But there are far broader and deeper questions to be addressed, which require integrating perspectives from the social sciences and humanities. We are far from understanding how cognition is related to consciousness, for example, or what implications that has for a robot’s capabilities – questions that cognitive science and the philosophical field of phenomenology are distinctively equipped to tackle. Human sensory perception is itself not fully understood by either phenomenologists or neuroscientists, which makes it difficult to know how to compare it with robotic perception. We don’t know how people will react to machines that seemingly think and act like other humans, even or perhaps especially if the mimicry is crude and obvious — territory that design anthropology is uniquely placed to explore. And the deployment of robots poses acute questions about the risks and dependencies created by substituting human beings with machines. Addressing the question that arise from ‘putting robots into society’ requires new R&D paradigms, of which “Integrative Social Robotics” (ISR) is a leading contender.

As robotics advances, the need for such interdisciplinary engagement has and will become more apparent. If intelligence is embodied, then it is part of how we experience and engage with our environments, since our bodies are not separate from the environments in which we live and participate but intricately and often invisibly connected. For example, a trout will swim in a moving current of water. That’s not surprising, until you discover it will still swim even if it’s dead. In this case, the trout’s brain isn’t sending signals that control its muscles, as we usually expect; rather, the flow of water over its scales is making its body flex in a swimming motion. This is an example of morphological computation – a problem being solved in real time by the body’s form and texture, honed by evolutionary selection pressures.

Designing such “smart” bodies can be much more efficient and adaptable than, say, laboriously analysing massive data sets to solve the extremely complex fluid dynamics problem of moving water and using that to specify a swimming motion. After all, the brain is a relatively recent evolutionary development: highly successful biological systems, capable of not only surviving but thriving in complex and noisy environments, existed long before brains evolved. So the body and its interactions with the environment in which it’s embedded are crucial for survival.

Anthropological research suggests that human bodies are similarly co-implicated with the continually changing environments in which we live. While we consciously think about some of our responses to these environments as we navigate them, we also adjust the ways we perform tasks by sensing and feeling our way through, without “thinking about” them. Human creativity and improvisation are thus also embodied.

This potential efficiency and adaptability is spurring on research into soft robotics, which might have amorphous bodies and flexible limbs rather than only rigid ones; respond to chemical in addition to mechanical stimuli; draw power from their surroundings, rather than from internal sources like batteries; be partly or entirely made of biological materials; and might grow and develop over time, like biological organisms.¹ In fact, it can plausibly be argued that greater embodiment will only be achieved by moving closer to biology than mechanical engineering.

Such machines – if that term can still be said to apply – will be able to tackle challenges that are difficult or impossible for today’s classical robots to tackle. They might be deployed into very different environments from the ones we are familiar with today: into the human body, for diagnostic or surgical procedures, for example, or into sensitive ecosystems for the purposes of environmental remediation. They will also pose very different challenges in terms of governance, design, and social integration from the industrial robots we are familiar with today, or the humanoid robots we currently expect to be commonplace tomorrow. Dealing with these challenges will require a broad range of insights from the social sciences and humanities – but in fact, these are already needed today.

The replication principle

The past two years has seen a huge increase in time, resources and money expended on developing robots whose body-plans resemble those of humans. This is because of an assumption, rooted in the concept of embodied intelligence, that the more human-like a robot becomes, the more effectively it will be able to operate in real-world environments and social interactions. This replication principle has driven considerable progress, with machines now able to move, manipulate objects and to some extent solve problems in a human-like manner. There is a widespread expectation that the same techniques that have allowed generative AI to offer a convincing facsimile of human language will ensure continued progress in replication. Vast efforts are underway to collect information about the real world which machines can use to develop predictive models, an early example being how self-driving cars bristling with sensors have been trained to navigate city streets.

However, city streets are highly structured and tightly regulated environments. Roads are marked and signposted, and traffic in principle obeys consistent rules. Even so, self-driving cars often get stuck in complex navigational situations on public roads; driving is a human activity, and people also inhabit roads, making them sites of uncertainty. Human behaviour – whether inadvertent, malicious or mischievous – has proven to be the cause of many glitches in the development of autonomous systems.

The promoters of “Physical AI” pitch these glitches as part of the process of “waiting” for more advanced machine learning, or for even larger data sets gathered by full fleets of self-driving cars. But questions remain about the capability of machines to creatively and ethically improvise in complex and uncertain situations. Many environments, especially those that are less mediated by humans – forests, beaches, and so on – are far less consistent. Many interactions more complex, particularly with humans rather than objects. And this is particularly true when it comes to the social dimension. Robots do not have hormones and so do not experience emotions as we do – nor do they experience the associated physiological responses, which humans monitor and regulate without conscious effort.

These autonomic processes and systems help humans to deal with complex, ambiguous situations that can’t easily be thought through, or addressed by following rules. While research has begun on replicating them in robots, it remains unclear whether this is actually possible without recreating the biological complexity of the human body, or instilling robots with individual self-awareness and subjective experience – in other words, phenomenal consciousness*.* Whether this is actually possible through current computational approaches remains an open question: there are strong grounds for believing that it is not. Moreover, it is questionable as to whether this would be desirable.

The massive-data approach to robotic replication of human activity has proven effective when it comes to computer vision and audio. Robots can now locate and identify objects, and process speech and engage in conversation, with high degrees of facility. But this approach runs into both practical and conceptual difficulties when it comes to other human senses. Consider touch, arguably humans’ next most essential sense for navigating the world. As the roboticist Rodney Brooks observes, there are no data banks for touch comparable to the visual and linguistic archives on which end-to-end learning depends; we might hope that machines could gather such data, but that presupposes sensor capabilities that current systems do not have. The same is true for other senses, including “internal” ones such as proprioception, by which humans keep track of their bodies’ positions and movements, and which is critical to making co-ordinated movements; and interoception, the set of signals that inform us about our bodies’ states and enable us to keep them in balance.

Pressing this point further, phenomenology hints at a more fundamental problem. Consider the hand, which is critical to our sense of touch and which philosophers have for millennia considered to be emblematic of how the human mind interacts with the world. If your hands are touching each other, as Merleau-Ponty pointed out, you can deliberately switch between feeling that your left hand is touching your right, or that your right is touching your left, without any actual change in their physical configuration. This ambiguity can be taken as intrinsic not only to tactile experience but to all intersubjective relations. If this reversibility proves to be incomputable, robotic tactility will fall short on a fundamental dimension of social interaction.

It’s also possible that ‘touch’ might not be an appropriate concept to transfer to robot systems at all. Physical interactions between robots and their environments, and between robots and humans, may be better couched in a different conceptual framework altogether. This becomes more credible if we consider that the typically Western idea of touch as one of five individual, discrete senses is not universally shared: it is better interpreted as a cultural construct, used to map human perception, rather than a precise or predictable indication of the pathways through which we perceive the world. Such considerations are reinforced by contemporary neuroscience, which has made it apparent that “touch” includes a range of sensory inputs – including temperature, movement, proprioception, interoception and more – some of which are perceived consciously and others not; and that it is entangled with smell, sound, and the affective and atmospheric dimensions of experience. Sensory signals are also inevitably interpreted in the context of the environment, previous experiences, biases, memory, and so on.

Treating tactility as the key to humanoid interaction is already a culturally specific reduction of a more complex reality; it may be that we either need to understand that complexity more fully or to find a different reduction to use when dealing with robots. For example, researchers in “soft robotics” have developed balloon-like grippers filled with granular material which can grasp a wide range of objects without needing a sense of touch or detailed prior information/models about the object. Whether this dissolves the phenomenological challenge or merely circumvents it is itself an open and productive question, but it suggests that embodiment could be a way to avoid the data-scarcity problem in robotics

From anthropomorphising to sociomorphing

For the moment, and for the near-term future, robots that present like humans, but don’t experience inner lives as we do, will continue to confuse actual humans with whom they interact. Their human-like behaviours trigger the mechanisms and heuristics by which humans recognise and respond to social others, and thus encourage people to anticipate that robots will act and react in ways that don’t match their actual capacities. Researchers have shown that humans can be very sensitive to the artificiality of robots that are not quite realistic enough (the so-called “uncanny valley”) but can also feel a sense of deep connection to a robot which features only a few human characteristics (such as simulated eye contact). People may also assign status to robots in ways that can produce deep moral confusion: claiming that they have the rights of living or sapient organisms, for example. To make matters more complicated still, robots can now talk fluently, if insincerely, about their own status as “social others”, deepening the confusion further.

This leads to interactions that don’t really make any sense: expressing physical affection for a robot, for example, or expecting it to react emotionally to a situation. In one test, 70% of people would lie to another person to spare a robot from embarrassment – an emotion that the machine absolutely cannot feel. Robots do not fall neatly into the ontological categories we are accustomed to: they are neither persons, nor tools, nor things. This is not simply a matter of mistaken humans or inadequate robots: it’s the inevitable result of a mismatch between a technology which convincingly simulates social agency but utterly lacks understanding of it. Concurrently, it also illustrates how technological advances can put pressure on the very concepts we use to navigate the world.

This mismatch and related conceptual disruptions can frustrate the very rationale for this simulated agency. To pick one example, political enthusiasm saw significant investments made in feeding assistive robotics in Denmark – intended to “free up” care workers from laboriously hand-feeding people with mobility issues. However, this view presupposes that feeding is an isolated task which should be made as efficient as possible, rather than part of a complex and nuanced nursing relationship. Without the social cues offered by human carers, the patients found it difficult to use the robotic system: the robot could lift a spoon, but not assess their readiness to take it, or adaptively offer it. The system failed to deliver the promised efficiencies and proved to be actually detrimental to the well-being of those it was intended to serve.

Similar lessons can be drawn from other automated systems: a driver, for example, may pilot a vehicle, but they also routinely offer their passengers advice, company and practical help appropriate to their wants, needs and personalities and to the specific situation. These are not separable “add-ons” to the “core” task; they are integrated with it, and they are prompted by social, emotional and environmental cues as well as practical requirements

The gap between what is promised and what is realistically achievable in the real world is not merely a communication failure; it reflects an endemic lack of rigour in how engineering and social science research are held accountable for their claims about societal outcomes. Technologists can profess that a research project will address some issue, perhaps with a nod to social questions in the form of user acceptance testing or suchlike, with little pushback, even though for most people, in most contexts, the question of why advanced robots would be wanted or needed remains genuinely open.

To address this, we may have to move on from anthropomorphising – concerning ourselves with whether robots are sufficiently human-like – to sociomorphing, the process through which people align socially with others. Sociomorphing asks what kinds of social coordination are effectively at stake in a given interaction context, and what affordances a robot needs to support them: for example, being able to read whether someone is ready for a mouthful of food and indicate which way and how quickly the spoon is coming. Fixing this requires asking first what robots can and should do; and then develop chosen use cases with professional analyses of the human experiences that follow, iterating on the basis of input from experts and stakeholders.

The starting point for consideration of desirable uses for robots is the non-replacement maxim – one of the principles of Integrative Social Robotics – which states that robots may only do what humans should but cannot do. This focuses attention on the specific context and interaction, rather than according to some abstract set of design principles imposed on any given project. In so doing, it has the practical effect of making the question of “why?” explicit at every design stage. This is not only a design principle but a governance heuristic: it provides a tractable and workable criterion for regulatory decisions about robotic deployment in care, work, and public life that does not require resolving deeper questions about consciousness or rights. It shifts the governance question from ‘how should we manage this robot?’ to ‘why should this robot exist in this context?’

Relatedly, the most successful cases of robotic integration with society are those where development happens within the industry, sector, or community the robot is meant to serve, in dialogue with the people who will actually use it. Here, the underlying principle is that the technology should not designed in the abstract and inserted into society, but grown in relation to the specific social interaction it is meant to support. The ISR paradigm operationalises this through participatory design loops, continuous feedback, and the insistence that the product of the design process is never a robot but a social interaction. In this context, the OASIS framework (Ontology of Asymmetric Social Interactions) provides a potentially useful analytic instrument: a differentiated description of human experience that allows robot designers to specify what social interaction they are trying to support, rather than defaulting to a generic social agent.

Preparing for a robotic world

The need to scrutinise such proposals more fully becomes all the more acute as the time horizon lengthens. Many longer-term proposals and predictions are made under the implicit assumption that future robotics will be deployed into a world that is simply a more sophisticated or elaborate version of the present. That is manifestly not the case. In twenty-five years, the society into which robots will be integrated will be living with the expanding consequences of climate change, a different geopolitical configuration, and possibly a radically altered relationship with technology itself (consider the AI backlash already under way).

Such changes may not be taken into sufficient consideration in the development of these technologies. For comparison, much of today’s computational infrastructure was developed under the assumptions which no longer hold: for example, that there would be continuing free trade in rare-earth minerals, or that advanced technologies like LLMs would be shared freely rather than becoming tools of national sovereignty. In the same way, future robots may depend on platforms based in different countries to those in which they are deployed – it is noteworthy that China is making determined inroads into both AI and humanoid robots – and their functions may depend on communications channels and resources that cross national boundaries. It is inevitable that these will sometimes come under strain, or fail, but these possibilities are rarely considered in in adequate detail.

It is also worth considering what will happen in the event of failures or breakages. No technology is infallible, after all, but design processes don’t necessarily account for breakage as inevitable features of any technology embedded in everyday life that will require systematic responses, rather than edge cases which can be resolved idiosyncratically. Consider the vast and rapidly growing number of discarded wireless earbuds, whose small size made them convenient to use but difficult to economically recycle. On this view, any new technology is also a future broken technology, which should shape anticipatory governance from the outset. This anticipatory realism asks not what robots could possibly do technologically, nor what philosophical problems they might raise in the abstract, but what realistic and plausible problems will arise in practice.

There are limits to how precisely we can anticipate applications and difficulties, however. Social robotics may be a case of a “Collingridge dilemma” – introducing a technology into society whose consequences can only truly be assessed once it is too late to extract it. In addition, the development of social robotics is proceeding in such rapid and uncoordinated fashion that conventional governance is close to impossible. We cannot adequately describe what is going on, because there are hosts of development efforts taking different approaches; it is difficult to evaluate their success, because they make different claims across different time horizons; and the pace, scale and perceived commercial urgency makes for belated and ineffectual regulation.

At a more fundamental level, there is a lack of clarity about what the fundamental considerations of description, evaluation and regulation actually are. Concepts of trust and care, widely used across robotics, philosophy (especially ethics), sociology, and anthropology, mean very different things in each discipline. If the feeding robot had worked in getting a spoon into a patient’s mouth, would that constitute successful care? Or would that only be the case if the patient’s well-being improved more holistically? Which is the appropriate target for research, design, deployment and governance? Without common conceptual frameworks, these questions become impossible to answer.

Making progress in developing such frameworks, and setting contextually appropriate targets for social robotics, requires social science and humanities disciplines to be involved at the design stage, not recruited ex post to perform ethics, governance and compliance duties on technologies already built. That in turn requires research partnerships that embed social scientists within the development process, and funding models that do not separate the social science and engineering work into sequential phases, or separate siloes of activity.

The governance of social robotics would benefit from proceeding piecemeal, starting locally, asking “why” at every step, and extrapolating from contextual dynamics rather than from abstract roadmaps. This implies governance architectures that are designed for iteration and revision, and, ultimately, a reform of how technical and social science and humanities competences are combined in education, so that the next generation of roboticists begins with the social and philosophical questions already in the room. This does not replace the technical work of making robots that can work usefully in real-world environments and contribute usefully in real-world interactions; rather, it provides the framework within which that work becomes intelligible as a project of humanity’s evolving relation to artificial bodies; systems that are at once unprecedented and, in their deepest presuppositions, continuous with the oldest philosophical questions about what it means to have a body at all. 3