The hidden labour behind AI belongs in our discussions of responsible technology use. In a 2025 reflection, Mark Graham, Oğuz Alyanak, and Jonas CL Valente argue that educators and universities need to examine the working conditions supporting the systems they adopt. Checking an answer or acknowledging AI assistance addresses only part of that responsibility.
Their article draws attention to people who label data, evaluate outputs, and moderate harmful content. These contributions can disappear from view when we encounter AI through a polished interface. I find their argument useful because it asks us to consider how a tool was produced alongside what it can do in a classroom.
Whose work sits behind our convenience?
We often introduce an AI tool by explaining the time it might save: drafting materials, summarising readings, or generating feedback. That account centres the experience of the person using it. The authors ask us to include the workers whose contributions made those functions possible.
A system can be convenient for a university while raising concerns about pay, security, or worker protection elsewhere in its supply chain. Its educational usefulness cannot settle those concerns.
This connects with my earlier discussion of why generative AI in education was never neutral. Adopting a technology involves relationships with companies, workers, and economic arrangements, even when those relationships remain distant from daily teaching.
Hidden labour behind AI: The paper’s argument
Graham, Alyanak, and Valente (2025), researchers at the Oxford Internet Institute, draw on existing Fairwork research and related scholarship to examine labour supporting AI systems.
They discuss cloudworkers, who work remotely through digital platforms, and data-enrichment workers, whose tasks include preparing and labelling data and reviewing algorithmic outputs. The article describes concerns including inadequate pay, unpaid work, insecure employment, and exposure to disturbing material.
The authors propose using Fairwork’s five principles (i.e., fair pay, fair conditions, fair contracts, fair management, and fair representation) to inform teaching and institutional decisions.
They explain that Fairwork evaluations combine documentary evidence, engagement with companies, and worker interviews. Their recommendations include teaching students about AI supply chains and incorporating labour standards into decisions about purchasing digital services.
Despite the journal’s “Original Research” label, the article presents a reflection based on existing research. Its data-availability statement explicitly says that its analysis does not rely on primary data.
It therefore offers an ethical argument and proposed responses, rather than a new evaluation of classroom practice. It also does not establish the working conditions behind every tool or demonstrate that adopting its recommendations changes those conditions.
Bringing labour into critical AI literacy
The authors recommend making the production of AI visible in teaching. I would begin with a question students can investigate: what human work was needed to develop and maintain this system?
Students could select a documented task, such as data labelling or content moderation, and explain its contribution. They could then examine what an AI provider discloses about the companies performing that work. This would introduce supply-chain questions without assuming students can reconstruct an entire industry from public information.
Fairwork’s principles provide a useful structure for that investigation. What evidence is available about pay? Are working conditions addressed? Can workers challenge decisions or organise collectively? Students should separate documented findings, company statements, and information they cannot establish.
That distinction matters. A classroom investigation cannot reproduce a Fairwork assessment simply by reading a company website. Worker interviews and independent evidence contribute information that promotional material may omit. Equally, missing public information warrants a request for transparency; it does not establish a particular violation.
These are suggested adaptations of the authors’ argument, not tested teaching interventions. They connect with my discussion of practical approaches to critical AI literacy and extend the wider AI Literacy conversation into the conditions of technological production.
Universities also have responsibilities that students cannot carry alone. The authors propose using labour standards and available Fairwork assessments in procurement. An institution could ask vendors what they disclose about subcontractors, worker protections, and mechanisms for addressing complaints. Any assessment should be checked for its date, scope, and the organisation it actually covers.
I would be careful about treating a labour score as a complete ethical verdict on an AI product. It addresses particular working conditions; other questions about data, accessibility, environmental costs, and educational purpose still require attention. The practical contribution is a set of questions institutions might otherwise leave out.
The paper asks educators to connect the benefits we experience with the work performed elsewhere. That connection should be visible when we teach students about AI and when universities decide which services deserve their support.
Reference
Graham, M., Alyanak, O., & Valente, J. C. L. (2025). Reflection AI: Feeding the machine—the hidden labour behind AI tools and ethical implications for higher education. Frontiers in Communication, 10, 1614817. https://doi.org/10.3389/fcomm.2025.1614817
