The research team, led by Rosemary Emegu Isoto of Makerere University together with colleagues including Joachim von Braun, framed child labour as one of the hidden social costs of agricultural production. While environmental externalities such as land degradation and greenhouse gas emissions are increasingly accounted for in debates about sustainable food systems, the social costs borne by farming families remain largely invisible in market prices. When households face low producer prices, labour shortages, and intense seasonal demands, they may turn to family labour, including children, to keep production going. The study set out to measure not just whether children work, but whether the work they perform is the kind that can harm their health, safety, and education.

The scale of the global problem provides important context. According to recent estimates from the International Labour Organization and UNICEF, 138 million children worldwide, about 7.8 percent of all children, remain engaged in child labour despite the Sustainable Development Goal target of eliminating the practice by 2025. Agriculture is the dominant sector, accounting for more than 60 percent of child labour among children aged 5 to 17, and the share is highest among the youngest children. Sub-Saharan Africa is home to two-thirds of all children in child labour, with a regional prevalence of roughly 21.5 percent. Uganda’s own rate, estimated at 28.2 percent in the 2019/20 national household survey, sits above the regional average even as rates decline across the continent.

What distinguishes this study is its methodological precision. Rather than treating all children involved in coffee production as child labourers, as many earlier studies did, the researchers built a coffee-specific classification of hazardous activities grounded in ILO Convention No. 182 and Uganda’s Employment (Employment of Children) Regulations of 2012. They grouped hazardous exposures into three categories: physical and mechanical hazards such as sharp tools, heavy loads, and land clearing with fire; environmental and chemical hazards including pesticide spraying, re-entering sprayed fields within twelve hours, dust, noise, and extreme heat; and working condition and time-related risks such as night work, excessively long hours, and working more than two hours on a school day. This granular approach allowed the team to separate children’s general participation in farm work from the specific conditions that make work harmful.

The data were collected face-to-face over twenty days in October 2024 using computer-assisted personal interviewing, covering 1,252 children aged 5 to 17 across households in fifteen districts spanning the Tooro, Elgon, Southwest, Buganda South, and Buganda North regions. Overall, 42 percent of children had participated in at least one coffee-related activity during the previous two agricultural seasons, with harvesting the most common task, involving between 77 and 86 percent of working children depending on the season, followed by weeding at around 54 percent. Among the 528 children who had worked in the seven days before the survey, 23 percent had been exposed to hazardous activities in the previous twelve months, with boys slightly more affected than girls, 25 percent versus 20 percent.

The injuries children reported are a stark reminder of what hazardous exposure means in practice. Nearly 9 percent of children engaged in hazardous work reported skin itches and scratches, while close to 8 percent each reported deep or long cuts and muscle pain. Other injuries included falls from ladders or trees, bruises and dislocations, animal or snake bites, head, back or neck injuries, broken bones, and eye or ear damage. Age patterns were pronounced: among children aged 15 to 17 who worked, 55.4 percent participated in physical and mechanical hazards and 70.8 percent faced environmental or chemical exposures, compared with much lower rates among the youngest children, suggesting that adolescents are systematically deployed in the most dangerous tasks.

The econometric analysis, using probit and multivariate probit models with standard errors clustered at the household level, identified several consistent drivers. Older children were significantly more likely to perform hazardous work, with those aged 15 to 17 showing a 15.5 percentage point higher probability than children aged 5 to 11 in the full sample. Poverty emerged as a powerful predictor: higher per capita household expenditure was associated with a substantially lower likelihood of hazardous work, reducing the probability by 3.7 percentage points in the full sample and 13.7 percentage points among working children, likely because better-off households can hire adult labour or purchase machinery that reduces drudgery. Kinship mattered too. Biological children and grandchildren of the household head were significantly less likely to be exposed to hazardous tasks than other children living in the household, a gap that widened to 19.3 percentage points in the working sub-sample, pointing to weaker bargaining power and lower investment incentives for non-biological children such as orphans and fostered relatives.

One of the most intriguing findings echoes what economists call the wealth paradox. Each additional acre of land owned increased the probability of hazardous work by 1.3 percentage points among working children, consistent with the idea that land operates not only as wealth but as a source of labour demand in settings where hired labour markets are imperfect. Geography and cropping system also mattered: children in the eastern region, where labour-intensive Arabica coffee dominates, faced significantly higher exposure to physical and mechanical hazards than children in the central Robusta belt, reflecting Arabica’s demanding cultivation on steep slopes, its vulnerability to pests, and its selective harvesting of unevenly ripening cherries. Gender shaped the type of hazard rather than overall participation: girls were less likely to face physical and mechanical dangers but significantly more likely to experience poor working conditions and time-related risks, likely reflecting their combined burden of farm work and unpaid domestic responsibilities.

The consequences for education were equally significant. The researchers measured educational outcomes using an education gap, the difference between the years of schooling expected for a child’s age and the years actually completed. Children engaged in hazardous work showed a significantly wider gap, falling more than half a year further behind their expected grade level than non-working peers, and among working children the association remained significant. When hazards were disaggregated, environmental and chemical exposure showed the largest coefficient, followed by poor working conditions and time-related risks, while physical and mechanical hazards alone did not reach statistical significance. The authors are careful to note that these are associations, not causal effects, since cross-sectional data cannot rule out reverse causality or omitted variables, but the pattern aligns with international evidence linking hazardous work to grade repetition, dropout, and poorer learning.

The study’s implications reach beyond Uganda’s coffee sector. Because hazardous child labour represents a social cost of food production that markets do not capture, the findings support the case for full-cost accounting in food systems, an agenda the research was funded to advance through the European Union’s FOODCoST project. The authors point to targeted policy responses: strengthening schooling through fee subsidies and school feeding programmes, raising parental awareness of specific hazardous exposures rather than blanket prohibitions that may simply push work out of sight, and giving particular protection to non-biological children in farming households. As global demand for coffee continues to grow, the study makes clear that sustainability certification and supply chain due diligence will need to look beyond sector labels and ask precisely which tasks children perform, under what conditions, and at what cost to their futures.

Subject of Research: Prevalence, determinants, and educational consequences of hazardous child labour in Ugandan smallholder coffee farming

Article Title: Hazardous child labour in smallholder coffee farming systems in Uganda: Prevalence, determinants, and implications for children's education

Article References: Isoto, R. E., Schilling, F., Shortall, P., Bashaasha, B., & von Braun, J. (2026). Hazardous child labour in smallholder coffee farming systems in Uganda: Prevalence, determinants, and implications for children's education. Journal of Agriculture and Food Research, 31, Article 103347. https://doi.org/10.1016/j.jafr.2026.103347

Image Credits: AI Generated

DOI: 10.1016/j.jafr.2026.103347

Keywords: child labour, Uganda, coffee farming, smallholder agriculture, hazardous work, education gap, ILO Convention 182, poverty, Arabica coffee, food systems, sub-Saharan Africa, household economics

The slow diffusion of AI into everyday work is only one part of the picture. Perhaps more revealing is how adoption is happening in the first place. The study finds that AI use is frequently driven by employees themselves rather than by strategic decisions made at the organizational level. Only 55 percent of AI users report that the tool they use most frequently was officially introduced by their employer. In other words, nearly half of workplace AI usage occurs informally, on the initiative of individual workers who experiment with chatbots, assistants and other applications without formal guidance, approval or support. This pattern has significant implications for how organizations manage technology, because it means that much of the practical knowledge about AI is accumulating outside official channels while companies are still catching up on governance, secure infrastructure and workforce training.

Florian Kunze, head of the study and professor of organizational behavior at the University of Konstanz, frames this as a gap between individual experimentation and organizational transformation. “In many organizations, the major AI revolution has not unfolded as a planned transformation process so far”, he explains. “Instead, many employees are experimenting with AI, while employers are still lagging behind when it comes to establishing clear policies, providing training and implementing secure AI solutions”. The image that emerges is not one of a top-down digital revolution but of a bottom-up, somewhat chaotic process in which workers teach themselves to use new tools while their employers struggle to formulate rules, safeguard data and build the institutional scaffolding that a genuine technological transition would require.

The unevenness of AI adoption extends across occupational boundaries. In office and knowledge-based work, 49 percent of employees now use AI, a figure that reflects the natural fit between generative tools and tasks involving text, analysis and information management. Among employees in production-related and manual occupations, by contrast, only 25 percent report using AI. Both groups have seen year-on-year growth, yet the divide between them remains largely unchanged. The researchers attribute this persistence at least in part to the differing scope for AI applications in the respective fields. Where the core of the work involves manipulating language, code or data, current AI systems offer immediate value; where the work is physical, embodied and hands-on, the technology has so far offered far fewer practical entry points.

Educational attainment produces an even sharper divide. While 56 percent of employees with a high level of education use AI in their work, only 21 percent of those with lower educational attainment do so. Employees with higher education also express more positive expectations about the technology and speak more openly about their adoption of it. The Konstanz researchers note that employees with higher educational qualifications are almost three times as likely to use AI as those with lower levels of education. This is a sobering statistic for anyone hoping that AI might act as a great equalizer in the labor market. Instead, the technology appears to be amplifying existing skill hierarchies, flowing most readily to those who already possess the credentials, confidence and task profiles that make adoption easy.

Organizational size adds a further layer of inequality. Small organizations are significantly less likely to offer AI training, establish formal guidelines for AI use or provide company-owned AI tools. The numbers are stark: only 11 percent of employees in small organizations report having received AI training, and just 10 percent report that binding rules on AI use exist in their workplace. In larger organizations, such support structures are far more common. The consequence is that workers in small firms are left to navigate a powerful and potentially risky technology largely on their own, without training, without clear rules and often without secure, company-sanctioned tools. This gap between large and small enterprises risks compounding the broader pattern in which advantage accrues to those already best positioned to benefit.

Elena Gerdiken, a doctoral researcher at the Future of Work Lab in Konstanz, warns that these disparities carry real consequences for social equality. “The findings show that the AI transformation is unfolding unevenly across society”, she says. “Employees in knowledge-intensive occupations, those with higher levels of education and those working in larger organizations are currently benefiting most from the new opportunities and advancement potential that AI offers. Without targeted support measures, there is a risk that existing inequalities will become further entrenched”. Her assessment points to a central tension in the current moment: a technology that is often discussed in universal terms is, in practice, being absorbed by the workforce in ways that track and reinforce pre-existing lines of privilege.

Attitudes toward AI reveal another striking divergence, this time between societal and personal perceptions of risk. Forty percent of employees expect AI and automation to have an overall negative impact on the labor market over the next ten years. Yet only 17 percent are worried about losing their own jobs. This gap between abstract concern and personal security is a well-recognized pattern in technology attitudes, but the Konstanz data show how pronounced it remains. Employees appear to view AI primarily as a future societal issue rather than as an immediate threat to their own work. Correspondingly, perceptions of concrete organizational change remain limited: only 12 percent of respondents think AI has reduced the number of entry-level positions, and just 23 percent perceive AI as already making a significant contribution to their organization’s value creation.

Carolina Opitz, also a doctoral researcher at the Future of Work Lab Konstanz, interprets this perceptual distance as a defining feature of the current phase. “Many employees view AI primarily as a future societal issue rather than as an immediate threat to their own work”, she says. “Both the risks and the economic potential associated with the technology are currently perceived more strongly at the societal level than in employees’ everyday working lives”. This disconnect matters for organizations and policymakers alike, because it suggests that the workforce may be neither sufficiently alarmed to demand protection nor sufficiently engaged to push for the training and infrastructure that would allow broader, safer adoption.

The Konstanz researchers close their report with a set of recommendations aimed at converting informal experimentation into managed, equitable transformation. They argue that companies should communicate AI initiatives more clearly and actively guide employees as they adapt to the new technologies. Key measures include establishing clear rules for the secure use of AI, expanding job-related training opportunities, providing targeted support for small and medium-sized enterprises, and developing AI solutions suited to a wide range of occupations rather than only to knowledge work. The study, funded by the DFG Cluster of Excellence “The Politics of Inequality” and conducted with the online panel provider Bilendi, builds on earlier waves surveyed in March 2024 with 2,019 respondents and in May 2025 with 1,024 respondents. Taken together, the three waves sketch a consistent trajectory: AI in German workplaces is growing, but slowly, largely from the bottom up, and along fault lines of education, occupation and organization size that society will need to address deliberately if the promised benefits are to be shared widely rather than concentrated among those already ahead.

Subject of Research: Workplace adoption of artificial intelligence and its social inequalities in Germany

Article Title: The use of AI at work: slow, informal and unequal

Article References: The use of AI at work: slow, informal and unequal. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, future of work, workplace adoption, digital inequality, Germany, University of Konstanz, labor market, generative AI, organizational behavior, employee training, occupational divide, education gap