Seven questions to ask before AI decides for you.

01Was this system independently tested by someone who does not profit from it passing?

The vendor's own numbers are the vendor grading itself. In case after documented case, the failure was visible in how the system behaved, and nobody independent had looked.

A good answer sounds like. Yes, here is the independent evaluation, who performed it, and what they were paid to find. Not a brochure. A named evaluator with nothing to sell but the result.

02What happens when it is wrong, and can the decision be undone?

Some wrong decisions are an inconvenience. Others, a denied home, a revoked benefit, an arrest, cannot simply be reversed. The harder a decision is to unwind, the higher the bar before a machine touches it.

A good answer sounds like. We classified this decision by how reversible it is, and the least reversible ones require a human sign-off before anything happens.

03Can I actually reach a human, and does asking for one work?

A right to appeal that leads to a phone tree is not a right. In the worst documented failures, tens of thousands of people were accused by a machine with no human between the flag and the penalty.

A good answer sounds like. Here is the path to a person, here is how long it takes, and the system is measured on whether that path works, not just whether it exists.

04Does it treat people like me the same as everyone else?

Systems have scored sicker patients as healthier, screened out older applicants automatically, and flagged families by nationality. The pattern rarely announces itself. It has to be tested for.

A good answer sounds like. We test outcomes across groups, neighborhoods, and applicant types, someone independent checks that testing, and here is what they found.

05Can they explain the decision, in writing, in words I can appeal?

A decision nobody can explain is a decision nobody can challenge. If the organization cannot state the reasons on the record, neither can you at your appeal.

A good answer sounds like. Every decision comes with stated reasons specific enough to contest, and a denial you cannot examine is treated as a defect, not a feature.

06How does it behave when nobody is watching?

Systems, like people, can behave differently under observation. The question is whether anyone has measured the difference, or whether the demo is all anyone ever checked.

A good answer sounds like. It was evaluated under conditions it could not distinguish from real work, not just on a stage.

07Is there a record of what it did, kept honestly, checked by someone other than its operator?

In Milwaukee, an officer tracked two people through a license-plate system by typing the word test into the justification field. The record existed. Nobody independent was reading it until another case forced an audit.

A good answer sounds like. Every action is logged, the log cannot be edited by the people it watches, and an outside party reviews it on a schedule, not after the harm.