Workers are 64% more likely to change employers in the period they update their professional profile (NBER, 2026). Not their current role โ their job history. The title of a position they left years ago, the description of work that was finished and signed off long before anyone thought to revise it.
That is the finding at the centre of a new NBER working paper by Nicholas Bloom, Gideon Moore, Lisa Simon and Caelan Wilkie-Rogers. Using monthly vintages of Revelio Labs data from 2020 to 2026, they show that 19.7% of established U.S. LinkedIn users retroactively edit the title or description of a job they have already left โ and that the median edit lands more than four years after the person walked out the door (NBER, 2026).
Two consequences follow, and most operations teams are exposed to both. The first is a retention signal that arrives months before the resignation letter. The second is a data problem sitting underneath every skills benchmark built on profile history.
What the NBER Study Actually Measured
Most profile research takes a single snapshot and treats it as the record. This study does something more useful: it compares each profile against its own earlier versions, month by month, across roughly 29.4 million U.S. profiles (TNW, 2026). That makes it possible to see not just what a profile says, but when it started saying it.
The authors call a retroactive edit "time travel," and the pattern is not marginal.
- One in five established users did it at least once since 2020 โ and the authors describe that as a lower bound, because their update file only begins in 2020.
- More than four years is the median gap between leaving a job and rewriting it. These are not people correcting a typo on the way out.
- Nearly one in three workers in technology and information time travel by the end of the sample, the highest rate of any industry; coverage of the paper puts it at 31.6% (TNW, 2026). Construction and real estate sit at the other end.
The concentration matters. The industries with the highest rates are the ones where mid-market knowledge work lives โ and where replacement is most expensive.
Retention Risk Shows Up Months Before the Resignation
The timing is what makes this more than a curiosity.
Relative to non-movers, workers who change employers are more than twice as likely to time travel even a full year before their move. From roughly six months out, the rate climbs further โ eventually reaching double its twelve-months-prior level, 2.4% against 1.1%. Within about three months of the move it falls back to baseline (NBER, 2026).
Read that as an operations leader. The signal leads the exit, not the other way round. By the time a resignation reaches a manager, the profile edit is often half a year old.
Why the lead time matters
Gallup's research on voluntary turnover frames the cost of missing that window. 42% of employees who voluntarily left their organization say their manager or organization could have done something to prevent it. 45% say that in the three months before leaving, neither their manager nor any other leader proactively discussed their job satisfaction, performance or future with them (Gallup, 2026).
Gallup puts replacement at around 200% of salary for leaders and managers, 80% for technical professionals and 40% for frontline employees. In a 150-person company, a handful of avoidable senior exits is a line item the CFO notices.
So the gap is not information about intent. Workers are publishing it. The gap is that nobody is having the conversation during the months when it could still change the outcome.
Your Job History Data Is Being Edited Toward AI
The second consequence is less visible and, for anyone buying talent intelligence, more expensive.
Retroactive edits are not random. After ChatGPT's release in late 2022, retroactive additions of AI-related language โ "AI," "GPT," "LLM," "artificial intelligence" โ rose more than sixfold, before softening in recent months. Remote-work and DEI language moved in the opposite direction. LLM-associated writing markers also surged after ChatGPT, most sharply among less-educated groups and MBAs from lower-ranked programs (NBER, 2026).
The authors offer a plausible reading of the mechanism, framed carefully as an interpretation rather than a finding: workers modify their rรฉsumรฉ to reflect what they believe employers want to hear (NBER, 2026).
The quantitative consequence is the one to carry into your next vendor meeting. Estimating AI-skill prevalence in 2022 from today's profiles, rather than from how those profiles looked in 2022, overstates AI skills by around 30% โ and again, the authors call this a lower bound (NBER, 2026).
What this does to a skills-gap baseline
If your workforce plan benchmarks your people against "market" AI-skill prevalence derived from current profile data, the market's history has been edited upward. The trend line is steeper in the data than it was in reality. Your team looks further behind than it is, and a training budget gets sized against a past that did not exist.
What it does to your hiring screen
The same logic applies to candidates. As Human Resources Director put it in its coverage, employers can no longer take an old job title at face value (Human Resources Director, 2026). A 2021 role that now reads "led AI-driven process automation" may have been written in 2025. Any screen โ human or model โ that treats job history as a contemporaneous record is rewarding the most recent rewrite, not the underlying experience.
The Objection: This Is Surveillance With Extra Steps
This is the pushback worth taking seriously, and it has three parts.
The signal is correlational. The authors are explicit that the link could be selection โ people polish their profile because they are already searching โ or treatment, where a polished profile improves the odds of landing a job. For a retention decision, the distinction barely matters: either way, the edit is information that someone is closer to the door than they were. But it does rule out treating an edit as proof of anything.
The absolute rates are small. A 64% relative lift on a low base rate is still a low base rate. A 2.4% monthly rate means the overwhelming majority of profile edits are not followed by an exit. Build an individual watchlist from this and you will generate far more false alarms than departures โ and you will spend managerial attention on the wrong people.
Monitoring corrodes the thing it measures. Employees who discover that someone is tracking their public profile will reasonably conclude they are not trusted. That accelerates exactly the exits you were trying to prevent. Depending on jurisdiction, it may also raise data-protection questions your counsel should answer before you build anything.
Take the objection seriously and it sharpens the conclusion. The profile edit is not the intervention. It is evidence that the real intervention โ a regular, substantive conversation about someone's future โ is happening too late or not at all. Gallup's 45% is the number that matters here, not the 64%.
What Separates a Signal From a Watchlist
The practical line is simple: use the finding to change the cadence of your conversations, not to build a monitoring system.
The study tells you three things you can act on without looking at a single employee's profile:
- Exit intent becomes visible roughly six months ahead. If stay conversations happen only at annual review, you are structurally late for most departures.
- Your highest-risk population is predictable. Technical and knowledge roles time travel most and cost most to replace. That is where conversation cadence should be tightest.
- Historical profile data is not a record. Any people-analytics model or skills benchmark built on it inherits an upward bias toward whatever the market currently rewards.
What to Do With Job History Data This Quarter
The move is smaller than it sounds and does not require new software.
Put a quarterly stay conversation on the calendar for every technical and senior role. Not a performance review. One question set: what would make you leave, what would make you stay, what do you want to be doing in eighteen months. The point is to close Gallup's 45% gap before the six-month window opens.
Ask every talent-intelligence vendor one question. Are your historical skill benchmarks built from point-in-time profile vintages, or reconstructed from today's snapshot? If the answer is the latter, discount the historical trend and ask for the correction factor. Treat around 30% as the floor for AI skills.
Re-baseline any skills-gap analysis that compares your team to "the market." If the market series came from current profiles, your team's gap is likely overstated. Resize the training budget before you commit it.
Stop treating job history as evidence in screening. Verify the experience that matters through work samples, structured interviews and references โ not through titles that may have been rewritten four years after the fact.
The Record Is Being Rewritten in Plain Sight
The most useful thing in this paper is not the 64%. It is the realisation that job history โ the data every workforce plan, skills benchmark and hiring screen quietly assumes is fixed โ is a living document, edited toward whatever the market rewards this year.
Your people are already telling you, months in advance, when they are reconsidering where they work. Most organizations have no mechanism to hear it, because the only conversation scheduled is the annual one.
Before the quarter closes, decide which roles get a stay conversation every ninety days โ and put the first round on the calendar this month.
