All of this raises a harder question. If careers-page language is mostly a reflection of the labor market, what does it say about worker power?
To find out, I overlaid four language trajectories — idealism, DEI, wellbeing, and performance — against the JOLTS quit rate from the U.S. Bureau of Labor Statistics, a rough proxy for when workers had the leverage to walk. It’s not an exact science, but the patterns are there.
The clearest signal is DEI. Worker-oriented language climbed through the late 2010s and crested in 2021, right on the 2021–22 generational peak in the quits rate, when workers had the most leverage to walk. It stayed elevated for four more years. Then, in 2026, after that leverage had drained back below pre-pandemic levels and the 2024 election turned the political winds, it was pulled back hard: the industry average nearly halved in a single year, to its lowest level in a decade. Meta’s DEI-register share went from near-saturation at the peak to zero; Starbucks’s went to zero too; Salesforce’s fell to less than half its peak. (Not universally — Apple’s held.) The language companies adopted when workers had power was unwound once they didn’t.
That retreat is the heart of it. The concessions weren’t immediate — DEI stayed high for years after the quits rate began falling — but they were sticky, not permanent, and the moment leverage was gone and the politics gave cover, they snapped back. Idealism rode the same cycle with a smaller swing; its crest also lands on the 2021–22 worker-power peak, for all that workers were always a little cynical about billionaires promising to change the world. Performance, which serves whoever can hire and fire, stays flat near the top no matter what.
There’s a fourth line now, and it’s the one that doesn’t follow the script. Wellbeing — the care language of balance, rest, and mental health — looks at first like another worker concession riding the same cycle. It isn’t. It spiked in 2020, with the pandemic, a year before the quits rate peaked — care language answered the emergency, not the bargaining table. And where DEI was pulled back hard once leverage was gone, wellbeing never was: it deflated slowly as the emergency faded, and in 2026 it still sits at roughly double its pre-pandemic average, a little over half its 2020 peak.
What got cut instead is quieter, and you can only see it if you ask who absorbs the care. The bottom panel tracks a second axis — individual versus structural locus of care, the difference between “here’s a meditation app, manage your own energy” and “we staff so no one returns to a backlog.” It drifts toward the individual across the decade, peaks right after the leverage peak, and stays there through 2025. The keyword record in benefits copy says the same thing more bluntly: mental-health and therapy-app language roughly tripled between 2019 and 2025 (from about 3% to 9% of benefit chunks) while family and caregiving language — the kind that costs the company staffing and coverage — collapsed from about 11% of benefit chunks in 2019 to about 1% by 2024. (2026 shows a partial caregiving rebound; that’s one year of data, too early to call.) The care that survived is the kind the worker performs on themselves. That’s not a concession being revoked; it’s a concession being converted into homework.
In 2026 you can’t trust what a company says about itself. But you can trust the pattern of when it stops saying it: the worker-serving language is the part that gets cut when leverage is gone. Cultural signaling doesn’t give workers power — it reflects the power they already have.
How this was measured
Each line is an industry mean over the same nineteen companies (2013–2026), every one scored on all four axes — no per-metric subsetting; the locus panel is a companion measure, not a fifth counterforce. (The three companies added in mid-2026 — Uber, Apple, and Nvidia — have thinner historical coverage, Uber especially.) The quits rate is economy-wide, a proxy for tech bargaining power; read this as co-movement, not causation.
Both counterforces track the quits rate strongly in raw terms (idealism r ≈ +0.72, DEI r ≈ +0.76), and both keep some of that under first-differencing, which strips out shared trend (idealism r ≈ +0.39, DEI r ≈ +0.44) — suggestive, not decisive, on thirteen year-pairs. Which of the two tracks the cycle more closely is not a stable result: DEI led in an eleven-company draft, idealism led at thirteen, and DEI leads again at nineteen — so I won’t hang anything on it. What the differencing understates for DEI is the lag: it peaked with worker power in 2021 but didn’t retreat until 2026, so the year-over-year ticks miss how well the overall shape — surge at the peak, collapse once leverage is gone — follows the cycle. (An earlier draft reported a much starker DEI number that turned out to be an artifact of incomplete data — one company’s mid-decade DEI language was unclassified and counted as zero — which is part of why I treat these correlations as supporting texture, not the finding.)
Performance is the line whose story got more complicated as the dataset grew. Its raw correlation (r ≈ +0.41) used to disappear entirely under differencing; at nineteen companies it doesn’t (r ≈ +0.39), so I can no longer say its wiggles carry no year-over-year co-movement. What still separates it from the counterforces is amplitude and unanimity: its total swing is a sliver — a relative range three to six times smaller than idealism’s or DEI’s — and firm by firm it’s a coin flip (10 of 19 positive, median ≈ 0), where idealism and DEI lean clearly positive (16 and 15 of 19). Performance never surged and never got cut; there was no concession there to revoke. And one honest confound runs through all of it: the 2026 DEI retreat coincides with both fading worker leverage and a political backlash, and this data can’t separate the two. What it can show is that the worker-serving language is the part that got cut.
Wellbeing is measured differently from the other three, and more cautiously. It’s an embedding axis (balance-and-rest against intensity-and-sacrifice) scored on the same mission-page chunks as the values fingerprint, so it shares whatever blind spots that corpus has; an earlier draft dropped it entirely because benefits copy is where most care language lives, but the mission-page signal turned out to be readable after all — the 2020 spike is the largest single-axis jump in the fingerprint set (z ≈ +2.1, ahead of inclusion’s +1.1). Its correlation with the quits rate is the weakest case on this page and I want to be plain about that: raw r ≈ +0.57, but under first-differencing it goes negative (r ≈ −0.25) — driven by 2020, when quits fell and care language exploded. So I’m not claiming wellbeing tracks leverage; I’m claiming the opposite, and that’s the point of including it. The locus shift rests on 78–194 benefit chunks a year, and nearly half of the wellbeing company-years sit on fewer than five mission chunks; the individual-locus axis and the keyword split are two imperfect instruments pointing the same direction, not one clean measurement. The fuller version of this argument, with the receipts, lives in the wellbeing story.