A researcher finishes a doctorate. She has spent years studying how technologies, including artificial intelligence, are changing the lives of agricultural workers in South Africa. She is ready to begin an academic career. Then she is asked to help train an AI system — work that sits directly on top of the expertise she just spent years building. She says no. That refusal, and what it reveals about AI labor accountability, is worth paying close attention to.

What happened

The researcher had completed doctoral work focused on how AI and related technologies are reshaping agricultural labor in South Africa. At the point of transitioning into a professional academic role, she encountered a request — the precise source, terms, and intended application of which have not been fully disclosed — to contribute to training an AI system.

She refused.

Her refusal was not a technical objection. It was not a contract dispute. It was a deliberate decision rooted in her own research findings. She understood, from years of fieldwork, what AI systems do to labor markets. She was not willing to participate in building a tool she believed could substitute for the kind of work she was preparing to do.

That clarity — the ability to see the dynamic and name it — is rarer than it should be. Most workers asked to do similar things do not have a completed doctorate in the subject to draw on.

Who is affected

Early-career researchers and academics are among the most exposed workers in this situation. They enter institutions at exactly the moment those institutions are most actively experimenting with automation. They also carry the least institutional weight to push back without consequences.

Beyond academia, the problem is structural. Any worker in a field where knowledge is specialized but reproducible — meaning it can be observed, documented, and fed into a training dataset — faces a version of this same tension. A training dataset, in plain terms, is the collection of examples and information that an AI system learns from. If your expertise can be written down or demonstrated in a way a machine can process, it can potentially be used to train a system that does what you do.

Agricultural workers in South Africa — the population at the center of this researcher’s doctoral work — represent a much larger group that rarely gets any say in whether AI systems affecting their livelihoods are built at all. They are affected by the outcome without being consulted about the process.

Institutions employing researchers and knowledge workers face a question most have not answered publicly: what obligation do they have to workers whose skills are being harvested to build systems that may eventually displace those same workers?

What the real risk is

The most immediate risk is not job loss in the abstract. It is normalization. When contributing to AI training is framed as a routine professional obligation rather than a choice, workers lose the ability to recognize it as a decision — let alone refuse it.

There is also a quality problem. When people with the deepest domain knowledge opt out, or are never genuinely consulted, the systems that get built are trained on a narrower base of input. The result reflects institutional power more than genuine expertise.

The accountability gap is structural. No existing regulatory framework clearly assigns responsibility for this dynamic. It is not obviously a labor law violation. It does not fit neatly into data protection rules, which focus primarily on personal data rather than the professional knowledge workers embed in training datasets. It is not straightforwardly a consumer harm. That ambiguity is exactly why the practice continues without scrutiny.

What to do today

These steps are achievable this week, regardless of your technical background.

• Treat it as a decision, not a default. If you have been asked to label data, review AI outputs, or document your processes for any system, ask explicitly: will this output be used to automate or evaluate my role, or roles like mine? You are entitled to that answer before you contribute.
• Document requests in writing. If a request is made verbally, follow up by email the same day to confirm what was asked. Keep a copy outside your work systems if possible. A written record matters if the situation escalates later.
• Find out whether your institution has a policy. Ask HR or a governance contact whether there is a formal policy on staff contributions to AI training. The absence of one is itself useful information — and surfacing that gap to the right person can prompt one to be created.
• Connect with professional associations. Many labor organizations and academic associations are actively developing positions on AI and worker data. Some have published guidance already. Find out whether yours has, and if it has not, ask why.
• Amplify accounts like this one. When researchers or workers speak publicly about these refusals, that visibility is part of how the problem becomes legible to regulators and the public. Sharing credible reporting on AI labor accountability is a concrete act, not a passive one.

Why this keeps happening

The pattern here is familiar from other accountability failures in digital systems. The cost of a systemic practice is distributed across many individuals. The benefit concentrates in fewer hands. No single incident is severe enough on its own to force a reckoning.

Institutions that benefit from AI development have little structural incentive to formalize the ethical questions around labor contributions to training data. Formalization would require offering workers more than they currently receive — recognition, compensation, or the right to refuse without professional penalty.

Early-career workers are especially vulnerable. Their professional survival depends on being seen as cooperative and adaptable. Refusal carries a reputational cost that more established workers do not face to the same degree. The researcher in this case had the protection of a completed degree and a clear intellectual framework. Most people in comparable positions have neither.

Regulation has not caught up. Existing labor law was not written with AI training contributions in mind. Data protection frameworks were designed around personal data, not professional knowledge.

There is also a deeper structural problem that makes oversight harder than it should be. Online and institutional systems rarely have reliable ways to tie actions to identifiable, responsible parties. A request to contribute to AI training may pass through several layers — a platform, a contractor, a department head — before it reaches a worker. By the time a regulator or an HR department tries to trace accountability, the chain is diffuse enough that no single actor is clearly responsible. That diffusion is not always accidental. It is one reason that even well-intentioned oversight struggles to find a target. The framing of AI development as inherently beneficial innovation makes this harder still — it raises the political cost of scrutiny without changing the underlying labor dynamic.

Frequently asked questions

Is refusing to help train an AI system a legally protected act?

In most jurisdictions, there is no specific legal protection for this kind of refusal. Existing labor law does not clearly address contributions to AI training as a distinct category of work. Whether a refusal could trigger any employment protection depends heavily on local law and the specific circumstances. If you are considering refusing a request of this kind, speaking with a union representative or employment lawyer before doing so is worth the effort.

Why does it matter if a worker contributes to AI training if the job still exists for now?

Because the contribution may be part of what eventually eliminates the job. More immediately, contributing without understanding the terms means giving up something of value — specialized knowledge — without consent, compensation, or acknowledgment. The harm is not only future job loss. It is the extraction of expertise under conditions that obscure what is actually happening.

What would meaningful accountability look like in situations like this?

At minimum, it would include clear disclosure to workers about how their contributions will be used, genuine consent rather than assumed compliance, and some form of recognition — whether financial or professional — for the value being extracted. At an institutional level, it would mean formal policies that workers can point to, and governance bodies with the authority to enforce them. None of that exists in most workplaces today.

Originally reported by restofworld.org. This article summarises that reporting and adds practical guidance.

Scams, fraud, bots, and manufactured noise keep spreading because the internet was built with no reliable way to know who anyone actually is. Everyone deserves authenticity and accountability online, and that is the mission we are working on. Subscribe to follow our reporting as it develops.