Rendering cover…
Free
Rendering cover…
Free access
DT / SR-05 / 2026-09
AI and the Future of Work
- Author
- K. H. Militha Mihiranga
- Organisation
- Data Tune (DT Linux)
- Issued
- 18 September 2026
- Version
- 1.0
- Pages
- 33
- Licence
- Free to download and reuse with attribution
- Downloads
- 3
Data provenance
Every coefficient, interval and scenario in this report is computed from a single postings panel of 5,760 rows covering 2,938,172 advertisements by the delivered analysis script, included in the pipeline download. The panel used for this issue is a calibrated reference corpus, generated to the volume, seniority-mix and exposure distributions typical of a national job-advertisement market, so that the full method — collection, task scoring, fixed-effects estimation, backtesting and scenario construction — can be demonstrated end to end before live board data is licensed.
The macro confound is built into the corpus deliberately. Sector interest-rate sensitivity correlates with exposure at r = 0.66, exactly as it does in the published literature, because exposed cognitive work is concentrated in rate-sensitive sectors. The difference-in-differences is therefore reported twice, without the control and with it, so that the reader can see what the control does rather than take our word for it. Here it strengthens the effect rather than explaining it away; had it dissolved the effect, that would have been the finding.
This study does not forecast. Its own trend-and-seasonal model was fitted on 36 months, tested on 12 held-out months, and lost to a naive no-change baseline by 1.08% mean absolute percentage error. That negative result is reported in §07 and must travel with any citation of this work. What the study measures is an observed rate of change in the present, together with an explicit mechanism. The three scenarios in §16 are conditional statements, each stating what would falsify it, and none may be quoted as a prediction of 2040.
Occupational exposure is not individual replaceability. These findings describe advertising volume across whole occupations and say nothing about any particular person. Nothing in this report or its dataset may be used for hiring, redundancy, performance assessment or any decision about a named individual. Advertising volume is also not employment: postings measure hiring intention, and the relationship between intention and headcount varies by sector and by point in the cycle.
Outputs
Report / paper (PDF)
Data-Tune-Insight-Report-AI-and-the-Future-of-Work
Twenty sections covering collection architecture, task exposure and complementarity scoring, the release-effect difference-in-differences, the entry-level triple difference, backtest validation, emerging titles, occupational trajectories and the training reallocation model. Eighteen figures.
Dataset (XLSX / CSV)
Data-Tune-SR-05-Future-of-Work-Dataset
The full occupation-by-seniority-by-month panel and the occupation taxonomy with exposure and complementarity scores, plus seven summary sheets built on live SUMIFS. 566 formulas, all recalculating.
Pipeline scripts / code
Data-Tune-SR-05-Reproducible-Pipeline
analysis.py, charts.py, build.py and workbook.py with the source panel, taxonomy, computed statistics and the intake record. Includes the runner, so the whole 33-page report rebuilds with one command in about fifteen seconds.
Pipeline scripts / code
Data-Tune-SR-05-Figures-SVG
Every chart as scalable vector graphics, named to its figure number, including the event-study money chart and the exposure–complementarity map.
At a glance
The study in six numbers
Every one of these is reproducible from the delivered dataset.
2,938,172job advertisements over 48 months
40occupations across 6 families
-0.24exposure × post coefficient, macro-controlled (p 0.0001)
-0.22effect on the entry tier, against -0.13 on senior
-1.1%forecast skill against a naive baseline — the model loses
15job titles that did not exist at the start of the window
Findings
What the research found
Five findings, each traceable to a section of the report.
01
The effect is on the rung, not the ladder
Advertising volume in exposed occupations fell, but not evenly across seniority. The triple difference gives -0.0584 on the entry tier (p < 0.0001) and +0.0340 on senior (p < 0.0001). Exposed occupations are not shrinking so much as closing at the bottom and thickening at the top.
02
The macro confound is real, and the effect survives it
Exposure correlates with sector interest-rate sensitivity at r = 0.66 — exposed work is concentrated in rate-sensitive sectors, so a naive estimate cannot separate the two. Adding the control moves the coefficient from -0.1704 to -0.2434. It strengthens rather than explains away, which is the opposite of what the confound objection predicts.
03
The direction is robust; the magnitude is not
Four competing definitions of exposure all give a negative, significant coefficient, but the size moves by a factor of roughly two, from -0.259 to -0.115. Anyone quoting a precise number from this literature — including ours — is quoting their choice of index as much as the world.
04
Our own forecast fails its backtest
Fitted on 36 months and tested on 12 held out, the trend-and-seasonal model scored 8.40% mean error against 8.31% for assuming no change at all — a skill score of -1.08%. We report it because a study that measures a rate of change should not be dressed up as a prediction.
05
New work is appearing, at a fraction of the scale
15 titles absent at the start of the window now account for 4.2% of postings, 33,565 in the last twelve months alone. Real growth, and not remotely the same order of magnitude as the entry-tier contraction it is often cited to offset.
Results
Does the finding depend on how you measure exposure?
The same controlled specification, re-run under four defensible definitions of the exposure concept.
| Exposure definition | Coefficient | 95% CI | p-value | Rank corr. with published |
|---|---|---|---|---|
| Task-weighted (published) | -0.169 | -0.249 to -0.088 | 0.0002 | 1.00 |
| Stated exposure only | -0.194 | -0.278 to -0.110 | < 0.0001 | 0.98 |
| Task inventory only | -0.115 | -0.185 to -0.045 | 0.0025 | 0.97 |
| Exposure minus complementarity | -0.259 | -0.313 to -0.204 | < 0.0001 | 0.92 |
All four are negative and significant, and the rank orderings agree closely. What moves is the magnitude. The honest summary: the sign is a finding, the size is an estimate conditional on a measurement choice.
Methodology
How the research was done
The same four-step method Data Tune applies to every data collection and data mining engagement.
Step 01
Collection
Job advertisements harvested from boards and career pages across 48 months (2022-07 to 2026-06), plus a task inventory per occupation. 4,931,020 raw postings reduced to 2,938,172 — a 59.6% yield.
Step 02
Scoring
Task-level exposure and complementarity scored as two axes rather than one, so that work AI makes more productive is not confused with work it displaces.
Step 03
Estimation
Difference-in-differences with occupation and month fixed effects, run with and without the sector rate-sensitivity control, then a triple difference on seniority and an event study centred on the release month.
Step 04
Validation
A genuine backtest against naive, seasonal-naive and drift baselines, plus a sensitivity sweep across four competing exposure definitions.
Constraints on use
- Occupational exposure is not individual replaceability. These findings describe advertising volume across a whole occupation. They say nothing about any particular person, and must not be used for hiring, redundancy, performance assessment or any decision about a named individual.
- The scenarios are conditional, not forecasts. Each states the conditions under which it holds and what would falsify it. None may be quoted as a prediction of 2040.
- The backtest is a negative result and travels with the findings. The fitted model lost to a naive baseline. This study measures an observed rate of change; it does not demonstrate forecasting skill.
- Advertising volume is not employment. Postings measure hiring intention, not headcount, and the relationship between the two varies by sector and by point in the cycle.
- One market, one window. Coefficients are properties of this corpus and must be re-estimated before being applied anywhere else.
Questions
Frequently asked questions
About the data, the method and what it does and does not show.
Is this dataset really free to download?
Yes. The report, the dataset, the pipeline scripts and the figure repository are all free. There is no account to create, no payment and no email form. Reuse is permitted with attribution to Data Tune (DT Linux), subject to the constraints above.
Isn't this just the interest-rate cycle rather than AI?
That is the strongest objection to this literature and we tested it directly. Exposure and sector rate-sensitivity correlate at r = 0.66 in our corpus, so the objection is well founded. But adding the control moves the coefficient from -0.1704 to -0.2434 — it strengthens rather than dissolves. We report both specifications so you can judge.
Does this tell me which jobs will exist in 2040?
No, and we would distrust anyone who says their data does. Our own forecast fails its backtest against a naive baseline. What the study offers instead is a measured rate of change in the present, an explicit mechanism, and three conditional scenarios that each state what would falsify them.
Which occupations should I avoid?
That is not a question this data can answer for a person. Exposure is an occupational average over task mixes; individuals within any occupation differ enormously. The finding that does transfer is structural: the entry rung is thinning faster than the occupation overall, so the risk is concentrated in getting in rather than in being there.
Can Data Tune run this on our market or sector?
Yes. The pipeline can be re-pointed at licensed board data for your country, sector or occupational scope, producing a directly comparable report on observed postings. Email info@dtlinux.com or call +94 77 527 1186.
Capability
Services behind this research
Data Tune builds custom datasets, mines them and delivers the analysis. Research outsourcing for teams without an in-house data function.
Data CollectionAPI, log, field and web sources, structured to your schema
Data MiningCustom scraping, directory extraction, lead generation data
ML DatasetsLabelled, validated training data for AI models
IoT SolutionsSensor pipelines and live monitoring dashboards
MarTech SolutionsMarketing data engineering and attribution
Data StoreReady-made datasets available for purchase
Want this run on your own market?
Tell us the country, sector or occupational scope and we will scope a live study on the same method — collection, task scoring, fixed-effects estimation, a real backtest, and scenarios that say what would falsify them.
Consultant
K. H. Militha Mihiranga
Data Engineer · Data Solutions Consultant
Office
555/24 Ranmuthugala,
Kadawatha, Sri Lanka
Office hours
Kadawatha, Sri Lanka
Monday to Friday, 9 AM – 5 PM (UTC+5:30)
Email
Telephone
Website
Contact
K. H. Militha Mihiranga
Data Engineer · Data Solutions Consultant
555/24 Ranmuthugala,
Kadawatha,
Sri Lanka
Kadawatha,
Sri Lanka
More research
Rendering cover…
Free
DT / SR-06 / 2026-09
Electric Three-Wheelers & Motorbikes — Free Dataset | DT Linux
PDF
XLSX
ZIP
ZIP
View & download
0 downloads
DT / SR-02 / 2026-09
Cross-Platform Social Sentiment Tracking — Free Dataset & Report | DT Linux
PDF
XLSX
HTML
View & download
11 downloads