Safety decisions stay with people

In Level 1 you learned that people decide. In logistics that principle is sharpest for three decisions, because a wrong call does not stop on paper:

• Vehicle roadworthiness. Whether the brakes, tyres, lights, steering and load are safe for this trip. The person who decides is the one who inspects the vehicle in person, usually a mechanic or inspector, following company procedure.
• Driver fatigue and working hours. Whether the driver has rested enough and is still within working-hour limits. There are limits from regulation and from company policy, and there are things no system records: a driver who did not sleep last night because their child was ill. The person who decides is the supervisor who talks to the driver.
• Dangerous goods. Whether a load counts as dangerous goods, how it is packed and marked, which vehicle may carry it, and which documents must travel with it. Carrying dangerous goods is specially regulated. The person who decides is an appointed, trained officer, based on the goods' safety data sheet and the rules in force.

All three share one thing: the key information is in the real world, not in text you can paste. A model only knows what you tell it, and it can still write an answer that sounds sure. That is the most dangerous form of hallucination in this field: a sentence like "this condition is usually still safe" that rests on nothing.

That does not mean AI is useless for safety. It can help a lot on the preparation side:

• tidying the daily vehicle checklist so it runs in order and is easy to tick in the field;
• drafting the questions a mechanic or supervisor should ask when a driver reports an odd noise;
• summarising last month's inspection notes (with vehicle codes, not people's names) into a list of repeat findings;
• drafting a rest reminder for drivers in polite, clear language;
• explaining terms on a safety data sheet in plain words, which you then match against the sheet.

The difference is the last sentence in each process: who says "cleared to go". That sentence never comes from a model. It comes from an authorised person, and it is recorded: who, when, based on which inspection. That record is a small audit log that answers questions after the fact, and also reminds the person signing that this decision is theirs.

A practical way to write it in a prompt: put in the Limits line something like "Do not state whether the vehicle is roadworthy, and do not judge whether the driver may continue." If the output still contains such a judgement, strike it out; do not pass it on.

The hardest part is not understanding the principle but holding on to it when the schedule is tight. The customer is waiting, there is no spare truck, and an AI answer saying "probably still safe" feels like a way out. At such moments, remember that the responsibility does not move to the model. If something happens, the question is who allowed that vehicle to leave.

So the principle of human in the loop needs procedure behind it, not just good intentions: a departure status only certain roles can change, a required record with the inspector's name, and a habit that a delay for safety reasons is never blamed on anyone. The customer message can wait; a safety decision must not be rushed by a draft a machine wrote.