7 in 10 AI Product Managers Have Rewritten Their Own Job History

Rank every profession on LinkedIn by how often people go back and rewrite a job they already left, and the one at the top is AI product manager. 70.6% of them have done it. At the bottom of the same list sit orthodontists and dental surgeons, at 2.4%.

If you sell B2B software, that gap is your problem, because the titles at the top of that list are your target account list. The distortion in the data you buy doesn’t spread evenly across the market. It pools in the exact roles you point campaigns at.

The finding comes from Time Travel on Professional Profiles, a National Bureau of Economic Research working paper by Nicholas Bloom and Gideon Moore of Stanford with Lisa K. Simon and Caelan Wilkie-Rogers of Revelio Labs. They compared monthly snapshots of 29,448,462 US LinkedIn profiles from 2020 to 2026 and counted how often a profile’s history changed after the fact.

Overall, 19.7% of established users had rewritten the title or description of a job they’d already left. The authors are direct about what that does to anyone reading the data as a record:

“We show that these records are not fixed historical snapshots, but mutable accounts that workers revise over time.”

The averages hide where the problem actually is

19.7% is the number every pickup led with, and on its own it reads like background noise. Table 1 of the paper is where it stops being background noise.

Under-30 workers rewrite at 38.1%, against 9.8% for the over-60s. MBAs sit at 29.3%. Technology and information workers hit 31.6%, per Business Insider’s read of the same table. And at the level of individual professions, the top five are AI product manager (70.6%), platform product owner (60.3%), user research (59.2%), design and qualitative research (56.6%) and AI leadership roles (56.6%).

Read that list again as a targeting export rather than a table. Product, design, research, AI leadership, under 40, tech sector, postgraduate. That’s the standard B2B SaaS buying committee, and it’s the half of LinkedIn whose stated history moves.

The bottom of the list is the useful contrast. Vehicle rental sales, chiropractors, orthodontists, dental surgeons: all between 2.4% and 3.3%. Nobody is retrofitting an AI skill onto a root canal. If your product sold to dentists, none of this would matter to you.

What it does to a number you might already be quoting

The paper runs one exercise that should worry anyone who has put a market-timing chart in a board deck.

Take the 2026 snapshot of LinkedIn, count the jobs from 2022 that mention AI, and you get a picture of how fast AI skills spread. Now compare that against what those same 2022 jobs said at the time. The gap is the retrofitting. By the authors’ estimate, the snapshot approach overstates AI skill presence before ChatGPT by around 30%.

The error has a direction, which is what makes it dangerous rather than merely noisy. The past always looks more AI-ready than it was. So every adoption curve built this way is flatter at the start than the truth, and the “we’re early” read you took off it was wrong in a predictable way.

Their own footnote pushes the same way. The update file only reaches back to 2020, so any rewriting done in 2019 or earlier is invisible, and 19.7% is a floor rather than a measurement.

The edits are a demand signal, which is the more useful half

The part worth stealing comes later in the paper, once the authors stop treating the mutability as a defect and start reading it as a message.

If a worker adds a skill to a job they left four years ago, they are telling you what they believe the market currently pays for. The paper puts it plainly: workers “modify their résumé to reflect what they believe employers want to hear.” That is a live readout of perceived demand, and it has a clock on it.

Time travel runs 64% higher in the period around a job change. In the six months before a move, the rate roughly doubles, from 1.1% to 2.4%, then falls back to baseline within three months of landing. Read that in reverse and a spike in retroactive editing inside an account tells you a population is getting ready to leave.

Set against what people are adding and dropping, the direction is clear. Terms like AI, GPT and LLM are up more than sixfold since late 2022. Work-from-home language stopped being net added at all, and DEI terms fell off sharply at the start of 2025, which Allwork.Space traced to the same window. Your buyers are re-describing themselves around AI competence, in their own words, unprompted, four years after the fact.

That’s the same shape as the problem in B2B deals being decided before the pitch: the signal you want exists, it just isn’t where the vendor dashboard says it is.

Who paid for this, and what the paper doesn’t establish

Two of the four authors, Lisa K. Simon and Caelan Wilkie-Rogers, work at Revelio Labs, the workforce-data company whose data the paper runs on and which sells labor-market intelligence. A finding that profile data needs careful vintage handling isn’t neutral to a firm selling carefully handled vintages. Read it with that in mind, the same way we read a vendor audit of AI citations last week.

Three limits the coverage skipped:

This is a working paper. Its own cover page says these are “circulated for discussion and comment purposes” and have “not been peer-reviewed.”

The headline profession stat rests on a small base. AI product manager is 1,321 profiles out of 29.4 million. The confidence interval is tight (68.1 to 73.0) but the group is narrow, and the authors capped inclusion at professions with at least 1,000 profiles precisely because sampling variation bites below that.

Age is “estimated from education” and gender is “predicted,” neither observed. The age gradient is the paper’s own inference, not a field on a profile.

What to do with it before your next campaign

None of this makes LinkedIn data unusable. It makes it a claim with a known bias rather than a record, and the bias has a direction you can correct for.

Two things are worth doing this quarter. First, stop treating a title-based audience as a fixed population when the title is one of the high-churn ones: product, AI leadership, UX, anything under 40 in tech. Those segments are being rewritten underneath your campaign at three times the rate of the market as a whole.

Second, if you are using “AI” appearing in a prospect’s history as an adoption or intent signal, the paper says roughly 30% of what you’re seeing before 2023 was added later. Discount it or drop the pre-2023 window entirely.

And there’s a use nobody is selling you yet. Retroactive editing inside a target account is a leading indicator of departure, at double the base rate six months out. If your CRM is quietly going stale because champions are leaving, that movement was visible on their profiles before it was visible in your pipeline. Pull the last two quarters of edits across your top 50 accounts and see whether the churn you got surprised by was signposted.