Module 1, Sep 2, Seminar and lab

Defining technology for social good

Background

The internet was announced as a great equalizer, and for a long stretch the evidence agreed. Speaking, organizing, and publishing got cheap enough that Benkler credited the network with a new, commons-based mode of production, with free software and Wikipedia as his leading exhibits.

Economists measured the claim in the field, and the cases that hold up best are the ones where information reached people who had been priced out of it. When mobile phones arrived along Kerala’s fishing coast between 1997 and 2001, price differences across fish markets nearly vanished, waste was eliminated, fishermen’s profits rose about 8 percent, and consumer prices fell about 4 percent, gains at both ends of the market at once. In Niger, mobile phone coverage cut grain price dispersion across markets by 10 to 16 percent, with the strongest effects between markets separated by the highest transport costs.

The measured gains are not only about prices. When submarine internet cables reached Africa’s coastal countries, employment rose in the connected areas, including for workers with little formal education, with little or no sign that the jobs were displaced from elsewhere. And when cable television reached Indian villages, reported acceptance of domestic violence fell, son preference fell, girls’ school enrollment rose, and fertility declined, shifts the authors sized, in some cases, as equivalent to about five years of education.

One strand of communication research frames the two readings as competing hypotheses. The social diversification hypothesis holds that marginalized and minority groups take up the internet to widen their networks and reach past physical and social barriers, while the social stratification hypothesis holds that digital tools mostly reproduce the inequalities already in place; researchers test the two against each other on whether online life narrows the gap or widens it. The stratification side has its evidence. Access and dependence still track income and race. In Pew’s January 2026 numbers, 94 percent of adults in households earning over 100,000 dollars a year subscribe to home broadband; in households under 30,000 dollars, 54 percent do. Smartphone dependence, owning a phone but lacking home broadband, runs 28 percent among Hispanic adults and 19 percent among Black adults, against 13 percent among white adults.

The network that promised to distribute power has concentrated it in a few companies whose reach now shapes what counts as development or social good. Khanal, Zhang, and Taeihagh call Alphabet, Amazon, Apple, Meta, and Microsoft “super policy entrepreneurs,” active at every stage of policymaking, and argue that the generative AI boom deepens the position because the same firms hold the compute, data, and talent. Telehealth is a good test case for the mixed record. A 48-study systematic review documents access gains next to persistent disparities, and a review of reviews finds the equity evidence thinnest exactly where digital literacy and infrastructure come in.

Michigan has already run one version of the experiment. In 2013 the state’s unemployment agency switched on MiDAS, a 47-million-dollar automated system that adjudicated fraud cases without human review. Within two years it had falsely accused about 40,000 residents of fraud, and when the Michigan Auditor General examined 22,000 of its determinations in 2016, 93 percent involved no fraud. Tax refunds were seized and wages garnished on the strength of those determinations, and a class action over the system settled for 20 million dollars, with final approval in January 2024.

The course holds both readings open all semester, giving each week’s proposition the strongest evidence on both sides.

Scoping technology

Before the course can ask what technology for social good is, it has to say what each half of the phrase means. This section starts with the first word, technology, and turns to the second, “social good,” at the end.

Emerging Technologies

Emerging technology is where the field concentrates its attention and its worry, and the term rewards a precise definition. Rotolo, Hicks, and Martin name five attributes: radical novelty, relatively fast growth, coherence as a recognizable field, prominent if still unrealized impact, and lasting uncertainty about where it goes. The category is temporary, and it resolves in more than one direction. Some technologies spread until they stop being emerging at all, the way the telephone, the computer, and the smartphone each moved from frontier to mundane infrastructure. Others draw years of attention and then recede: Google Glass pulled its consumer version in 2015, 3D television lost manufacturer support by 2017, and the Segway ended production in 2020 without the mass future it was sold on. Others settle into a lasting niche, like the fax machine that about 70 percent of US healthcare providers still use. Everett Rogers’ diffusion of innovations is the standard account of how an innovation moves through a population, from a small band of early adopters to a majority, to a niche, or to abandonment, and today’s emerging AI could still follow any of these.

A technological solution, though, is not always an emerging one. By a serious definition a paper intake form is technology as fully as a large language model is, and David Edgerton argues that histories of technology overweight invention and underweight the cheap, mundane, already-in-use things that do most of the world’s work, with corrugated iron and the bicycle outlasting the celebrated failures of their eras. Leo Marx traces how the word “technology” is barely a century old as an abstraction, one that lets people speak of technology acting on its own while the specific people and institutions responsible drop out of view, and Andrew Russell and Lee Vinsel show that the culture rewards innovation stories while the maintenance of old systems, which sustains far more of daily life, goes unfunded and unseen. The Gartner hype cycle that sorts technologies by expectation has little predictive power once its curve is checked against what technologies actually did.

In social services the consequential systems are often the settled ones. MiDAS was rules-based automation of a kind decades old. Bovens and Zouridis trace welfare administration from street-level bureaucracy, where a frontline worker held discretion, through screen-level, where the worker acts through a terminal, to system-level, where the software decides and staff handle only the exceptions, a shift that predates the current AI wave by two decades. Geoffrey Bowker and Susan Leigh Star show that a classification scheme such as the ICD, a paper standard long before it was a database, decides who counts as having a real diagnosis and therefore who can reach care, and Herd, Hoynes, Michener, and Moynihan document a 49-page mail-in Medicaid recertification form that dropped roughly 250,000 still-eligible Tennessee children when families who move often could not return the paperwork in time.

From artifact to infrastructure

The eligibility system that a caseworker’s laptop connects to is not a single tool but an infrastructure. Susan Leigh Star describes the features that set infrastructure apart: it is embedded in other systems, it is noticed only when it breaks down, and it depends on an installed base that cannot be rebuilt from scratch. Its data-sharing agreements, legacy backend, and federal reporting requirements fit that description, and the design choices inside it usually surface only when something goes wrong, in a shutdown or a wrongful denial.

The device and its algorithm

A fall sensor and the software that decides when its readings count as an emergency are often sold together as one product, but they are separate technologies. Jean-François Blanchette notes that digital information is not immaterial: every bit is stored on a physical medium. Wendy Chun describes the opposite tendency, the impression that software has no physical form, and argues that this impression is produced and draws attention away from the hardware and labor involved. The distinction matters in care settings. Clara Berridge and colleagues report that 94 percent of older adults want the ability to pause an in-home monitor, but fall sensors, GPS trackers, and cognitive-decline recorders are usually set to run continuously, a setting determined in the software, not the hardware. A device and its software have to be evaluated together, and Nick Seaver argues that an algorithm cannot be understood apart from the practices in which it is used.

Naming the good

The second word needs the same scrutiny as the first. “Social good” is not language social work generates on its own; the profession speaks of social justice, social welfare, and the common good, and this phrase came from elsewhere. Its roots run from ICT for development in the 1990s through the corporate philanthropy of Google.org in 2005. “Tech for good” took hold as a named movement in Britain between about 2012 and 2016, through accelerators and funders such as Bethnal Green Ventures and Comic Relief. “AI for good” is a later branch: the ITU’s AI for Good summit opened in 2017 and Google’s and Microsoft’s programs followed in 2018, and online interest in the phrase has climbed steeply since.

Social work took it up on its own terms. In 2015 Berzin, Singer, and Chan proposed the Grand Challenge later titled Harness Technology for Social Good, and a scoping review of the scholarship it produced reports mostly positive outcomes. AI tools are already in direct practice: Reamer’s inventory includes the Woebot therapy chatbot, the VA’s Annie app, and the Trevor Project’s crisis-simulation trainer, each with open questions about consent, privacy, and over-deference. The definitions still have gaps: reviewing how “social good” gets used, Abu-Elyounes and Gentelet find them aspirational and rarely specific about whose good is meant, set mostly by tech firms and philanthropies.

At the same time, many organizations plainly benefit from technology. TechSoup supplies discounted software to nonprofits in more than 200 countries, Code for America’s GetCalFresh has moved billions of dollars in food assistance to eligible Californians, and findhelp connects millions of people to local services. Some of this work is built by communities rather than corporations: volunteer Code for America brigades, community-run networks such as NYC Mesh, and organizing groups such as Data for Black Lives build and run technology within the communities it serves.

Key ideas

Artifacts configure their users

Every artifact acts on the person using it. James Gibson named the affordance, the relation between an object and an actor that makes an action possible whether or not anyone notices it. Donald Norman narrowed the term to the affordances a design makes perceptible, then added the signifier, the visible signal, a “Pull” label or a worn path, that tells a user where the affordance is. The distinction matters for public systems: a portal that technically permits a step but never signals it is, in practice, closed to people who cannot guess it is there. Madeleine Akrich describes how designers inscribe a script into an artifact, a scenario that casts the user in a role with assumed competencies, the way a screenplay assigns parts. A phone tree scripts its caller as English-fluent, touch-tone-equipped, and patient, and the politics surface the moment someone departs from the script. Steve Woolgar called this configuring the user: through usability trials and manuals, designers settle who counts as a legitimate user and design everyone else out. Don Ihde and Peter-Paul Verbeek describe the user’s side, where technologies mediate perception and action instead of sitting inert between a person and the world; a caseworker reading an electronic file meets the client through the system’s representation of them. Pinch and Bijker add that an artifact’s meaning is not fixed at birth, since different groups read the same object differently until social processes close the interpretation, so the line between neutral and political is itself a historical outcome.

Design and power

Langdon Winner argued in 1980 that design encodes power. His lead example was Robert Moses’s low parkway overpasses on Long Island, built, Winner said, to keep the buses that carried poor and Black New Yorkers from reaching Jones Beach. He separated cases like the overpasses, arranged by people with agendas, from technologies he called inherently political: the mechanical tomato harvester, whose adoption pushed California agriculture toward large-scale operations, and atomic weapons, which require centralized, hierarchical control no matter who owns them. A benefits portal that assumes a smartphone, a fixed address, and fluent English has made a decision of the same kind about who it serves.

The overpass story itself is contested. Joerges reviewed the record in 1999 and found that parkway design conventions predated Moses, and that tolls and competition from railroads better explain who could reach Jones Beach. He accepted that artifacts can carry politics; his objection was to letting a rhetorically powerful anecdote stand in for evidence.

Angela Glover Blackwell’s curb-cut essay runs the other direction. Design aimed at people at the margins, like the curb cuts Berkeley’s disability activists forced in 1972, tends to end up serving everyone. Her evidence includes seat belt laws, first passed to protect children, credited with saving an estimated 317,000 lives, and bike lanes, which cut pedestrian injuries by 40 percent; she estimates that closing the racial wage gap would add 2.1 trillion dollars a year to U.S. GDP. Chan traces the first curb cuts further back than Berkeley, to Kalamazoo, Michigan, where disabled World War II veteran Jack Fisher’s activism got them installed in 1945, and extends the effect to email, captions, and OCR, each built first for disabled users and later useful to everyone.

The curb-cut argument has critics of its own. Reid argues that decades of selling accessibility policy through its side benefits for non-disabled people has erased disabled people from innovation and disability-law discourse; his name for the pattern is accessibility without disability. Ruha Benjamin argues that technology reflects its creators’ values and reproduces racial inequity unless equity is actively designed in, as in a healthcare resource-allocation algorithm that favored white patients while appearing objective; her phrase is “the default setting of innovation is inequity.”

powell, Menendian, and Ake give the policy logic underneath the curb cut a name, targeted universalism: universal goals pursued through targeted processes. Their example is Chicago Public Schools, which held a single educational goal constant while tailoring implementation to the barriers different communities faced. This is not a design course, and it moves past single objects quickly.

Kranzberg’s first law

Kranzberg’s first law of technology states that technology is neither good nor bad, nor is it neutral. Melvin Kranzberg, a historian of technology, published all six laws in a 1986 article in Technology and Culture. The first law rules out dismissing a tool as a neutral instrument, and it also rules out judging one good or evil in itself. What a system does depends on the institutions around it: who deploys it, under what rules, and with what recourse when it is wrong.

Person-in-environment

In social work, person-in-environment holds that a person cannot be understood apart from their situation. This course treats databases, algorithms, and platforms as part of that environment. Technology decisions get made from a specific community’s situation and constraints.

A well-grounded decision still does not control what a technology becomes. Collingridge’s dilemma, stated in his 1980 book The Social Control of Technology, holds that a technology’s consequences are hard to predict early, while it is still cheap to change, and hard to steer late, once it is entrenched. Jasanoff’s “technologies of humility” describes a habit of appraisal that plans for that limit. She contrasts it with the technologies of hubris, the predictive methods such as risk assessment and cost-benefit analysis that governments use to reassure the public that consequences are calculable. Her alternative names four focal points for assessing any project that intends to alter society: framing, vulnerability, distribution, and learning, which she glosses as four questions: what is the purpose, who will be hurt, who benefits, and how can we know.

Technosolutionism

Technosolutionism is the habit of recasting political problems as technical ones, so homelessness becomes an app and poverty becomes a dashboard. Evgeny Morozov, who coined the term, defines it as treating complex social problems as neatly defined problems with definite, computable solutions. His example is a smart trash bin that photographs what you throw away and awards points for recycling; the user gets rewarded and the waste system stays as it was.

Not every technical remedy earns the label. Sætra and Selinger reserve “techno-fix” for a remedy that requires minimal political power and leaves social norms intact, an electric vehicle for instance, and “techno-solutionism” for the overconfidence that technology alone can transform society, as in proposals for AI-run governance or geoengineering.

Abeba Birhane’s critique of “AI for social good” traces who produces that narrative and who is missing from it. The systems marketed under the banner have a record. Indiana’s automated welfare-eligibility system denied more than a million applications and, in Virginia Eubanks’s account, broke the relationship between caseworkers and clients; her counterexample is mRelief, which lets people check whether they qualify for benefits before submitting personal data. The word “good” usually carries the most unexamined assumptions.

The amplification thesis

In Geek Heresy, Kentaro Toyama argues that technology amplifies preexisting differences in wealth and achievement. Where institutions and motivation are already strong, a new tool compounds the advantage; where they are weak, the tool changes little. His cases are One Laptop Per Child and Khan Academy, which he says failed to close the inequality gaps their backers promised, and his target is the silver-bullet mentality he found among philanthropists and tech leaders.

The Pahlka case

In 2012, Jennifer Pahlka argued on the TED stage that government services should work like good software, permissionless and open, and that small, cheaply built apps could reconnect people with their governments. The civic-tech movement she built through Code for America, founded in 2010, shipped services people still use.

GetCalFresh is the reference case. When it launched in 2014, only 66 percent of eligible Californians were receiving food assistance, fourth lowest among states. The tool simplified the online application for CalFresh, California’s SNAP program, and grew from covering half the state’s counties to all 58 by June 2019; more than seven million Californians have used it, and it carries about 73 percent of the state’s online SNAP applications. The other widely cited win was IRS Direct File, free federal tax filing built inside government: 90 percent of users rated it excellent or above average in a GSA survey, and it generated over 90 million dollars in refunds before it was discontinued for the 2026 tax season, a shutdown Forbes attributed to commercial tax-prep lobbying and political opposition.

In 2025, the delivery-and-efficiency playbook was picked up by DOGE, the federal cost-cutting office that absorbed the U.S. Digital Service Pahlka had co-founded under President Obama. MIT Technology Review reported that DOGE granted ICE access to IRS taxpayer data, planned to rebuild the Social Security Administration’s COBOL-based systems in months against expert estimates of about five years, and moved to eliminate Direct File.

Pahlka has criticized the result as “irresponsible transformation” making “large, very indiscriminate cuts” that threaten Medicaid and SNAP recipients, distinguishing illegally stopping payments Congress authorized from making an IT system work better. In her own writing she argues the mass-layoff strategy is self-defeating, since federal “last in, first out” rules cut the recent, digitally skilled hires before the long-tenured staff most resistant to change. Her preferred diagnosis, argued in Recoding America, is that government’s digital failures stem from a rigid, industrial-era culture more than from a lack of resources or technology. The same methods ran through both projects in one career.

Before class

In class

Seminar

Vote on the proposition, then a working definition exercise: we build a board of everything that might count as “technology” (a spear, a paper intake form, a spreadsheet, an LLM), sort each candidate by the scoping axes (which mode, how emerging, artifact or infrastructure), and see where the boundary stops being defensible. Winner and Blackwell get read against each other, the Pahlka case tests the neutrality claim, and we end on whose narratives dominate “tech for social good.”

Lab: the founding glossary

In small groups you draft the class’s own definitions of technology, AI, and technology for social good, each with one example that fits and one edge case that breaks it. Each group then asks an LLM to define its term, marks what the model smuggled in, and verifies one factual claim it made. Each group also trains a quick image classifier in Teachable Machine on a handful of examples it picks and runs it on cases it left out. We map the examples on a shared 2×2, post an anonymous wall of fears and skepticism, and revote. The materials are saved; Week 14 reopens the glossary, the map, and the wall.