The human-centered movement in Britain, as the editors recount through the work of founding editor Karamjit Gill, has its origins in the LUCAS Plan of the early 1970s. That plan rested on two planks: socially useful technology, and the user as both producer and user of technology. Gill, writing in 1996, formalized these into two conditions for any technology to be considered human-centered. Technical feasibility is a necessary condition; social desirability is a sufficient one. Both must be satisfied. On this view, satisfying the client, the user or the consumer is not enough, because broader societal implications must be weighed by anyone affected, not only by those directing development from the top. The editors contrast this with the machine-centered model, in which researchers develop products for markets, deploy them, and then assess the results on the affected population. In that model, they argue, needs are not answered by technology; needs are exploited to market technologies.

The editorial opens with a quotation from Mike Cooley, who warned in 2007 that either we follow the techno-centric route and automate people, or we pursue the human-centered route and informate them. Cooley also cautioned that proceeding on a machine-centered basis would produce mistakes that are truly profound. The editors argue that such mistakes are now being made. They point to computer vision as a textbook example of technology-driven development: nobody asked for cameras on every corner, yet advances in image processing fed a surveillance industry that has morphed into real-time tracking, with civilian targeting integrated into military decision-making. They note public pushback, including people removing telemetry systems from modern cars and, in the United States, destroying Flock surveillance cameras installed by government. Computer vision, the editors write, is one thing; ubiquitous surveillance is another.

The authors also describe how conditions can be created that make machine-centered solutions appear necessary. They cite a report in The San Francisco Standard describing how San Francisco, once a leader in the fight against police surveillance, has become what the report called a laboratory for it, with local leaders embracing Flock cameras. If conditions are poor and dangerous enough, the editors observe, pervasive tracking and armed robotic systems can seem to mandate themselves. But they argue that when the outcome of a technology on society is not best for that society, the conditions for its development can never be sufficient. Infrastructure built this way, they warn, cements existing problems into machine automation, and future generations will inherit it as the material consequences of present intentions.

A substantial portion of the editorial draws on survey evidence to document a widening divide between expert and public expectations of AI. The Stanford Report on AI development summarizes results from the Longitudinal Expert AI Panel, or LEAP, an initiative of the Forecasting Research Institute that surveys a fixed pool of respondents monthly. In the first wave, comparing expert and non-expert predictions over two timespans, experts were 25 percentage points more likely than non-experts to anticipate a major mathematical discovery by AI before 2030. The largest disparity concerned overall impact: on what the report describes as a technological Richter scale of disruption, experts estimated a likelihood of a transformative technological event before 2040 some 30 percentage points above non-experts. The expert panel itself was heavily invested in AI’s future: more than half of industry respondents came from the largest AI companies, the median citation count among polled computer scientists exceeded 7000, and 75 of 119 policy respondents worked for think tanks or AI policy institutes.

The editors argue that this divide matters less as a difference in knowledge than as a barrier to communication, precisely at the moments when open communication and coordination among affected communities are most critical. They connect it to the systematic application of a technology-driven philosophy at the highest levels of policy. As an example, they cite US Presidential Executive Order 1431 of July 2025, which commits federal lands and resources to accelerating data center construction in service of what the order calls the prosperity and security of the American people. Yet a Pew Research poll of more than 8500 Americans found that almost 40 percent viewed data centers negatively with respect to energy costs and the environment, and another 30 percent noted negative impacts on quality of life for nearby residents. Only between 4 and 6 percent held favorable views on those dimensions. The editors ask whose prosperity and security the order represents, given that so few respondents associated data centers with their communities’ interests.

Robotics and agriculture supply the editorial’s most concrete illustrations. In LEAP’s seventh wave, covering robotics and physical AI, 214 experts, 55 superforecasters and 630 members of the public were surveyed between March 9 and March 31, 2026. Experts expect a robot to make coffee in an unstructured home environment by 2034, superforecasters by 2032, and the public by 2039. Unassisted surgery is expected by experts by 2040 and by superforecasters by 2035, with the public again at 2039. Experts also tempered their projections of warehouse automation by tolerable rates of social change, noting that people can absorb only so much labor displacement in a given span. At Amazon, sales in US dollars per employee are projected to rise roughly 15 percent by 2030 and roughly 80 percent by 2040, with superforecasters projecting more than 100 percent. The editors accept that warehouses, being flat, predictable and climate-controlled, are paradigm cases where robotics could genuinely help workers, but they question whether capital centralization and the rare-earth mining that robots require satisfy the sufficient conditions of social desirability on a holistic view.

On farming, the contrast sharpens. The machine-centered model is effective for large tracts of mono-cropped industrial grains, the editors note, exactly the applications demanding the most excessive use of pesticides, herbicides and other chemicals. They cite the National Biosolids Data Project, which reports that roughly one-sixth of farmland acres in the United States are licensed for use of wastewater treatment sludge that often contains PFAS and other so-called forever chemicals, and that more than half of wastewater sludge is dumped somewhere in the country. Meanwhile, the Social Change Lab surveyed almost 4000 people in the United States and Britain and found that the more people knew about factory farms, the more they wanted them stopped: at least 73 percent of British respondents opposed factory farms for chickens, pigs or cows, and more than half of British and 45 percent of US respondents would ban factory farming completely. The editors argue that automating such an industry, rather than replacing it, is not socially desirable, and that this divergence between public preference and technological trajectory is a hard fork, not merely a difference in expectations about timing.

The editorial also questions the framing of survey questions themselves. LEAP questions ask when a robot will first perform an appendectomy without human help, not whether it should. From a human-centered perspective, the editors ask why automate appendectomies at all, a procedure humans have performed since at least 1735, rather than pursue preventive care or less invasive alternatives. They cite a 2025 study by Budzyń and colleagues finding that physicians’ unassisted cancer detection rates declined by 6 percent after six months of relying on AI diagnostics, and suggest instead augmenting human surgeons with image processing for precision, helping people care for people rather than replacing them with machines that do not care at all. They propose repurposing the inherited surveillance infrastructure, inverting the panopticon to inform everyday people of the energy, environmental and quality-of-life implications of policy choices, simulating possible futures and letting communities choose among them.

The editors situate their argument within forthcoming work in the journal. Brauner and colleagues, in 2026, measured perceived likelihood, value, risks and benefits across 71 AI scenarios and found that experts anticipate realization of the technologies as more probable than non-experts do, with lower risks and greater benefits, and warn that systems designed primarily by experts trained in benefit-oriented mental models may unintentionally tilt toward technical affordances rather than broader societal values, a bias they call procrustean AI, after the Greek myth of Procrustes, who forced his guests to fit a standard bed. Brian Kelters argues that only by rigorously curtailing AI proliferation can what he terms systematic alignment decay be reliably averted, while David Jacob Harrison describes a reverse alignment problem in which humans are manipulated to become more machine-like rather than machines being aligned with human values. The editorial closes with Gill’s 2025 clarion call that philosophers have long interpreted the world and that it is time to change it, inviting contributions that repurpose available and anticipated technologies toward the holistic human-centered tradition the journal has pursued for four decades.

Subject of Research: Human-centered versus machine-centered approaches to AI development, surveyed in an anniversary editorial for the journal AI & Society

Article Title: Human-centered systems versus the machine

Article References: White, J., Seberger, J., & Hayes, P. (2026). Human-centered systems versus the machine. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03388-w

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03388-w

Keywords: AI & Society, human-centered design, machine-centered systems, LEAP survey, public opinion, surveillance, data centers, factory farming, subsidiarity, socially useful technology, Karamjit Gill, Mike Cooley