The question behind the AI panic is not whether the tools work. It is what happens to workers after they do.
Christoph Trattner says that when an employee uses AI to finish a report faster, the upside can be immediate: better assignments, a pay rise or promotion. But the downside is just as clear. A manager may decide that a team of three can become a team of one.
That fear is not theoretical. Glean surveyed 6,000 people who mostly work on computers in the U.S., the U.K. and Australia, and found that 36% of workers who said AI helped them get more done and produce better work also said they hid how much it helped them. A 2024 Microsoft and LinkedIn survey found that more than half of people using AI at work worried that using it on important tasks would make them look replaceable.
Trattner compares the moment to factory life more than a century ago, when workers held back production because they did not want managers to learn how much they could really produce. Frederick Winslow Taylor described that behavior in his 1911 book on scientific management and called it 'soldiering.'
His point lands in the AI era because the reward structure has not changed as much as the technology has. If a worker shows how much faster the job can be done, the company may simply raise expectations and keep the savings.
Norway is trying a different model. On October 2, the Norwegian government and eight employer and union organizations signed an AI agreement running through 2029. It calls for employees to be involved when AI is introduced and says the gains should benefit workers as well as businesses and society.
The agreement is only a statement of intent. It does not guarantee pay rises or a share of savings. But it does point to a basic principle that free enterprise works better when incentives are clear: if productivity gains go somewhere other than the people creating them, workers will hide them.
There is also a hard reality in the numbers. In 2025, 52.6% of Norwegian employees belonged to a union, compared with 10% of American workers and 5.9% in the U.S. private sector. That matters because a U.S. employer can copy the conversation, but not the enforcement.
Trattner suggests a simple test: if workers share their AI shortcuts, do they get a say before management uses the savings to raise targets or shrink the headcount? Or are they asked for feedback after the decision is already made?
That is where the market problem sits. Capital rewards clear rules, but workers respond to the rules in front of them. If AI is going to lift productivity, the people using it need a believable path to benefit from the gains. Otherwise, the incentive is not innovation. It is concealment.
The next fight over AI will not be about the software. It will be about leverage, wages and who gets the first claim on the savings.
