Here is a survey result that should change how organisations plan their AI training.
89.5% of AI users surveyed say checking AI answers is fully or mostly their own responsibility.
That is almost everyone. People are not confused about whose job it is. They accept that if the AI is wrong, it is on them to find out.
Now the second result, from the same survey.
70.1% of AI users surveyed check AI answers only sometimes, rarely or never.
So most people agree that checking is their job, and most people do not do it every time. Both are true at once.
It is not a knowledge problem
The usual response to AI risk is training. Teach people that AI makes mistakes. Teach them to check. Give them a short course and a set of rules.
But the survey shows that people already know. They do not need to be told that checking matters. They have told the researchers so themselves.
The same pattern appears at work. In a 2025 study, KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. Of the employees who use AI at work, 66% said they had relied on AI output without evaluating its accuracy at least once. Of the same employees, 42% said they did so sometimes or more often.
The gap between knowing and doing is not closed by more knowing.
Why habit wins
Checking takes time. The AI answer takes seconds. Most of the time, the answer is good enough. Each skipped check feels reasonable, and each one is usually fine.
So people learn, one task at a time, that checking rarely changes anything. The habit forms quietly. Then one day the AI is wrong on something that matters, and the habit does not switch off.
This is not a failure of character. It is how habits form under time pressure. It is also why a course on Monday rarely changes what someone does on Friday at 17:00. By then there is a report due and an answer on the screen.
What the law asks, and does not
Regulators have noticed that AI use needs human skill. The EU AI Act, in Article 4, requires providers and deployers to take measures to support their staff's AI literacy.
It is worth reading what Article 4 does not require. It sets no required level. It sets no duty to measure staff. It sets no certificate. It asks organisations to act, and leaves the form open.
That leaves room for an honest question. If most people already know they should check, what kind of measure would change what they actually do?
From knowing to doing
Awareness is a start. But a habit changes through practice and feedback, not through information alone.
- Practise on realistic cases. People need to make the decision themselves, with an AI that is sometimes wrong, under some time pressure. A quiz about AI risk is not the same.
- Show people their own route. A person who sees what they checked, and what they let through, learns more than one who is told what to check.
- Talk about it as a team. When a team compares how each person checked the same case, habits become visible and open to discussion.
- Make it safe. People play honestly when the exercise cannot count against them.
The question for leaders is not "Have my people been trained on AI?" Most have, or soon will be. It is "When a real answer is on the screen and time is short, what do my people do?"
Sources
- F-Secure, AI chatbots: navigating distrust and usefulness in everyday life. Source: F-Secure, AI chatbot survey, April 2026, 1,500 AI users. Page.
- KPMG and the University of Melbourne (2025). Trust, attitudes and use of AI: A global study 2025. Survey of 48,000+ people in 47 countries, November 2024 to January 2025. Press release. Country insights.
- EU AI Act, Article 4, as amended. European Commission, AI literacy questions and answers.
