06-reference/research

technological displacement career vs generation

2026-07-28·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
labor-displacementautomationeconomic-historyai-labor-marketcareer-strategy

Within a career or across a generation: what eight displacements actually show

The question

Verbatim: "Across 6-8 documented technological displacements (power looms, NYSE floor traders, newspaper journalism, bank tellers, travel agents, switchboard operators, typesetters), was the displaced cohort absorbed WITHIN a career or only across a GENERATION — and what structural variable predicts which?"

Raised by the founder 2026-07-26 ~22:00 after the ATM-vs-farming exchange, with a candidate predictor to test, not assume: whether the successor job required the displaced job's skill.

Short answer: across a generation, essentially always. The candidate predictor fails. What actually predicts the difference is the rate of decline measured against workforce turnover, plus whether the move to the successor job was administrative (same employer, no job search) or market-mediated.

What we already know (from the vault)

The per-case table

Citations are numbered refs to the Sources section; every one is a working URL. "Not established" means no source I fetched supports a figure. Nothing here is estimated to make a cell look full.

# Case & window (a) Onset → trough (b) Same people or different? Evidence for (b) (c) Total sector employment rose? (d) Successor needed the old skill?
1 English handloom weavers, power loom, ~1815 → 1860s ~40-45 yrs Different. Mills preferentially hired women and children; male handloom weavers largely were not absorbed. Peak ~240-250k in the 1820s [unverified — every primary mirror was 403-blocked] INFERRED from aggregate counts and contemporary testimony. No linked-individual study located. Yes. Power looms automated ~98% of the labor per yard of cloth, yet factory weaving jobs increased, because cheaper cloth expanded demand faster than labor per yard fell [2] No. Handloom craft skill did not transfer. Bessen's twist: factory weavers built substantial human capital, but on the job, not carried in [19, abstract only]
2 US agriculture, mechanization, 1870 → 1970 ~100 yrs; absolute headcount peaked in 1910, 40+ yrs after the share began falling Different — generationally. See the arithmetic below; the pattern is mechanically incompatible with incumbent expulsion. INFERRED, but from an unusually strong structural signature (share falls while headcount rises = exit by non-entry). No individual tracking located. Yes. Total gainful workers 29.07M (1900) → 53.30M (1940) while agriculture went 10.95M → 9.00M [9] No. Farm to factory transferred essentially nothing.
3 Telephone switchboard operators, dial cutover, 1920 → 1940 (paper window) Step function, not a ramp: operator employment for women 16-25 in a cutover city fell 50-80% immediately and permanently [1] BOTH, and this is the decisive case. Incumbents were dispossessed: "less likely to be in the same job the next decade, less likely to be working at all, and conditional on working were more likely to be in lower-paying occupations." Subsequent cohorts were fully absorbed: "automation did not reduce employment rates in subsequent cohorts of young women, who found work in other sectors" — typists, secretaries, some food service [1] LONGITUDINAL — the only genuine one in the table. Genealogy-based census linking (FamilySearch trees) follows 16,253 named women 1920→1930 and 11,220 1930→1940, solving the surname-change problem that defeats standard linkage. Cohort half is cross-sectional city-panel. Not established. Paper does not quantify total communications-sector employment over the window. Operators peaked at 180,000 in 1930 [1] Partial. Typist/secretary is adjacent clerical work at similar wages and demographics [1] — and incumbents still were not absorbed.
4 Typesetters / compositors, hot metal → computerized composition, ~1964 → 1990 ~25 yrs. NYT's last hot-metal issue was 1978 [14]; ITU membership halved 1984-87 [12] Incumbents held in place by contract; successors never hired. The 1974 NYC newspaper agreement traded automation rights for lifetime employment guarantees to existing situation-holders [unverified — NYT source unfetchable]. A distinct 1964 NYC contract had already created an employer-funded "automation fund" [18]. The occupation ended when the guaranteed incumbents retired. INFERRED, plus a single anecdote (the NYT's last Linotype operator moved to Mac/InDesign ad production, retired 2016). No study analogous to [1]. Not established for printing/newspaper employment in this window. Partial. Typographic judgment transferred; hot-metal keyboard and casting skill did not.
5 Bank tellers, ATM 1969-2010, then mobile 2010-present Two distinct shocks. ATM era: no trough — tellers/branch fell 20→13 (1988-2004) but urban branches rose 43%, so headcount held [2]; +11.6% 1997-2007 [4]. Mobile era: 502,700 (2016) → 347,400 (2024), BLS projects -13% 2024-2034 [3][4] Same people, during the ATM era only — the role re-tasked in place toward sales and relationship work [2]. The mobile-era decline is genuine and ongoing. CROSS-SECTIONAL. Burning Glass Institute claims a "Worker Career Histories" longitudinal analysis (64% of tellers promoted within 3 yrs, only 4% into higher-paying roles like loan officer) [17] — report its existence, not its findings; methodology unverified. Not established. Could not fetch a clean total commercial-banking employment series. Yes. Teller → relationship banker transfers customer and product knowledge directly.
6 Travel agents, commission cuts 1995/2002 + online booking ~5 yrs, the fastest in the table. Agency count 46,659 (1998) → ~24,000 (2003), roughly -49% [unverified — archive sources were search-snippet only] Not established. CROSS-SECTIONAL only. No study following individual displaced agents located. Probably, but not established to standard. Leisure/hospitality grew while the occupation collapsed to 65,700 (2024) [5] Yes. The surviving job is the same job for a narrower clientele — BLS now projects +2% growth 2024-2034 on "expertise in recommending options... for personalized travel experiences" [5]
7 NYSE floor traders, Reg NMS + Hybrid Market, ~1997 → 2010s ~10-15 yrs. Floor headcount fell roughly an order of magnitude; sources irreconcilable (5,000+ in the 1970s vs 4,000 vs 2,100 vs ~700, because each counts a different population). The current NYSE floor-broker directory lists ~23 contacts across ~20 firms [16] Not established. ANECDOTAL ONLY. Named individuals in Crain's/Forbes/FIA (a Goldman desk going from 56 floor staff to "one or two"; a 25-year broker becoming a Morgan Stanley advisor). No systematic dataset. Yes, directionally. NAICS 523000 total employment 1,064,060 (May 2023) [6]; 1,167.1k (June 2026) [7]. Pre-1995 baseline not established. Partial. Market intuition transferred; physical open-outcry execution skill did not.
8 Newspaper journalism, ~2000 → present 25+ yrs and still declining — no trough reached. Newspaper publishing (NAICS 51111): 400,000+ in the late 1990s → 86,000 in Dec 2024 [15, secondary] Not established. Two conflicting indirect signals: UK data that >half of laid-off newsroom staff were over 40, versus a US survey finding layoffs fell disproportionately on younger journalists. Unresolved. CROSS-SECTIONAL. No individual-tracking study located. NO — the important exception. Pew: total newsroom employment 114,000 (2008) → 85,000 (2020), -26%. Newspapers -57% (71k→31k); digital-native +144% (7.4k→18k). Digital gains of ~+10,600 nowhere near offset newspaper losses of ~-40,000 [8] Yes, near-completely. The successor job is the same job on a different substrate.

Testing the candidate predictor: does skill transfer explain the variation?

It does not. Three independent failures, one of them logically fatal.

Fatal failure — the predictor is single-valued but the outcome bifurcates. In the best-identified case in the entire literature, switchboard operators, skill transfer to the successor job (typist, secretary) is a single fixed value, yet the outcome splits cleanly in two: incumbents lost jobs, earnings and labor-force attachment, while the very next cohort of young women was absorbed with no measurable employment penalty [1]. One input cannot produce two outputs. Whatever explains the split, it is not skill transfer, because skill transfer was held constant across the two groups by construction.

Failure two — the highest-transfer case has the worst outcome. Newspaper journalism is the maximum of column (d): digital reporting is the same craft on a different substrate. It is also the only case in the table where total sector employment fell rather than rose, by 26% [8]. If skill transfer predicted absorption, journalism should be the easiest case in the set. It is the hardest.

Failure three — the same predictor value yields opposite outcomes across time. Teller-to-relationship-banker skill transfer did not change between 2000 and 2024. The outcome reversed completely: growth through 2007, then 502,700 → 347,400 by 2024 with -13% projected [3][4]. A time-invariant predictor cannot explain a time-varying outcome.

Skill transfer is the story workers tell themselves. The evidence does not support it as a protective variable.

What actually predicts it

1. Rate of decline measured against workforce turnover (the strongest candidate)

Compute the annualized decline rate from the cited figures and the ordering is striking:

Case Annualized decline Incumbent harm observed?
Agriculture, 1910-1940 -0.77%/yr (11.35M → 9.00M) [9] None visible
Tellers, 1997-2007 +1.1%/yr (growing) [4] None
Tellers, 2016-2024 -4.5%/yr (502.7k → 347.4k) [3][4] Moderate, ongoing
Newspaper publishing, ~1997-2024 -5.5%/yr (400k+ → 86k) [15] Severe
Travel agencies, 1998-2003 -12.4%/yr (46,659 → ~24,000) [unverified] Severe
Switchboard, at city cutover -50 to -80% in a step [1] Severe, measured longitudinally

That ordering tracks incumbent harm far better than column (d) does. The mechanism is simple: every occupation loses some percentage of its incumbents each year to retirement and voluntary exit anyway. If the occupation shrinks slower than that natural attrition, it empties by non-replacement and literally nobody has to be displaced. If it shrinks faster, incumbents are expelled regardless of how well their skills transfer.

Agriculture is the clean demonstration. Its share of the labor force fell from 53.0% (1870) to 16.9% (1940), yet its absolute headcount kept rising until 1910 [9]. Forty years of collapsing share with growing headcount is mechanically impossible via incumbent expulsion. It can only happen one way: farmers' children did not become farmers. That is a generational transition by definition, and it is why nobody experienced it as a displacement event.

Caveat on this candidate: the benchmark attrition rate is the number the whole hypothesis pivots on, and I did not fetch a source for it. The ordering above stands on its own arithmetic; the threshold does not. See follow-up 1.

2. Whether the transition was administrative or market-mediated

The teller case is usually read as proof that skill adjacency saves you. It is better read as proof that not having to look for a job saves you. A teller became a relationship banker at the same employer, in the same building, without a job search, during a period when that employer was opening 43% more branches [2]. A handloom weaver had to persuade a different employer in a different town who preferred to hire women and children. Same-firm redeployment is an administrative act; everything else requires a market to clear, and markets clear badly for a 45-year-old with one occupation on the résumé.

3. Step functions hide inside smooth national aggregates

The national telephone-operator series looks like a gentle ramp. The lived experience was a single-day cutover per city, with a 50-80% local collapse [1]. This is precisely why Feigenbaum and Gross could identify the effect at all — the local step function is the natural experiment. Any national-aggregate reassurance about AI and jobs is committing this error. The aggregate is not the experience.

4. Institutional buffering works, but it converts rather than solves

The 1974 NYC printers' agreement is the only case in the table where incumbents were demonstrably protected, and it was protected by contract, not by market absorption. It did not save the occupation: ITU membership halved 1984-87 and the 134-year-old union voted 29,740 to 7,265 to dissolve into the CWA in 1986 [12][13]. Buffering bought the incumbents a career and let the occupation die with them. That is a generational transition purchased at the employer's expense, not a within-career absorption.

The reframe

"Within a career" absorption is not a natural outcome that sometimes occurs. Across all eight cases it is either mislabeled (the occupation did not die, it re-tasked while keeping its name — tellers 1990-2010, factory weavers) or purchased (an institution paid to hold incumbents in place — NYC printers 1974). Where an occupation genuinely died and nobody paid to protect the incumbents, the incumbents were not absorbed. The absorption showed up one cohort later, in people who never entered the occupation at all.

Correcting the received narratives

The ATM/teller datum — the most-misquoted number in this debate. The popular version is "ATMs didn't reduce teller jobs, therefore automation doesn't kill jobs." Bessen's actual claim is narrower and conditional: tellers per branch fell 35% (20 → 13 between 1988 and 2004), and total employment held only because urban branch counts rose 43% [2]. That is a specific, fragile mechanism — cheaper branches made more branches worth opening — not a general law. And it has since reversed: the occupation went 502,700 (2016) → 347,400 (2024) with BLS projecting -13% through 2034 [3][4]. The datum universally cited to prove automation does not kill jobs now describes a job that is dying.

The Luddites. The received telling is that they were irrational technophobes disproved by history. The historical work (Binfield's collected Luddite writings; Merchant's Blood in the Machine) holds that they were skilled croppers and framework knitters objecting to specific uses of machinery by employers cutting wages and evading apprenticeship custom, grounding their claims in the Charter of the Company of Framework Knitters and in statutes repealed in 1809 — not to machinery as such. Flagged: I reached this only via search synthesis; every primary and review source was 403-blocked. But note the sharper point, which does not depend on their motives: on the narrow question of whether they personally would be absorbed, the Luddites were empirically correct. The received version gets both the motive and the outcome wrong.

Kodak — explicitly not researched. The founder flagged it as commonly mis-told, and I agree, but Kodak was not among the eight cases and I fetched no source on it. I am not going to assert a corrected version I cannot cite. Flagging it as an open gap rather than filling it with a plausible-sounding retelling.

Re-verifying the founder's two figures

Agriculture: partly wrong, and wrong in an interesting direction. Against the Census/NBER gainful-workers series (Series D 47-61, Historical Statistics of the United States 1789-1945) [9], agriculture was 10,950k of 29,070k gainful workers in 1900 = 37.7%, not 41%. He is about 3 points high. The load-bearing half of his correction holds emphatically — it was nowhere near 80%; the series shows 53.0% as far back as 1870 and falling steadily. The "under 2% today" half holds: USDA ERS puts agriculture at 2% of total US employment in 2017, down from 13% in 1948 [10][11]. The fact he should actually take from this: agricultural employment peaked in absolute terms in 1910 at 11.35M, forty years after its share started falling. That gap is the signature of generational rather than career displacement, and it is more useful to him than the exact 1900 percentage.

Tellers: substantially right, one endpoint unverifiable. "-13% for 2024-2034" is confirmed exactly [3]. "Killed by mobile banking rather than ATMs" is confirmed as BLS's own stated rationale — mobile deposit, online banking and video-teller kiosks, not ATMs [3]. "Declined through the 2010s" is confirmed [3][4]. But "~500k → ~550k between 1980 and 2010" is not established from any primary source I could reach; the specific 1980 and 2010 endpoints circulate in retellings without traceable provenance, and Bessen's own IMF piece contains no national employment-level series at all. What BLS does support is +11.6% growth 1997-2007 and 502,700 in 2016 [4], which suggests the peak was later than 2010, not at it.

The methodological problem, stated plainly

Of eight cases, exactly one has genuine longitudinal individual-level evidence for column (b): switchboard operators [1]. One more (tellers) has a claimed longitudinal analysis whose methodology I could not verify [17]. The other six are cross-sectional job counts plus, in two cases, journalistic anecdote.

This is not a gap in my search. It is the state of the field. Occupational data counts jobs, not humans. When a source says "displaced workers moved into X," it almost always means "occupation A shrank while occupation B grew," which is compatible with the actual humans in A having retired, exited the labor force, or taken worse jobs entirely.

The single most important consequence: Feigenbaum and Gross found the incumbent harm precisely because they built the linkage. Their contribution was a genealogy-based method to follow women across censuses through surname changes at marriage. Before that method existed, the switchboard case looked from the aggregate data exactly like the other seven — an occupation that shrank while the workforce found other work. The harm was always there; the instrument to see it was not. The correct prior is therefore that the other seven cases also contain unmeasured incumbent harm, not that they were benign. Absence of evidence here is a known property of the instrument.

Synthesis for RDCO

The comforting story about AI and data work is skill adjacency: my SQL and modeling judgment transfers to whatever the agent-augmented version of the job is, therefore I am fine. This brief says that story has no empirical support. Skill transfer was highest in journalism and journalism is the worst case in the table. It was moderate for switchboard operators and the incumbents still lost jobs, earnings, and labor-force attachment. The variable that actually separated winners from losers was whether the occupation shrank slower than people naturally leave it, and whether the move to the successor role required a job search or just a reassignment.

That reframes the phData bet in a specific and favorable way. The protective factor in the teller case was not that relationship banking uses teller skills; it was that the same employer redeployed the same person without a market transaction, during an expansion. The founder is currently inside a firm that is doing the redeploying rather than being redeployed, on a cert escalator that is the firm's formal instrument for that redeployment ([[project_phdata_cert_escalator_path]] context). That is the administrative-transition position, and it is the only position in eight cases that reliably produced within-career absorption without a union contract forcing it.

It also sharpens the four-walls spec. "Build-then-TEACH" is not just a preference about how the founder likes to work — on this evidence it is the structurally correct move. In every case in the table, the people who came out ahead of a displacement were not the ones whose skills transferred; they were the ones positioned where the next cohort had to come through them. Teaching is that position. The generational transition is the thing that actually happens, so owning the on-ramp for the generation that arrives after the transition is worth more than defending the skill that preceded it.

The practical instrumentation follows directly. Rather than asking "will AI replace data engineering," which the aggregate cannot answer, compute the annualized rate of change for the specific SOC codes adjacent to the founder's lane and locate them against the rate table above. If the relevant codes are declining slower than natural attrition, the shape of the event is agriculture and nobody gets fired. If they are declining faster, the shape is travel agencies and skill adjacency will not help. That is a number we can pull, not a debate we have to hold.

Last caution, and it applies to us: national aggregates hide local step functions. The switchboard series looked smooth nationally and was a cliff in each city on cutover day. If the AI shock arrives client by client, or team by team, the national data-employment series will look reassuring right up until it happens to a specific person in a specific place. Do not let a smooth aggregate stand as evidence about an individual trajectory — including the founder's.

Why this is in the vault

This is the evidence base for whether the phData cert-escalator bet ([[project_phdata_cert_escalator_path]]) is protection or comfort. It answers a specific decision input: the bet is defensible on administrative-position grounds (inside the firm doing the redeploying), and is not defensible on the skill-adjacency grounds it is usually justified by. It also supplies the rate-vs-attrition test as a reusable instrument for evaluating any future "will X be automated" question, replacing narrative analogy with a computable number.

Open follow-ups

  1. RESEARCH question (answerable from public sources). What is the empirical distribution of annual occupational exit rates — retirements plus voluntary separations — across US occupations, and at what decline rate does measurable incumbent harm begin? BLS publishes occupational separations projections (Table 1.10 of the Employment Projections program) alongside the OES time series, which together should let us calibrate the threshold the entire rate-vs-attrition hypothesis pivots on. This is sharper than the parent question because the parent's answer currently rests on an ordering with no calibrated cut point.
  2. BUILD/TEST task (not a research question — it is a data pull we execute). Compute annualized 2019-2024 employment change from the BLS OES series for the SOC codes adjacent to the founder's lane (15-2051 data scientists, 15-1243 database architects, 15-1242 database administrators, 13-2011 accountants and auditors as a control), and place each against the rate table in this brief. Output is a single chart locating the founder's occupation between the agriculture case and the travel-agency case.

Deliberately not proposed: a broader "more displacement cases" study. The binding constraint is not case count, it is that seven of eight cases lack longitudinal evidence. Adding a ninth cross-sectional case adds nothing.

Related

Sources

Verified and directly fetched:

  1. Feigenbaum, J. & Gross, D., "Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation," NBER WP 28061 (published QJE 139(3), 2024, 1879-1939). Ungated PDF: https://www.nber.org/system/files/working_papers/w28061/w28061.pdf
  2. Bessen, J., "Toil and Technology," IMF Finance & Development, March 2015: https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm
  3. BLS, Occupational Outlook Handbook, Tellers: https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm
  4. BLS, Beyond the Numbers vol. 7, "In the Money: Occupational Projections for the Financial Industry": https://www.bls.gov/opub/btn/volume-7/in-the-money-occupational-projections-for-the-financial-industry.htm
  5. BLS, Occupational Outlook Handbook, Travel Agents: https://www.bls.gov/ooh/sales/travel-agents.htm
  6. BLS OES, NAICS 523000, May 2023: https://www.bls.gov/oes/2023/may/naics4_523000.htm
  7. FRED / BLS CES5552300001: https://fred.stlouisfed.org/series/CES5552300001
  8. Pew Research Center, "U.S. newsroom employment has fallen 26% since 2008" (2021): https://www.pewresearch.org/short-reads/2021/07/13/u-s-newsroom-employment-has-fallen-26-since-2008/
  9. US Census Bureau, Historical Statistics of the United States 1789-1945, Ch. D, Series D 47-61, "Labor Force — Industrial Distribution of Gainful Workers (NBER): 1820 to 1940": https://www2.census.gov/library/publications/1949/compendia/hist_stats_1789-1945/hist_stats_1789-1945-chD.pdf
  10. USDA ERS, Farm Labor: https://www.ers.usda.gov/topics/farm-economy/farm-labor
  11. USDA ERS, Amber Waves, Feb 2022: https://www.ers.usda.gov/amber-waves/2022/february/increases-in-labor-quality-contributed-to-growth-in-u-s-agricultural-output
  12. The Conversation, "What today's labor leaders can learn from the explosive rise and quick fall of the typesetters' union": https://theconversation.com/what-todays-labor-leaders-can-learn-from-the-explosive-rise-and-quick-fall-of-the-typesetters-union-214519
  13. UPI Archives, ITU merger vote, 26 Nov 1986: https://www.upi.com/Archives/1986/11/26/Members-of-the-134-year-old-International-Typographical-Union-voted-overwhelmingly/9757533365200/
  14. Presshaus LA, "1978: the last NY Times paper printed with Linotype machines": https://presshaus.la/journal/2016/9/20/1978-the-last-ny-times-paper-printed-with-a-linotype-machines
  15. NYSE Trading Floor Broker Directory (primary artifact, ~23 contacts / ~20 firms): https://www.nyse.com/publicdocs/nyse/NYSE_Trading_Floor_Broker_Directory.pdf
  16. Otherforms, "Typography, automation and the division of labor" (1964 ITU automation fund): https://otherforms.net/typography-automation-and-the-division-of-labor/

Secondary or unverified — flagged in-text, do not cite downstream without re-verification:

  1. Report Earth, "The lost newspaper jobs of 2024" (cites BLS CES NAICS 51111; the underlying BLS page was 403-blocked): https://reportearth.substack.com/p/the-lost-newspaper-jobs-of-2024-and
  2. Burning Glass Institute, "The case of the vanishing teller" (claims a longitudinal "Worker Career Histories" analysis; methodology not verified): https://www.burningglassinstitute.org/bginsights/the-case-of-the-vanishing-teller-how-bankings-entry-level-jobs-are-transforming
  3. Bessen, J., "Was Mechanization De-Skilling? The Origins of Task-Biased Technical Change," SSRN — abstract only, full text CAPTCHA-gated on every mirror attempted: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1789688

Blocked or paywalled, attempted and abandoned per protocol (no retries): SSRN full texts (CAPTCHA), bls.gov direct WebFetch (403; routed via reader proxy where noted), web.archive.org (unreachable from this toolchain), the June 2016 NYT feature on the 1974 ITU automation agreement (no live URL found), CEPR/VoxEU (403), Taylor & Francis handloom-weaver counts (403), LA Review of Books on Binfield (403), NBER chapter c1567 and WP w24235 (fetched but returned unparseable binary).