References: Will AI take my job?

Source material

This lesson is an original adaptation. Its durable spine, the task-based way of thinking about jobs (a job is a bundle of tasks; AI acts on tasks, not titles; a task is exposed when AI can reach it to automate or augment), comes from the future-of-work session of a Harvard Kennedy School course. The current statistics, the fresh 2026 labor evidence, the hedges, and the prose are our own. One framing carries through everything below and is the point of the whole lesson: exposed does not mean doomed, it means reachable, and reachable can be a gift as easily as a threat. Clawdemy is independent of Harvard, which has not reviewed or endorsed this track.

• The Science and Implications of Generative AI (HKS DPI-681M), Harvard Kennedy School, Spring 2024. Faculty: Sharad Goel, Dan Levy, and Teddy Svoronos. This lesson adapts Class 10, the future of work from the Spring 2024 course site, whose content is licensed under Creative Commons Attribution 4.0. Class 10 is the course’s future-of-work session: it breaks jobs into tasks, defines exposure as subject to automation or augmentation, asks whether we will run out of work, and weighs who gains and who loses. Its session videos are the source for our framework.
• Official course lecture playlist on YouTube, Harvard Kennedy School. The full lectures, free to watch. We mean it when we encourage you to take the original course alongside this track: good teachers deserve more students.
• Provenance note: the framework and classroom examples in this lesson were drawn from the transcripts of the official Class 10 lecture videos in the playlist above, obtained from the official Harvard sources only. No third-party re-uploads or mirrors were used.

A note on how these figures are cited

Freshness is the whole discipline here. This field moves fast, so every dated figure below is re-verified live before publish and on each freshness sweep. The framework is durable, but the numbers are not, and none is presented as a settled fact about the present. The 2023 Goldman figure is a dated forecast, always. The 2026 labor evidence is early and genuinely contested, always, in both directions: never “AI is destroying jobs” and never “AI has had no effect.” And the junior versus senior split in particular is carried only alongside its own contestation, never as a settled pattern. Exposure, throughout, means a task reachable by AI in either direction, never job loss.

The 2023 Goldman Sachs forecast, dated

• Goldman Sachs, “The Potentially Large Effects of Artificial Intelligence on Economic Growth” (Briggs, Kodnani, Hatzius, and colleagues), published late March 2023. The report estimated that shifts in workflows could expose the equivalent of about 300 million full-time jobs to automation, and that broad AI adoption could eventually raise annual global output by around 7 percent over roughly a decade. Discipline that must survive: this is a projection made in 2023 about the decade ahead, not a measurement of anything that has happened. Exposure is not job loss; the report stressed that most exposed work would be complemented rather than substituted, and the authors did not fold these estimates into their own baseline forecast given the uncertainty. No derivative reports 300 million as jobs lost or 7 percent as a current fact. Sources: Goldman Sachs Insights summary.

The 2026 labor evidence, early and contested

• Stanford Digital Economy Lab, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” (Brynjolfsson, Chandar, Chen), originally August 2025, updated February 2026. Using United States payroll records, it reports that early-career workers in the most AI-exposed occupations saw a marked relative decline in employment since late 2022, while older, more experienced workers in the same occupations were stable or growing, with declines concentrated where AI automates rather than augments. Limitations that must travel with it: this is one study and it is contested. The lesson reads it through the framework as the framework predicting, not as settled causation, and it carries the contestation in the same breath rather than letting the junior versus senior split stand alone (see the next two entries). The body says “a marked drop” rather than reciting a percentage, to keep the number from hardening into fact. Sources: Stanford Digital Economy Lab publication page.
• Federal Reserve Bank of New York, Liberty Street Economics (May 2026). Analysis of job postings finds that the divergence in AI-exposed occupations began before the arrival of widely used AI writing tools, so AI is not the main driver of the recent hiring slowdown, and finds no divergence between junior and senior positions within highly exposed occupations. This is the singular source, kept singular, for the rebuttal to the junior versus senior split: another 2026 analysis of the same jobs finds no such divide between younger and older workers. Sources: Liberty Street Economics.
• The Canaries authors’ own February 2026 update softened the causal timing of the early decline, tracing part of the late-2022 and 2023 drop to other, non-AI factors, rather than to AI alone; the AI-linked effect becomes clearer only later. This is why the lesson says the study’s own authors now trace part of the early drop to other forces, and never lets the early window read as settled AI causation.
• Humlum and Vestergaard (2025), Denmark. A detailed study of Danish payroll and survey data found effects on earnings and hours near zero. Carried as part of the balancing picture: the field is early, and no single study, including the ones showing large effects, should stand as the settled answer.
• Peterson Institute for International Economics (PIIE), 2026. A review argued that research on AI and the labor market is still in its early innings and that current evidence is inconclusive, with claims of harm to particular groups premature. Paraphrased here; carried as a marker of how early the evidence is. Sources: PIIE realtime economics.
• Yale Budget Lab (May 2026). Analysis concluded that AI is probably not yet the reason for recent labor-market weakening. Paraphrased; carried in the balancing cluster. Sources: Yale Budget Lab research.

The exposure comparison the course cites

• World Economic Forum analysis (as cited by the course). The course’s reflection question, which is more exposed to AI, a software developer or a human resources manager, resolves toward the software developer, because the potential for both automation and augmentation is greater there while human resources leans on judgment and interpersonal work today’s tools reach less easily. The source is characterized generically as a World Economic Forum analysis; no dated product specifics are carried forward. This is the course’s own worked reflection, kept and updated.

The course’s own worked examples

• The bank-teller and telephone-operator history is the course’s illustration, drawing on labor-economics work (including David Autor’s), that automation can raise employment in a job rather than end it: after automatic teller machines spread, the number of bank tellers rose for years as cheaper branches multiplied, while telephone operators, by contrast, faded away. Carried as a durable historical pattern, not a dated claim.
• David Autor’s three forces for why new work keeps appearing (insatiable demand as prices fall, augmentation of remaining human work, and the invention of wholly new work) are the course’s frame for why the whole economy is unlikely to run out of jobs. Autor is named as the course names him.
• William Baumol’s cost disease and Robert Solow’s productivity paradox are the two economic ideas the course uses to explain why the boom may arrive slower than expected. Both are delivered as intuitive stories and fully paraphrased; neither is rendered in quotation marks, because a spoken lecture transcript is not an authoritative source for a verbatim line.
• Acemoglu and Johnson, “Power and Progress” (2023), grounds the lesson’s closing note that shared prosperity is not guaranteed by technological progress, and has historically emerged only when societies steered the direction of technology and the sharing of its gains toward the many. This is carried institutionally, as a contested economic idea and not an endorsement of any policy, with names kept generic in the body; the lesson stays clear of partisan content entirely.

On this site

• Should AI do this task?, lesson 4 of this track. The judgment lens for whether a task should be handed to AI at all; this lesson turns that same lens on your own working life.
• The risk map, lesson 6 of this track. It puts the future of work on the four-category risk map as one square big enough to earn its own lesson; this lesson is that lesson.
• How models are pretrained, from Clawdemy’s AI Foundations track. Why language is the frontier these systems reach most easily: underneath, they are prediction engines trained on oceans of human text. This lesson needs only that fact; that lesson covers how.

Going deeper

• Official HKS course listing for DPI-681M, Harvard Kennedy School. The in-person course behind the open online materials this track adapts.

Adjacent topics

• The next lesson turns from what AI does to your work to what AI does to what you can trust. The same technology that drafts your emails can also flood the world with convincing false words, images, and voices, at a scale we have never seen. It is where the track lands, because the point was never fear. It was standing on solid ground.