AI & Human Capital Bits

Who will hold the new jobs of the AI future? Technological waves both wipe out traditional roles and create new jobs that didn’t exist before. In What Makes New Work Different from More Work? David Autor and co-authors show that in the United States such new jobs have been largely held by the young and those with more schooling—and pay a sustained wage premium. Assuming this pattern continues to hold, investments in education will continue to be critical to prepare people for the AI era.

May is Mental Health Awareness Month, and mental health promotion, prevention, care, and recovery support are critical components of Universal Health Coverage (UHC). On leveraging AI for mental health support: a recently published RCT tested the Kai.ai chat agent and found significant drops in anxiety and depression. Kai operates with a human in the loop model, which includes a trigger function to enable immediate human intervention.

This new NBER paper tested whether personalized recommendations and collaborative filtering could increase interaction and use of an English-learning app for children. The pilot RCT engaged 7,750 young learners in India. Those with personalized recommendations boosted engagement by 60%. When scaled app-wide, the five-month impact was 2.5 times larger than in the original trial; a striking result in a field where scaled-up interventions routinely underperform. The paper suggests that algorithmic personalization can defy the usual “voltage drop”. Yet the gains were front-loaded: returns diminished as students exhausted the best content in the app, indicating that content availability rather than algorithmic sophistication becomes the binding constraint.

Stanford released the 2026 Artificial Intelligence Index Report; 423 pages of independently sourced data on AI's trajectory. The headline from co-chairs Yolanda Gil and Raymond Perrault sets the tone: "The data does not point in a single direction. It reveals a field that is scaling faster than the systems around it can adapt." AI experts and the public are very far apart on whether AI will help people do their jobs: 73% of experts say yes, only 23% of the public agrees. Consider the report required background reading for anyone shaping AI-related policy or operational work.

If AI alone defaults to the easiest methods, the real value is augmentation, not automation. Two economists are showing what that looks like. Chris Blattman launched claudeblattman.com — free AI workflows for research and project management, no code required. Scott Cunningham’s Claude Code series (40 installments) documents using AI agents to learn econometrics, replicate papers, and audit results. The tools won’t do the thinking. They free up time for it.

Should students control when AI helps them? Probably not, according to a new Wharton field experiment with over 200 chess club students. Students were randomly assigned to either a self-regulated condition, where they could request help at any time by clicking a button, or a system-regulated condition, where the platform automatically provided AI tips at key moments. The result? Those with self-regulated on-demand help trained less, felt less accomplished, and quietly became AI-dependent, even knowing it was happening—ultimately achieving much lower performance gains. For human capital development, the implication seems to be that smarter guardrails matter as much as smarter tools.

A study published in Nature Health evaluated an LLM-based clinical decision support system deployed across 16 primary care clinics in Kenya. The AI advice was largely sound: management guidance aligned with local guidelines in 99% of cases, hallucinations were rare (3.4%), and the per-consultation LLM inference costs were low. But what stands out is that the bigger challenge remains the human-AI interaction. Clinicians rarely changed decisions after reviewing AI output, and when they did act on AI advice, harmful suggestions were more likely to be followed than beneficial ones. The work ahead is not just building better models. It is designing systems clinicians trust enough to use when it matters, even while costs continue to go down.

Our analysis of recent Anthropic data shows that a quarter of Claude use in MICs is associated with teaching (and half is associated with ICT). This matches other data suggesting AI use by teachers in developing countries is already widespread. Also, across all countries, Claude usage by occupation correlates with measures of “AI exposure” with one major exception: managers have high exposure but modest adoption.

In Togo, when COVID-19 hit and no social registry existed, machine learning trained on mobile phone data helped identify the poorest households in weeks rather than months. The approach improved targeting precision by 42% compared to geographic methods alone. But here’s the nuance: the researchers found traditional registries still outperform phone-based targeting when time permits. Meanwhile, an OECD review warns of algorithmic bias, lack of transparency, and the risk of reinforcing structural inequalities. Importantly, only 40% of citizens in OECD countries welcome AI-based benefit processing.

Anthropic CEO Dario Amodei's essay The Adolescence of Technology wrestles with the AI transition period where we gain transformative power of AI before developing the maturity to wield it safely. AI’s rapid scaling could reshape economic structures, state capacity, and global security far faster than existing institutions can adapt. The essay also turns to economic & job disruption and subsequent inequality challenges. In the company’s blog series, you will also find an informative story on tests for technical candidates in Anthropic’s engineering team, wrestling with how to identify human candidates to work on problems that their coding model becomes increasingly better at solving.

AI is being used in at least 38 cases in low-and-middle income countries to support community health workers. A pathbreaking study by World Bank colleague Anja Sautmann and co-authors tests whether a LLM in Nigeria helps them provide better care. They find the tool improved the quality of clinical notes and was valued by the workers. But measured quality of care, as assessed by on-site doctors, did not. We still have a lot to learn about how to make these tools effective in low-resource settings. [The study was presented at the Knowledge Symposium on AI and Human Capital in the Global South.]

Transformative AI is racing ahead, raising prosperity and risk. Volume 2 of ‘The Digitalist Papers’ is now out and maps the terrain: global governance, fair competition, worker safety nets, fiscal resilience, and education reform. Treat AI as a public good, experiment responsibly, and prepare people for new roles. Policy, not hype, will shape equitable outcomes.

The World Bank’s Jose Cuesta breaks this down in a just-published article with an input-output framework where he shows how AI serves as a mediating technology between biased inputs and societal predictions which lead to gender biases emerging in AI systems. He considers why AI can exhibit such bias, its consequences, and both technical and policy solutions.

Stanford’s Erik Brynjolfsson and co-authors lay out nine Grand Challenges for policymakers as AI accelerates from task-automation to truly transformative systems. They warn that while AI could multiply growth rates several-fold, without new labor laws, safety nets, and global governance frameworks, it risks fueling inequality, concentration of power, and instability.

Anthropic’s Economic Index report came out last week and shows that Claude AI usage at the country level is highly correlated with GDP per capita (see below). A separate new paper from “How People Use ChatGPT” confirms that while ChatGPT usage rates have grown rapidly around the world, they remain very low in the poorest 30 percent of countries.

Ludwig et al. (2024) show that simple, off-the-shelf algorithms applied to real-world programs—like pre-trial release and patient test referrals — can yield increased social welfare while and reduce income losses for policy makers. And Kaiser et al. (2024) finds that fine-tuned LLMs can accurately predict the direction of outcomes in 80% of empirical studies on dietary interventions — like food choices, sales, or waste.

With the help of LLMs, could you test fewer ideas and use those insights to drive smarter policymaking?

An MIT-led trial looked into the brain activity of 54 students assigned to write an essay. It found that those who let GPT-4o compose their essays from the start showed the weakest neural engagement compared to those that used a standard search engine without AI as help, or those that used only their brains. Meanwhile, with 55% of UK households already dabbling in generative AI, the Alan Turing Institute urges robust AI-literacy programs for children and guidance for parents to mitigate the risks of generative AI and make AI a tool that benefits young people.

AI is knocking on the clinic door, with high stakes and hopes for solutions. Stanford is spearheading over 50 projects, with pragmatic tools like ChatEHR in an initial pilot test while others remain under the impact microscope. Meanwhile, in Kenya Penda Health’s clinical AI is now being evaluated at scale. It could be one of the first randomized trials to provide evidence for L/MICs.

ILO has updated its AI occupational exposure index methodology (from 2023), as shown in the figure below, the new global estimates confirm that clerical occupations have the highest exposure levels and add some strongly digitized occupations. Overall, women are more likely to have greater exposure to AI. The positive bit, it is more about transformation of jobs than full replacement of workers. For the full interactive experience, click here.

World Bank staff are advancing the knowledge frontier on AI for education. One Bank team just published a working paper from a randomized control trial which shows substantial learning gains in English instruction from use of AI for in-school tutoring in Nigeria. And Microsoft CEO Satya Nadella featured another Bank pilot in his keynote address at a recent event. This one uses AI to assist teachers in Lima, Peru. See the writeup here. A formal evaluation is ongoing.