable result?
5. Who will review and take responsibility for the decision?

A family experiment

Give an AI a harmless narrow request, such as:

“Create the fastest possible morning routine.”

Examine what the answer sacrifices. Does it leave time for breakfast, conversation, missing items or a younger child who needs help?

Rewrite the prompt to include what the family actually values.

The lesson from King Midas is not that goals are dangerous.

It is that a goal becomes safer when we make its limits visible and remain willing to correct it.