Number of the Week: AI and Older Workers

Number of the Week: AI and Older Workers

11.1%. That is the share of programmers over 55 who left their jobs within a year, measured since ChatGPT showed up. Before, it was 8.7%. A jump of more than 25%, and this week the tech press drew the same conclusion from it: AI is pushing seniors out the door.

The number checks out. We went and verified it, line by line, against the source. The problem sits two pages later, in the same document: spray painters leave their jobs at a rate of 13.7%. The supposed victims of AI are quitting less often than the workers AI barely touches.

This is the first installment of a new column

Starting today, every week we take one AI number that everyone is repeating, trace it back to its primary source, and publish the gap between what the source actually says and what got said about it.

The premise fits in one sentence: the number is almost always right, it is the story around it that goes off the rails. We are not hunting for lies. We are hunting for the exact spot where a careful study turns into a dramatic headline.

The idea did not come from a desire to lecture anyone. It came from an accident: three days into monitoring the news cycle, we ran into the same pattern twice. The first time, we nearly republished it ourselves. An accident is forgivable. A pattern is worth documenting.

Two rules follow from that: nothing runs here without a full read of the primary source, never its summary, and if a write-up is faithful to the source, there is no column that week. We are not going to manufacture a gap that does not exist just to hit a deadline.

Where the number comes from

The study is called "Are the Careers of Older Workers Being Cut Short by AI?", written by Geoffrey Sanzenbacher at Boston College's Center for Retirement Research (Issue Brief 26-13, June 2026). First detail worth noting: that title is a question, not a conclusion.

The methodology combines the Current Population Survey, the same monthly survey used to calculate the US unemployment rate, with an AI exposure index built by Tufts. 102,097 observations, workers 55 and older, tracked year over year, before and after ChatGPT. The index does not measure whether a job can be replaced. It measures how well AI can perform the tasks that make up that job: a programmer scores 97 out of 100, a spray painter scores 5.

The study establishes an association, not causation, and the author does not hide it: "Like any early analysis of the impact of AI, caution is in order." He also flags two biases pulling in opposite directions: Trump administration cuts to R&D funding, which could inflate the measured effect, and the AI startup boom, which is driving up demand for data scientists and may be masking what happens once that growth cools off.

Unemployed, not retired

The decisive detail is buried in the appendix, page 9. The study tests four outcomes: no longer working, unemployed, out of the labor force, retired. After ChatGPT, the effect of AI exposure is statistically significant on the first two. On the other two, it is not. Those bars do not say anything.

Exposed seniors are not heading off to enjoy retirement. They are out of work, and looking for another job. A 58-year-old who is unemployed and a 58-year-old who is retired are two very different public policy problems.

What the study actually measures: a fading premium

Line up the two numbers again. Programmers go from 8.7% to 11.1% in exits. Painters go from 13.5% to 13.7%. The gap between the two jobs shrank from 4.8 points to 2.6, nearly cut in half.

Before ChatGPT, AI-exposed jobs handed their older workers something valuable: a safer late career than average. Less physical, better paid, more credentialed. In the sample, exposed workers earn $1,410 a week versus $869 for everyone else.

That is the gap AI is grinding down. Not the career. The premium. Sanzenbacher says it plainly: "AI exposure may reduce the gap in career length between low- and high-paying jobs."

The real story is less dramatic than "AI is firing seniors." It is more unsettling than that. The premium was a return on scarcity: a skill few people had protected the person who had it. Once a machine learns to do that job too, the protection erodes long before the job itself disappears. The floor does not collapse. It just tilts.

Six headlines, one PDF

The brief came out in June. Rediscovered in mid-July, it produced this in 72 hours.

Futurism, July 11: "AI Is Pushing Older Employees Straight Out of the Workforce." Leaving the labor force is exactly one of the two non-significant results.

InvestmentNews, July 13: "AI may be nudging some older workers into early retirement." Retirement is the other non-significant result. The article itself correctly notes that the rise shows up in transitions to unemployment. The headline contradicts the story underneath it.

The Next Web, July 13: "AI is ending older workers' careers early, and it is coming for the well-paid ones first." The study's opening question has been rewritten as a fact.

CNBC, July 13: "AI is changing older workers' careers, research finds." So it was possible, after all.

The distortion lives in the headlines, not the articles

This is the heart of this first column, and fairness demands saying it: the articles themselves hold up. TNW quotes the author's "caution is in order," correctly notes the increase shows up in unemployment rather than retirement, and even points out that programmers still leave their jobs less often than painters. It is all there. Even the "coming for the well-paid ones first" line is not invented. It is in the brief, word for word. What the headlines cut is the sentence that comes right after.

The distortion, then, lives in the headline layer, the one that has eight words to earn a click, and the only one most readers actually read. Nobody lied. The system took in a careful question and put out a certainty. We have already seen how fear of AI has a market value of its own.

What still holds

The fact-check does not demolish the study. It just puts it back on its feet. Seniors in AI-exposed jobs really are leaving work more often than before. The increase is real, measured, significant.

And the early-retirement story is not wrong, just premature. Sanzenbacher points back to his own 2019 research: an involuntary job loss often leads, eventually, to retiring earlier than planned. Today's unemployment could become tomorrow's early retirement. It just does not show up in the data yet.

That is the whole distinction: what we know versus what we fear. A serious study keeps the two apart. A headline mixes them together.

What nobody put in a headline is the real signal here. For forty years, the advice for anyone hoping to work until retirement has stayed the same: move up the skill ladder, head toward knowledge work, get away from physical labor. That advice still works. It just works 2.2 points worse than it used to.