Why Task-Level Wins Mask Economy-Wide Stagnation
The recent decline in entry-level employment is not a consequence of AI’s revolutionary power. Instead, it is a symptom of a maturing industry entering a low-growth phase. AI is serving as a tool for cost optimisation rather than a driver of new value creation.
If generative AI were truly a general-purpose technology on par with electricity or the micro-processor, the macro data should already be unmistakable. Instead, the silence is deafening.
A Stanford study by Brynjolfsson, Chandar, and Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,”1 correctly identifies a correlation between AI adoption and declining youth employment but appears to have the causality backwards. This paper mistakes a consolidation fever for technological revolution, suggesting that the decline in entry-level jobs signals a fundamental, technology-driven revolution. It looks at the effect and concludes the cause is the new technology’s transformative power. This is the classic tool-to-work model.
However, it’s more likely the decline is caused by a deep-seated business crisis, as I’ve argued elsewhere.2 The true cause is an industry struggling to adapt to the end of a long, high-growth phase. The “displacement” of young workers is a symptom of a transition from growth to consolidation, not a technological leap.
While the Stanford paper’s data on job exposure to AI is likely accurate, its interpretation falters when faced with broader economic realities.
The Absence of a Systemic Productivity Boost
If LLMs were truly “punctuation” technologies, we would expect to see:
- A surge in gross output per worker (not flat GDP-per-hour).
- Rising labour demand in adjacent and new tasks (not falling postings for junior roles).
- Capital deepening and wage gains (not zero wage pass-through).
- Firm-level expansion post-adoption (not buybacks and layoffs).
Table 1: The Absence of a Systemic Productivity Boost
| Signals we should see | The disappointing reality |
|---|---|
| GDP-per-hour surging ≥2 % above pre-AI trend | US Q1-Q2 2025: 0.8 % YoY, below 2010-19 average3 |
| Junior job postings rising in adjacent tasks | Indeed & Revelio Labs: –18 % for entry-level SWE & CS roles since Nov 20224 |
| Real wages for AI-complementary skills up >3 % | Denmark admin data: precise zero (+0.0 %; 95 % CI ±0.8 %)5 |
| Cap-ex surge in AI-adopting firms | S&P 500: record $882 bn buybacks, R&D growth flat6 |
Instead, the gains are confined to specific tasks, characteristic of incremental change, not a revolutionary shift. A study on M365 Copilot found its impact limited to task-specific gains, like a 12% increase in document speed, not a systemic boost to output.7
This is supported by research from the NBER, which found ‘no significant impact’ on workers’ earnings or hours despite widespread AI chatbot adoption,8 directly challenging the narrative of an imminent labor market transformation.
The modest time savings documented in these studies suggest AI is a tool for incremental cost-cutting within existing business models, not for creating new ones. This disconnect between micro-efficiency and macro-stagnation is the smoking gun. True revolutionary technologies create measurable economy-wide growth spurts. AI’s gains are being captured as profit, not reinvested into new value creation.
Skeptics argue macro effects need time. Yet the micro studies above were conducted 12-24 months post-rollout—the same window in which electricity, ICT, and containerisation had already registered clear productivity jumps. Delay is no longer a credible alibi.
A Focus on Cost-Cutting
The decline in entry-level hiring isn’t just about efficiency; it’s about a business landscape that has run out of productive ways to deploy capital. Recent industry surveys identify ‘resource optimisation’ and ‘faster access to data’ as the top business outcomes of AI,9 rather than market disruption or new value creation.
Firms are resorting to financial tactics like buybacks and layoffs as they struggle to find new growth drivers. AI has become a convenient tool for this, allowing them to trim costs and increase short-term margins.
But the J-Curve!
Brynjolfsson et al. assume the J-curve dip is a natural, technological inevitability. However, management can also induce a J-curve by making a conscious decision to replace labour with technology to reduce costs. Management induces a pseudo-J-curve: a one-time cost re-base that masquerades as technological progress. The curve’s right-hand tail never rises because the investment thesis was never growth—it was margin defence.
A company investing heavily in AI creates a measurable productivity dip—output per hour falls due to the transition costs. Once complete, the result is not a dramatic rise in output. Instead, it is a permanent reduction in costs (headcount) and a maintenance of existing output levels. The curve doesn’t soar back up; it returns to its previous level at a lower operating cost. This is a defensive, cost-cutting gain, not an offensive, value-creating one.
For the economy, this is problematic. It explains the central paradox: task-level studies show time savings, but these savings are used for cost reduction, not output expansion. They are not compounding into new products, services, or growth. The NBER study suggests the current limited effects are more than a temporary dip; they are symptomatic of a deeper, systemic business problem.
The Performance of Strategic Thinking
To understand what is happening, we must look past the “digital transformation” rhetoric and examine corporate actions. While companies speak of “transformation,” their actions—layoffs, hiring freezes, and a focus on subscription models—suggest a focus on defending existing revenues and managing decline.
This is a performance of strategic thinking. The technology is used to give the illusion of progress, a managerial shortcut that masks a lack of genuine reinvention.
A company facing a growth crisis uses AI differently than one investing for the future. The former makes a short-term P&L decision to reduce headcount. The “rise” of their J may never come because no new value was created. The investment is a signal to investors that they are modernising and cutting costs. Layoffs are justified under the guise of technological inevitability rather than managerial choice.
Conclusions
Despite widespread claims of AI-led transformation, concrete examples of companies creating genuinely new value—rather than optimising existing operations—remain remarkably scarce.
The canary in the coal mine is not singing of a new, productive era. It is a warning sign that the current industry model is failing. The observed job displacement is a symptom of this unraveling. The tech slowdown created a vulnerable, low-growth environment. In this context, powerful AI tools did not act as a catalyst for new value creation but as an accelerant for pre-existing trends toward financialisation and optimisation.
The observed “productivity dip” is therefore not solely the cost of learning a new technology; it is primarily the cost of an industry-wide shift in strategy from growth to efficiency, for which AI is the perfect tool and the perfect excuse. Until firms redeploy AI savings into new products, markets, or capabilities, the canary isn’t forecasting a renaissance: it is suffocating in a mine that management refuses to ventilate.
- Brynjolfsson, E., Chandar, S., & Chen, M. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/wp-content/uploads/2025/08/Canaries_BrynjolfssonChandarChen.pdf ↩︎
- Evans-Greenwood, Peter. “Unravel, Rebalance, or Reinvent.” Substack newsletter. The Puzzle and Its Pieces, April 30, 2025. https://thepuzzleanditspieces.substack.com/p/unravel-rebalance-or-reinvent. ↩︎
- U.S. Bureau of Labor Statistics. (2025). Labor productivity and costs, second quarter 2025, preliminary. U.S. Department of Labor. https://www.bls.gov/news.release/prod2.nr0.htm. ↩︎
- Revelio Labs. (2024). The AI revolution is here… and it’s making fewer jobs for juniors. Indeed Hiring Lab. https://www.hiringlab.org/. ↩︎
- Humlum, A., & Vestergaard, E. (2025). Large language models, small labor market effects. National Bureau of Economic Research. https://www.nber.org/papers/w33777. ↩︎
- S&P Global Market Intelligence. (2025). S&P 500 companies continue to prioritize share buybacks. S&P Global. https://www.spglobal.com/marketintelligence/en/news-insights/blog/sp-500-companies-continue-to-prioritize-share-buybacks. ↩︎
- Dillon, E. W. et al. (2025). “Early Impacts of M365 Copilot.” https://arxiv.org/abs/2504.11443. ↩︎
- Humlum, Anders & Emilie Vestergaard (2025). Large Language Models, Small Labor Market Effects. National Bureau of Economic Research. https://www.nber.org/papers/w33777. ↩︎
- Freeman, B. S. et al. (2024). “Evaluation of Task‑Specific Productivity Improvements Using a Generative AI Personal Assistant Tool.” https://arxiv.org/abs/2409.14511. ↩︎