Although artificial intelligence can raise workers’ productivity and leave them free from interference, it can still erode their autonomy. This has an uncomfortable implication for how policy-makers regulate AI at work, since a ‘keep a human in charge’ approach turns out to be the wrong test.
This article is part of a collection on good jobs by the Policy Hub for the Huth Initiative for a New Political Economy.
When we ask whether AI will be good or bad for work, we usually mean one of two things: will AI put too many people out of a job? And how will it affect wages? A third question is now being added to the public debate: does AI make the work itself better or worse?
To address this third question, my co-author and I look at how AI affects worker autonomy. We develop an account of autonomous working with AI assistance tools (or ‘agents’) that can guide the development and deployment of human-complementary AI. We argue that workers need to remain authors of their work, as well as capable of being answerable for what they do.
How are AI agents changing work?
Our research starts with data about how AI agents are changing workers’ experience of their work, and what ‘good work’ is to them. To gather these data, we built an AI interviewer agent to interview IT professionals and data scientists. We chose these workers because they are among the first to feel the full force of generative AI coding tools.
Many workers reported positive experiences about working with AI agents, and positive impacts on their work. Workers said that they were more productive, could accomplish tasks they couldn’t before, had more control over their time and could shift their work to tasks that they found more interesting.
And yet, the same workers voiced a persistent unease that the productivity story does not explain. A recurring fear was of becoming, in one interviewee’s words, ‘a babysitter for AI’. A data analyst worried that colleagues would come to see her as ‘a messenger, not an owner’. Others described feeling ‘redundant’, or reduced to ‘marking work’, even though none of their formal powers had changed.
Workers in our study also worried that important aspects of their work would be taken over by AI. Two themes were particularly prominent here:
- First, they wanted to ensure that they were still in charge of making normative judgments, such as managing trade-offs or understanding a client’s values.
- Second, they wanted to retain responsibility for the work, but worried about being put in the position of being held accountable for a decision in which they were insufficiently involved.
Here is the puzzle. Standard accounts of autonomous work focus on the problem of being bossed around. The dominant way of thinking about freedom at work, in philosophy and in policy alike, is built around what we call the problem of interference: the worry that someone with power over you, such as your boss, can order you around, coerce you, or interfere with your choices (Anderson, 2017).
Debates about algorithmic management and surveillance extend the same picture to AI: the fear is that algorithms hand managers new tools to monitor, direct and discipline workers (Kellogg et al, 2020; Muldoon and Raekstad, 2023). By these control-based measures, the autonomy of the IT professionals and data scientists that we interviewed is not reduced. The AI issues no orders, and the manager’s authority is unchanged. So, what were workers worrying about when they worried about their autonomy?
How can people work autonomously with AI agents?
The non-interference account of autonomy at work is apt when the threat really is a boss with asymmetric power. But AI agents pose a different kind of threat. The coding agent does not dominate the worker or curtail her control. Instead, it quietly takes over the substance of the work while leaving her formally in charge. If we assess job quality only by who holds control and who is free from interference, AI-mediated work can look fine, or even improved, while something that workers value drains away.
What these workers are defending, we argue, is better understood as standing: the normative position within an organisation from which one can make, own and defend one’s work. It requires a domain of work tasks over which you are genuinely in control but adds two additional components.
Authorship
The first component is authorship: genuinely directing the work, or being able to manifest your own judgment, skill and point of view in it, rather than supervising a system that does the substantive thinking. This is why our interviewees guarded judgment-heavy tasks so closely and resented being left with line-by-line validation.
The most positive experiences came when people used AI to draft something that they then made their own; and the most negative came when AI did the thinking and left them to check it.
Tellingly, authorship can be lost without anyone taking it. Several workers described letting AI do too much and finding, as one put it, that they ‘did not own the output’. No boss is interfering – the worker has simply abdicated the position of being the boss of her own work.
Answerability
The second component is answerability: being recognised by others as the person who must defend the work and answer criticism. Crucially, answerability is a position, not a feeling.
Compare someone privately weighing the pros and cons of a medical treatment with a doctor recommending that treatment to a patient. The reflection may be identical – what differs is that the doctor can be called on to justify her recommendation, owes that answer to someone whose interests are at stake, and her answer carries weight.
Work is like this. ‘I need to be able to put my name to something’, as one manager told us. Workers wanted to remain accountable, but only where they had genuine authorship. To be answerable for an AI’s output that you did not shape meaningfully is the worst of both worlds: responsibility without authorship.
Two lessons follow:
- First, autonomy is something that organisations and regulators can build or dismantle.
- Second, autonomy in this sense belongs on the list of the goods of work, alongside pay, security and fair treatment.
What does this mean for policy?
If autonomy is standing, the reflexive policy move to keep a human ‘in the loop’ is not enough, and can even harm autonomy. The European Union AI Act’s flagship safeguard for high-risk systems is a requirement of human oversight (Article 14): the ability to monitor, intervene and override. A worker who can veto an AI but contributes nothing of her own has oversight and no standing. From the standpoint of autonomy, hollow human control is worse than full automation.
For policy aimed at good jobs, three implications follow:
- Design for authorship, not just oversight: job quality standards and public procurement rules for workplace AI should ask whether a tool leaves workers with substantive, judgment-rich work to author, rather than merely whether a human can press stop.
- Match accountability to authorship: workers should not be made answerable for outputs that they had no real hand in shaping, and when they had no control over the choice of whether to use an AI product for a given task.
- Give workers a voice in how AI is introduced and used: which tasks a system absorbs, how credit and responsibility are reallocated, and whether workers have the flexibility to use AI in line with their judgment determine whether workers retain autonomy. Those decisions should be shaped with workers and their representatives.
It is perfectly possible that AI could enhance or indeed erode worker autonomy – and it is not the technology’s raw capability that settles the matter. If a good jobs policy keeps asking only whether a human holds formal control, it will end up certifying workplaces in which workers have quietly ceased to be the authors of their own work.
Where can I find out more?
- Applying AI to rebuild middle-class jobs: This paper by MIT economist David Autor presents an accessible economic case that AI’s effect on good jobs is a design choice, not a destiny.
- Power and Progress: Our thousand-year struggle over technology and prosperity: 2023 book by economics Nobel laureates Daron Acemoglu and Simon Johnson on why technologies augment or replace workers depending on the choices that we make.
- The goods of work (other than money!): 2016 paper Anca Gheaus and Lisa Herzog in the Journal of Social Philosophy.
- Moral deskilling and upskilling in a new machine age: 2015 paper by Shannon Vallor in Philosophy and Technology on how relying on machines can erode the skills through which we make work our own.
- Freedom at work: understanding, alienation, and the AI-driven workplace: 2022 paper by Kate Vredenburgh in the Canadian Journal of Philosophy.
Who are experts on this question?
- Daron Acemoglu and David Autor (MIT) – the economics of automation, AI and the conditions for good jobs.
- Elizabeth Anderson (University of Michigan) – power, interference and freedom in the workplace.
- Lisa Herzog (University of Groningen) and Anca Gheaus (Central European University) – the non-monetary goods of work.
- Iñigo González-Ricoy (University of Barcelona) – workplace republicanism and economic democracy.
- Jeremias Adams-Prassl (University of Oxford) – employment law, algorithmic management and the future of work.
- Karen Levy (Cornell University) – the sociology of workplace surveillance and automation.
- Shannon Vallor (University of Edinburgh) – the ethics of AI, skill and moral deskilling.
- Kate Vredenburgh (LSE) – AI, worker autonomy and the idea of autonomy as standing.
