MH
← All essays
Originally on Substack

AI With a Human Face

The case for appropriate AI and a plurality of futures worth choosing

In Consuming Reality: The Religion of Progress, I asked whether we were developing technology to serve human flourishing or gradually redefining human flourishing to serve technological development. I ended on a troubling possibility: that we might build systems which make meaningful human participation increasingly unnecessary, then mistake our own displacement for progress.

I still think that danger deserves to be taken seriously. But I’ve been wondering whether there is something missing from the way I’ve been approaching it. If we spend all our time describing the forces pushing us towards a future we don’t want, we risk making that future seem inevitable. We may inadvertently reproduce the very sense of helplessness we’re trying to challenge.

So I want to try something more constructive. What might it look like to develop AI on different terms? Not to reject its capabilities, nor to resign ourselves to wherever they happen to lead, but to ask what kind of relationship with these systems would actually enrich our lives. Perhaps there need not be one answer, or one inevitable AI future. Different people and communities might reasonably choose different relationships with the technology.

There’s an older tradition of thinking about technology that might help us answer that question. And it begins with an idea that sounds almost quaint in an age of frontier models and autonomous agents: perhaps the best technology isn’t always the most powerful technology. Perhaps it’s the technology most appropriate to the people and purposes it serves.

Technology with a human face

In Small Is Beautiful, the economist E. F. Schumacher challenged the assumption that technological development should always move towards greater scale, complexity and capital intensity. Writing in the context of industrialisation and economic development, he advocated what became known as intermediate or appropriate technology: tools suited to local circumstances, available resources and human needs, rather than imposed simply because they represented the latest industrial advance.

His concern wasn’t that sophisticated machinery was inherently bad, or that everyone should return to a romanticised past. It was that technology should be evaluated in relation to the lives it was supposed to improve. A technically impressive solution could be wholly inappropriate if it created dependencies, displaced livelihoods or required an institutional structure that the people using it could neither afford nor govern.

Schumacher asked whether we might develop a “technology with a human face”. More than half a century later, that seems like an extraordinarily relevant question to ask about artificial intelligence.

Ivan Illich approached the issue from a related but more radical direction. In Tools for Conviviality, he argued for tools that allow people to act creatively, exercise their own judgement and shape the world in cooperation with others. He called such tools convivial. The word wasn’t about pleasantness or friendliness. It described a relationship in which people remain capable of directing the tools they use, rather than organising their lives around the requirements of technical systems.

Illich recognised that a technology could be beneficial at first and become restrictive as society reorganised around it. Cars can extend our freedom of movement, but a city built almost entirely for cars can leave people with little choice but to drive. The tool that once expanded our options can eventually eliminate the alternatives.

That possibility is strikingly familiar in the age of AI. An assistant that helps us write, learn or make decisions may initially expand what we’re able to do. But if education, employment and public services are progressively reorganised around automated systems, the freedom to work or think without them may quietly disappear. The question changes from whether the tool is useful to whether we can meaningfully choose the terms on which we use it.

We could call this approach appropriate AI. There are already related ideas in research on convivial AI, frugal AI and human-centred technology, and I don’t pretend to be inventing an entirely new philosophy. What interests me is bringing their insights together around a deceptively simple proposition: AI should be judged not only by what it can do for us, but by what it enables us to become and remain capable of doing for ourselves. It also suggests something more pluralistic than a single vision of technological progress: different forms of AI, and different choices about using it, can coexist.

More capability is not always more progress

We usually evaluate AI by measuring its performance. How accurately can it answer questions? How quickly can it write software? How cheaply can it perform a task that previously required human labour? Those measurements matter. They tell us what the technology is capable of achieving.

But they tell us remarkably little about whether its use is appropriate.

Imagine an AI tutor that can solve almost any school mathematics problem. On one measure, it is an extraordinary success. A student can obtain answers and explanations that would otherwise have been inaccessible. But suppose the student becomes so accustomed to receiving completed solutions that they never develop the confidence to struggle productively with a problem themselves. We have improved the immediate output while possibly undermining the purpose of the activity.

Now imagine the same system used differently. Instead of supplying the solution at once, it asks the student what they’ve tried, helps them identify a misconception and offers a hint when they’re stuck. It adapts to their needs while preserving the opportunity to understand something for themselves. Success is no longer measured simply by how many questions they answer correctly with assistance, but by what they can understand and do afterwards.

This isn’t an argument for making learning unnecessarily difficult. For a child struggling with mathematics, accessible and patient AI assistance could be genuinely liberating. The point is that the tool should be designed around learning, rather than treating the production of correct answers as equivalent to learning.

I encounter a similar distinction in software engineering. AI can now produce functioning code with astonishing speed. Sometimes that’s exactly what I need. Automating repetitive migrations, generating scaffolding or investigating unfamiliar libraries can free me to concentrate on more interesting problems. But there are also occasions when obtaining the implementation too quickly means I’ve missed the process through which I would have developed an understanding of the design.

An appropriate use of AI might be to explore competing approaches, expose assumptions, critique my reasoning or help me learn an unfamiliar system. In other circumstances, handing over the whole task may be sensible. The difference is not whether a human typed every line of code. It’s whether we have thought seriously about what capabilities and forms of judgement the work requires us to preserve.

The same principle extends well beyond education and software. AI could help a clinician interpret evidence while allowing more time for conversation with a patient. It could reduce administrative burdens on teachers without replacing the relationships through which students develop confidence and belonging. It could make specialist knowledge more accessible to people and communities who have historically been excluded from it.

These are not small ambitions. They involve using highly sophisticated technology. But their success depends on outcomes that cannot be captured by a benchmark score or a productivity dashboard alone.

The relevant measure of progress is not simply how much work the machine takes over. It’s whether the resulting arrangement makes human lives richer, freer and more capable.

The freedom to participate — and the freedom not to

There is a danger of making this argument sound like a moral obligation to do everything ourselves. That would miss the point. Some tasks are exhausting, repetitive, dangerous or simply uninteresting. There is no intrinsic virtue in preserving unnecessary drudgery, and no reason to deny someone assistance merely because another person believes the activity is character-building.

This matters particularly for accessibility. For someone who cannot easily write, speak, see or perform a particular activity, AI might open up possibilities for independence and expression that were previously out of reach. Insisting on unaided human performance as the ideal would turn a philosophy of empowerment into its opposite.

Appropriate AI cannot mean that everyone must use technology less. It has to mean that people gain meaningful freedom to use it in ways that serve their own lives. For one person, that might mean automating a task entirely. For another, it might mean keeping a hands-on role in an activity they value. Often, it will mean moving between those approaches as their circumstances change.

But individual choice isn’t enough when the surrounding institutions are changing. If an employer expects work to be delivered at a pace only achievable with AI, opting out becomes a professional disadvantage. If a public service provides meaningful access only through an automated interface, people may be forced to accept a system they cannot question. And if an organisation has outsourced so much expertise that it can no longer evaluate its providers independently, the nominal option to switch tools may offer little real freedom.

This is why appropriate AI must be about institutions as well as interfaces. We should ask whether people can understand and challenge important automated decisions, whether they have access to a human or other meaningful remedy when things go wrong, and whether organisations retain the skills and infrastructure needed to govern the systems upon which they depend.

There are encouraging signs that these questions are finding a broader audience. A 2026 report from the UN Office for Partnerships, summarising interdisciplinary conversations on AI and human development, draws attention to three revealing tests: can people understand and challenge AI systems, can they manage meaningfully without them, and are they more capable over time because of them? The report records proposals and discussions rather than establishing a binding global standard. Nevertheless, it suggests that the conversation is beginning to move beyond what AI systems can achieve towards what their adoption does to human agency.

The Commonwealth of Learning’s work on Frugal AI points in a related direction. It advocates using only the computational capability needed for a defined purpose and building local institutional capacity instead of defaulting to permanent dependence on external platforms. A locally operated model will not always be the right solution, and smaller systems can still be used in harmful ways. But the emphasis on proportionality, autonomy and long-term capability is important.

The right to participate must be accompanied by the right to decline participation in any particular technical arrangement. Otherwise, we are not choosing how to use AI. We are simply adapting to a world increasingly designed around it.

Who gets to decide what is appropriate?

There is an obvious objection to everything I’ve said so far: appropriate according to whom?

A company might consider an AI system appropriate because it reduces its costs. A worker might see it as an erosion of expertise or bargaining power. A patient might welcome an automated service because it makes care more accessible, while another might fear being denied a conversation with a clinician. Even within a single organisation, people can reasonably disagree about which activities are worth preserving and which are best delegated.

There is no universal answer that can be handed down by technologists, philosophers or policymakers. Indeed, allowing a small group of experts to decide what forms of human participation are valuable would recreate the very problem this approach is supposed to address. A pluralism of AI futures would make room for different priorities: one community might favour locally governed, modest systems; another might welcome extensive automation; a third might choose to keep AI out of particular activities altogether. None should have to treat the preferences of the most powerful developers as destiny.

That is why appropriate should describe a process of judgement, not a fixed category of approved technologies. Pluralism does not mean that anything goes: basic rights, safety and accountability still matter, and one group’s choices should not remove another’s meaningful options. Within those constraints, the people affected by a system need a say in what it is for, what it replaces, what must remain possible without it and what evidence would justify changing course. The choices should be revisable rather than treated as settled by technical progress alone.

In practice, that might mean involving teachers and students in the design of educational AI, not merely testing whether they like a completed product. It could mean workers helping decide which tasks should be automated and how the benefits are shared. It could mean public institutions procuring systems with clear appeal processes, interoperable data and the ability to replace a provider without losing essential services. And it could mean evaluating AI deployments against measures of learning, autonomy, wellbeing and trust alongside efficiency and cost.

None of this is straightforward. Competition can punish organisations that take a slower or more participatory approach. Public institutions have limited budgets. Open systems still require investment, expertise and responsible governance. And there will be genuine disagreements about the trade-offs.

But those difficulties reveal something important. The future of AI is not solely a technical problem waiting for a better model. It is also a question about the distribution of power, the organisation of work and the purposes we want our institutions to serve.

We cannot solve those questions by asking AI systems to become more capable. They require human decisions, made together.

What would we do with the time?

I keep coming back to the promise of liberation from labour. If these systems enable us to produce more with less effort, why should the inevitable outcome be that we produce ever more? Why couldn’t some of the gains become shorter working weeks, more secure livelihoods, better public services and greater freedom to care for one another, learn, create or simply live?

There is nothing automatic about such a future. In competitive markets, saved time can quickly become another demand to deliver more. Those who own productive systems may capture much of the benefit. An organisation may cut costs without giving its workers more autonomy, and a society may grow wealthier without becoming more equal or more at ease with itself.

Making AI appropriate therefore means more than designing friendly assistants. It also means deciding who benefits from increased productivity and whether people have genuine power to shape how that productivity is used. An AI system that reduces administrative work could create breathing room for a teacher or nurse. It could just as easily become a reason to increase their workload. The technology does not settle the question. Our institutions do.

This is where I find the most hopeful possibility. We don’t have to measure the success of AI by the sheer quantity of activity it generates. We could use it to reclaim time and attention for experiences that are valuable precisely because we participate in them: making something, understanding a difficult idea, looking after someone, solving a problem together, or contributing to the place where we live.

I don’t mean that every such activity must remain untouched by technology. I mean that its value cannot be inferred from how efficiently the final result is produced. A conversation with a friend is not improved simply because it can be summarised more quickly. Learning to play an instrument isn’t rendered pointless by software that can generate flawless music. Caring for another person is more than the completion of a sequence of tasks.

There are dimensions of life in which the experience of doing matters as much as, or more than, the output. The ability to distinguish those dimensions from activities that genuinely benefit from automation may be one of the most important forms of judgement we need to cultivate.

Futures worth choosing

I don’t imagine that appropriate AI offers a complete blueprint. It doesn’t resolve the difficult questions of global competition, environmental costs, concentrated ownership or the possibility of future artificial minds with morally significant experiences. Nor does it require that human beings remain the only forms of intelligence worthy of consideration. But recognising the possible value of other minds does not mean that ours must become dispensable, or that superior problem-solving ability confers an automatic right to govern.

What this approach offers is a different starting point. Instead of asking how far we can push automation and then trying to manage the consequences, we begin with the kinds of human lives and shared worlds we want to sustain. There need not be one blueprint for everyone. Only then do we ask which technologies might help us get there, how powerful they need to be, who should control them and what ought not to be sacrificed in the process.

This is not a rejection of ambition. Helping more people learn, giving people with disabilities greater independence, accelerating medical discovery, making valuable knowledge widely available and freeing human beings from dangerous or degrading work are ambitious goals. They may demand extraordinarily capable AI. The point is that capability remains a means to something worth pursuing, not an end that justifies itself.

In my previous essay, I worried that we were becoming consumers of realities constructed on our behalf. The alternative isn’t to abandon the tools that can help us understand and shape the world. It’s to insist that they leave us more able to participate in it.

I still feel uneasy about the direction in which some parts of the AI industry are moving. But I don’t think that unease has to end in resignation. The future will be shaped not only by breakthroughs in laboratories, but by decisions in schools, workplaces, communities and public institutions about what these systems are allowed to replace and what we insist on preserving.

We may not be able to choose the pace at which every new capability appears. We can still contest the purposes to which it is put. And we can refuse the idea that all of humanity must converge on a single technological destination.

Perhaps the most hopeful question we can ask of artificial intelligence is not what it will eventually be able to do instead of us, but what it might make possible for us to do, understand and become — together.

References and further reading