AI is taking the raise, not the job

Most AI-and-jobs studies guess at exposure. This one measured it, using real usage across 321 occupations from 2015 to 2025, with a clear break after ChatGPT in 2023. In the roles most exposed to AI, real wage growth fell 6.7 percent compared with less-exposed roles. Employment did not move in any way you could pin on AI.

The cost is not shared evenly. The bottom wage quartile lost 10.7 percent of wage growth and service workers 24.3 percent, while the top quartile saw no significant effect. Firms are keeping the productivity gain and passing very little of it back, and the people with the least cushion feel it first.

The public fear has been machines taking jobs. The near-term reality in this data is quieter: the same output for less pay growth. For a worker, the risk is not a pink slip, it is becoming cheaper to employ without noticing. The people who capture the productivity themselves, rather than having it captured around them, keep their bargaining power.

If your role is AI-exposed, assume the productivity gain is visible to your employer. Make the part you contribute visible too.

The defensive move is to become the person who produces the gain, not the person the gain is quietly extracted from.

Lower-paid and service roles are exposed sooner than the headlines suggest. Re-skilling is most urgent where it is least discussed.

Keep a simple record of what you now do faster or better with AI. It is the raw material for any pay or promotion conversation.

Turn one repeated task a week into a reusable method you own. Owned methods are leverage; ad hoc speed is not.

Track outcomes, not hours saved. 'I ship this in a day now' beats 'AI helps me a lot.'

This is early, observed, US evidence. The wage effect could grow or fade as adoption matures.

Reading it as 'AI is safe for jobs' is a mistake. No job loss so far is not a forecast.