We Are Not Machines: Craft, Care and Collective Agency

Sarah O’Connor’s We Are Not Machines: The Fight for the Future of Work accompanied me on my summer holiday, but it has continued to unsettle how I think about work. It is an engaging, accessible book, and precisely the kind that prompts a recalibration of what you want your own work to be about. In the age of AI, it offers a useful lens for deciding what we should automate, what we should protect, and who gets to make those choices.

I am not anti-AI; I am pro-human. I routinely use AI tools, my research involves designing hardware for AI, and I welcome their use where they genuinely enhance life. The difficult questions are which parts of life they enhance, who gets to decide, and what exactly we mean by “enhancement”.

O’Connor examines the intersection of work and technology through our collective struggle to preserve some of our most cherished human traits: “to pass skill and knowledge from one generation to the next; to create, and to delight one another with our creations; to care for each other when we are too weak to care for ourselves”. The book is structured into three sections: “Mind”, “Body”, and “Soul”, which explore autonomy in cognitive work, safety in physical labour, and the intrinsic value of skill, creation, and care. Moving across warehouses, mines, care homes, and creative studios, O’Connor repeatedly returns to three questions: who shapes technology and its adoption, under what incentives, and in whose interests?

For me, the book’s deeper lesson is that these choices are rarely individual. What workers can protect depends on whether they have any collective power over how technology is introduced. The title itself comes from the 1969-70 strike by miners at the Swedish state-owned company LKAB. Their rallying cry was Vi är ej maskiner: we are not machines. Tellingly, the book’s most hopeful narratives are those in which people act collectively to reclaim agency over how their work is organised.

Craft and “Vibe Knitters”

A concept that particularly resonated with me is O’Connor’s use of Karri Saarinen’s account of craft: “the deliberate attention put into making something excellent, not because someone is checking, but because it matters to the maker”. O’Connor also cites Saarinen’s wider economic point that when something becomes cheaper to build, the default outcome is often simply to build more of it, making us less critical of what actually deserves to exist. I recognise this tension from my own vibe coding. The speed and ease are exhilarating, yet they risk detaching the act of production from the rigorous exercise of judgement.

O’Connor restores the historical context around the stockingers who became Luddites. Far from being ignorant reactionaries smashing unfamiliar machines, they were highly skilled framework knitters reacting against particular uses of machinery that degraded both their trade and its products. She refers to the employment of unapprenticed “colts” to turn out cheap goods as “vibe knitters” of their day. The historical question, then as now, is about the social relations in which machines are deployed, the kinds of work they enable, and the products they are used to create.

This discussion connects to a much older debate about the division of labour. O’Connor cites Frederick Winslow Taylor, whose The Principles of Scientific Management sought to convert workers’ accumulated craft knowledge into “rules, laws, and formulae” held and applied by management, displacing individual judgement. Against this, she places John Ruskin’s 1853 warning in “The Nature of Gothic”: it is not merely labour that is divided, but people themselves, broken into “fragments and crumbs of life”.

Karl Marx’s theory of alienation has much to add to this discussion. In the 1844 manuscripts, Marx describes four connected dimensions of alienation: the worker is estranged from the product, from the activity of labour, from other people, and from the human capacity for free, conscious creation. His observation that the worker’s activity “belongs to another” and is “a loss of his self” is therefore not simply a claim about who owns the product. It highlights the worker’s estrangement from their own activity – work experienced as external, compelled, and no longer self-directed. By 1848, Marx and Friedrich Engels wrote in The Communist Manifesto that the worker “becomes an appendage of the machine”.

Each of these thinkers helps illuminate what changes when workers are separated from meaningful responsibility for both their labour and its result. Specialisation enables production on a scale impossible for a solo craftsperson. But those gains ring hollow when work is so fractured that nobody can see, or care about, the whole. Something human is extracted from the process, even as measurable output soars. Unless, perhaps, there is collective and living control over the entire enterprise.

What Should We Automate?

O’Connor quotes an interviewee who poses a striking question: why are we so eager to automate the cognitive activities our minds excel at, and which nourish our minds, rather than the physical tasks our bodies struggle with and which cause physical damage? This question stayed with me throughout the holiday.

Expanding on this, she cites Andreas Schleicher on generative AI in education. A system can effortlessly throw an answer back at us without revealing its provenance or how it was constructed. I recently experienced my own somewhat jarring version of this. I found myself spending more time investigating the intellectual lineage of a mathematical proof generated by GPT-5.5 than I spent proving the theorem from scratch in Lean. It inverted my sense of what takes time in research. Producing the formal mathematical object was almost instantaneous; establishing its origins, understanding its mechanics, and deciding whether I actually thought it illuminating remained stubbornly slow.

This is not necessarily a negative development, but it shifts where the craft lies. Unless we notice that shift, we risk mistaking possession of an answer for understanding.

Seeing the Whole Person

Perhaps the book’s most compelling case study is the Dutch home-care organisation Buurtzorg. Its model relies on small, self-managing teams of nurses who take holistic responsibility for the entire care process. A highly trained nurse might administer complex medication, dress a wound, and then make the patient a sandwich. Viewed through a strictly Taylorist lens, making a sandwich is a gross misallocation of expensive skill and ought to be delegated to cheaper labour. Viewed as part of human care, however, it becomes an opportunity to observe how that person is living and assess their broader needs. The simple task is inseparable from the skilled one.

The same pattern extends beyond nursing. I see it in the creeping deskilling of teaching in England, where the educator’s role is increasingly carved into separately managed tasks. An Ofsted study of teacher wellbeing found that limited influence over policy – teachers feeling “done to” rather than “worked with” – contributed to a sense of de-professionalisation. The same logic shapes the ticketing systems through which large institutions, including my own, manage HR and ICT support. It also shapes higher education, where interactions with staff and students are increasingly packaged, routed, and siloed. In each case, relationships that depend on continuity and judgement are divided into discrete, measurable transactions.

Specialists matter, and a functioning ticketing system is preferable to chaos. Yet any honest accounting of efficiency must include the hidden cost of nobody seeing the person – or the problem – as a whole. It must also account for what workers lose when their roles become too narrow for them to exercise judgement or care about the outcome – and, centrally, when they cannot reach beyond those roles through collective control of the production process as a whole.

Who Is the “We”?

O’Connor skewers what she describes as a cliché that technologies are “only tools” and that what matters is how we choose to use them. Her counter-question is simple: who exactly belongs to the “we” making those choices?

It is striking that almost all the battles for human-centred work detailed in the book are collective ones, even when their victories are ultimately codified as individual rights. The Swedish miners did not optimise their relationship with machinery through individual negotiation. More recently, during its 148-day strike, the Writers Guild of America won rules in the 2023 agreement governing AI: AI could not write or rewrite literary material under the agreement; AI-generated material was not source material; and writers could not be required to use AI. These were foundational choices about technology, won through organised labour.

Because such victories are codified as rights exercised by individuals, their collective origins are easy to forget. Nicos Poulantzas is useful here. In State, Power, Socialism, he treats the state as a material condensation of the balance of forces among classes and class fractions. This offers one way of seeing how rights that appear to belong to isolated individuals can bear the imprint of collective struggle. The right may be granted and exercised individually; the power that made it possible was collective.

The “we” defining technological choice must include those actually performing the work. For people divorced from the ownership and control of the technology, exhortations to preserve their craft are nearly useless if they control neither the tools nor the targets and metrics by which their work is judged.

More Than Machines

The final chapters include perhaps the book’s most unsettling idea: that what some call artificial general intelligence (AGI) might arrive not because machines advance, but because humans retreat from what we are capable of when faced with apparently superior machine performance. Organisations may treat us as machines, but we may also come to understand ourselves as defective machines, slow and inconsistent approximations of systems whose strengths we have adopted as the gold standard.

I closed the book wanting to identify – and ruthlessly prioritise – the crafts I most enjoy, at work and beyond. I remain eager to use AI to automate genuine human drudgery. Yet the current temptation is to automate whatever is easiest to automate, rather than what we actually wish to relinquish, on the vague promise that doing so will free us up for… what exactly?

Much of life’s joy resides in craft. Those of us fortunate enough to retain some autonomy over our work and lives should think carefully before surrendering it. O’Connor’s narratives show that preserving human craft requires both individual choice and collective action. Remaining more than machines is not merely a personal preference, it is a collective and deeply political task.

Sources and further reading

Quotations not otherwise linked are drawn from Sarah O’Connor’s We Are Not Machines: The Fight for the Future of Work.

People

Primary texts

Broader context

AI tools were used to support copy-editing and to help collect and organise the links above.

Do Your Best: A Social Question?

I’ve always struggled with the concept of “doing your best”, especially with regards to avoiding harm. This morning from my sick bed I’ve been playing around with how I could think about formalising this question (I always find formalisation helps understanding). I have not got very far with the formalisation itself, but here are some brief notes that I think could be picked up and developed later for formalisation. Perhaps others have already done so, and if so I would be grateful for pointers to summary literature in this space.

The context in which this arises, I think, is what does it mean to be responsible for something, or even to blame? We might try to answer this by appeal to social norms: what the “reasonable person” would have done. But in truth I’m not a big fan of social norms: they may be politically or socially biased, I’m never confident that they are not arbitrary, and they are often opaque to those who think differently.

So rather than starting from norms, in common with many neurodivergents, I want to think from first principles. How should we define our own responsibilities when we act with incomplete information? And what does it mean to be “trying one’s best” in that situation? And how do we not get completely overwhelmed in the process?

Responsibility and Causality

One natural starting point is causal responsibility. If I take action A and outcome B occurs, we ask: would B have been different if I had acted otherwise? Causal models could potentially make this precise through counterfactuals. This captures the basic sense of control: how pivotal was my action?

But responsibility isn’t just about causality. It is also about what I knew (or should have known) when I acted.

Mens Rea in the Information Age

The legal tradition of mens rea, the “guilty mind”, is helpful here. It recognises degrees of responsibility, such as:

Intention: I aimed at the outcome.
Knowledge: I knew the outcome was very likely.
Recklessness: I recognised a real risk but went ahead regardless.
Negligence: I failed to take reasonable steps that would have revealed the risk.

It’s the final one of these, negligence, that causes me the most difficulty on an emotional level. A generation ago, a “reasonable step” might be to ask a professional. But in the age of abundant online information, the challenge is defining what “reasonable steps” now are. No one can read everything, and I personally find it very hard to draw the line.

If we had knowledge of how the information we gain increases with the time we spend collecting that information, we would be in an informed place. We could decide, based on the limited time we have, how long we wish to explore any given problem.

From Omniscient Optimisation to Procedural Reasonableness

However, we must accept that there are at least two levels of epistemic uncertainty here. We don’t know everything there is to know, but nor do we even know how the amount of useful information we collect will vary based on the amount of time we put in. Maybe just one more Google search or just one more interaction with ChatGPT will provide the answer to our problem.

In response, I think we must shift the benchmark. Trying one’s best does not mean picking the action that hindsight reveals as correct. It means following a reasonable procedure given bounded time and attention.

So what would a reasonable procedure look like? I would suggest that we start with the most salient, socially-accepted, and low-cost information sources. We then keep going with our investigation until further investigation is unlikely to change the decision in proportion to its cost.

In principle, we may want to continue searching until the expected value of more information is less than its cost. But of course, in practice we cannot compute this expectation.

A workable heuristic appears then to be to allocate an initial time budget for exploration, and if by the end the information picture has stabilised (no new surprises, consistent signals), then stop and decide.

I suspect there is a good Bayesian interpretation of this heuristic.

The Value of Social Norms

What then of social norms? What counts as an obvious source, an expert, or standard practice, is socially determined. Even if I am suspicious of social norms, I have to admit that they carry indirect value: they embody social learning from others’ past mistakes. Especially in contexts where catastrophic harms have occurred such as in medicine and engineering, norms, heuristics and rules of thumb represent distilled experience.

So while norms need (should?) not be obeyed blindly, they deserve to be treated as informative priors: they tell us about where risks may lie and which avenues to prioritise for exploration.

Trying One’s Best: A Practical Recipe

Pulling these threads together, perhaps “trying one’s best” under uncertainty means:

Start with a first-principles orientation: aim for causal clarity and avoid blind conformity.
Consult obvious sources of information and relevant social norms as informative signals.
Allocate an initial finite time for self-investigation.
Stop when information appears stable. If significant new evidence arises during investigation, continue. The significance threshold should vary depending on the potential impact.
Document your reasoning if you depart from norms.

Responsibility is not about hindsight-optimal outcomes. It is about following a bounded, transparent, and risk-sensitive procedure. Social norms play a role not as absolute dictates, but as evidence of collective learning obtained in the context of a particular social environment. Above all, “trying one’s best” means replacing the impossible ideal of omniscience with procedural reasonableness.

While this approach still seems very vague, it has at least helped me to put decision making in some perspective.

Acknowledgements

The idea for this post came as a result of discussions with CM over the last two years. The fleshing out of the structure of the post and argument were over a series of conversations with ChatGPT 5 on 26th October 2025. The text was largely written by me.

Ontology and Oppression

This Autumn I read Katharine Jenkins’ book Ontology and Oppression. The ideas and approaches taken by Jenkins resonated with me, and I find myself consciously or subconsciously applying them in many contexts beyond those she studies. I therefore thought it was worth a quick blog post to summarise the key ideas, in case others find them helpful and to recommend you also read Jenkins’ work if you do.

Jenkins studies the ontology of “social kinds” from a pluralist perspective – that there can be many different definitions of social kinds of the same name, e.g. ‘woman’, ‘Black’ – and that several of them can be useful and/or the right tool to understand the world in the right circumstances. After a general theoretical introduction, she focuses on gender and race to find examples of such kinds, but the idea is clearly applicable much more broadly.

Jenkins begins by describing her “Constraints and Enablements” framework, arguing that what it means to be a member of a social kind is at least partly determined by being subject to certain social constraints and enablements, which Jenkins classifies in certain ways. These can be imposed on you by (some subset of) society or can even be self-imposed through self-identification as a member of a given social kind. Jenkins defines two types of wrong that can come about as a result of being considered a member of a given social kind, ‘ontic injustice’, where the constraints and enablements constitute a ‘wrong’, and a proper subclass, ‘ontic oppression’, where the constraints and enablements additionally “steer individuals in this kind towards exploitation, marginalisation, powerlessness, cultural domination, violence and/or communicative curtailment”. She argues that a pluralist framework can be valuable as a philosophical tool for liberation, and studies how intersectionality arises naturally in her approach.

The race and gender kinds Jenkins studies, she classifies as ‘hegemonic kinds’, ‘interpersonal kinds’ and ‘identity kinds’. I find this classification compelling for wanting to really understand power structures and help people rather than simply shout about identity politics from the sidelines – a form of intervention that sadly characterises much of the ‘debate’ in ‘culture wars’ at the moment. It also provides a useful toolbox to understand how a social kind (e.g. ‘Black’, ‘woman’) can be both hegemonically oppressive and yet corresponding interpersonal and identity kinds can sometimes serve an emancipatory function.

Ultimately, Jenkins’ description allows us to break away from some of the more ridiculous lines of argument we’ve seen in recent years, trying to ‘define away’ issues. At the end of the book, Jenkins takes aim at the ‘ontology-first approach’: the idea that one should first settle ‘the’ meaning of a social kind, e.g. ‘what is a woman?’ and from that apply appropriate (in this case gendered) social practices. This approach, so widespread in society, Jenkins shows does not fit with her framework. She challenges us to ask: what do we actually want to change about society? And from that, to understand what kinds make sense to talk about, and in what context, and how.