AI đang tạo ra việc làm cũng như xóa sổ chúng – nhưng mức độ hữu ích của chúng ra sao?

Don’t believe the hype about AI taking over all forms of work. Most of the evidence to date suggests large–scale job losses are limited. But what is changing is the way work is organised.

One area of increased employment is in people “cleaning up” after large language models (LLMs) , or quality controlling them.

The sales pitch is that these AI systems can program code, write reports, analyse data and automate routine decisions faster and better than humans, all at negligible marginal cost. In theory, this should free people to focus on higher–level thinking, and leisure.

In practice, AI agents produce vast amounts of “almost right” output – but there are still errors, hallucinations and odd gaps in judgment that a real person has to catch.

The resulting jobs, often known as AI remediation, vary in their skill and remuneration levels. Some AI companies are now hiring experienced professionals in many different fields to ensure their agents are producing work of requisite quality – and to help them change that if not.

But there is also a more subtle shift in job responsibilities within many organisations. Some white–collar workers in industries such as finance, tech and retail are suggesting that AI can add to, rather than reduce, their workload, because of the need to monitor so–called “workslop”. This can also undermine trust in colleagues who are using AI agents.

A 2025 Harvard Business Review report quoted a retail director who lamented spending “more time following up on the information [provided by AI] and checking it with my own research. I then had to waste even more time setting up meetings with other supervisors to address the issue.”

Public sector slop

The public sector risks being drawn into this “AI repair” logic. The UK government (among many others) is experimenting with AI for tasks such as summarising consultation responses, drafting correspondence and analysing sentiment in citizen feedback. Official guidance for civil servants encourages cautious use of generative AI, while warning that its outputs can be misleading and must be independently verified.

This means, if public services start producing large volumes of AI-generated analysis and letters, they will also have to create new positions whose main responsibility is to monitor, check and correct that output before it reaches citizens or feeds into policy.

How much of the apparent AI efficiency gain will be consumed by people silently tidying up after systems that are sold as automatic?

We already know what can happen when AI automation goes wrong in government. In the Netherlands, tax authorities used algorithms to generate risk scores for supposed fraud and automatically flag families for investigation. Tens of thousands of parents – many from minorities – were wrongly accused of fraud, forced to repay benefits they did not owe, and pushed into debt or bankruptcy.

Here, the problem was not just the output but the absence of robust human supervision, able to question the machine’s verdict before lives were destroyed. Any public–sector wave of generative AI adoption will have to invest heavily in exactly that kind of supervision, to avoid repeating the same harms at a larger scale.

There is a deeper trap here. The tokens (the basic building blocks of data processed by LLMs) that generate all this output are heavily subsidised by debt–laden AI companies – but the human time needed to check and fix it is not.

Many companies are deliberately pricing access below cost to build dependence on their systems. Once enough organisations have reoriented their workflows and services around a particular AI model, the company behind it can start to raise prices, limit free tiers, change terms or segment features. So the cost–benefit analysis of AI today may look very different in the future.

Hidden humans

There has already long been a global dimension to the “hidden humans” underpinning AI. Some of the most unpleasant aspects of AI work – chatbot moderation, data labelling, low–level annotation of images – has been outsourced to low–wage countries for many years.

This work has been found to take a substantial mental toll on some of these workers, because of the disturbing nature of graphic text and images of violence, self–harm and the like. Employees in such sectors and geographies also often lack any legal protection.

If AI workslop becomes the norm inside organisations, it is easy to imagine a second wave of outsourcing: the routine checking and correction of AI-generated text shipped to offshore centres – packaged as quality control, but experienced on the ground as endless, underpaid digital piecework.

We can already see the work-reorganisation side of this in translation services. In many agencies, the main workflow is now machine translation plus “post–editing”. The AI does a first pass; freelance human translators are paid less money to correct its mistakes.

Research suggests that post–editing is often paid worse than traditional translation, requires similar or greater cognitive effort, and is widely perceived as tedious and deskilling. Yet those doing this role simultaneously raise concerns about even that position being replaced in the longer term.

How to push back

Popular culture shows how workers can try to resist being pushed into low–status AI clean–up roles. In 2023, Hollywood writers went on strike partly because they feared studios would use AI tools to generate first drafts and then hire humans to polish them on the cheap.

The resulting agreement with the Writers Guild of America set guardrails: AI cannot be credited as a writer, cannot be forced on writers, and cannot be used in ways that undercut human pay and recognition. In effect, the union fought to stop a slide into AI-janitor work, where creative professionals would spend most of their time tidying up after algorithms rather than originating stories themselves,

However, translators, freelancers working in fields such as video–gaming, and other workers have far less collective bargaining power, which makes it easier for AI-driven business models to quietly make repair work as the new normal.

In my view, the problem is not only economic – it is also about meaning. Most people would like to feel their work actually contributes something in the world. They want to believe they are solving problems, helping people, or contributing to making something that would not exist without them.

But for at least as long as access to powerful AI systems remains cheap (and heavily subsidised) , more and more people will be diverted to cleaning up the mess these systems inevitably create. How meaningful these jobs are is, I believe, a question we cannot afford to dodge.

Akhil Bhardwaj does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.