Module 4, Dec 2, Seminar and debate lab. The plan outline and metrics are due.

Digital equity and environmental costs

Background

The Digital Equity Act, the largest federal investment in digital inclusion ever made, was terminated in May 2025 with the money unspent in most states, and the litigation to restore it is still running. The Affordable Connectivity Program had already lapsed in June 2024 after serving more than 23 million households with $30 to $75 monthly internet subsidies, and the Congressional Research Service weighed the remaining alternatives, FCC Lifeline at $9.25 a month, VA telehealth-connectivity programs, and ISP low-cost plans, and concluded that no single one would fully replace it. The Digital Equity Act’s termination did not make the underlying need disappear.

The environmental question needs its numbers read honestly. A single prompt costs very little, and Google’s own study puts a median Gemini text prompt at 0.26 mL of water and 0.03 grams of CO2 equivalent. The cost shows up in aggregate and in specific places: US data centers drew 4.4 percent of national electricity in 2023, projected to reach 6.7 to 12 percent by 2028, and roughly two-thirds of new US data centers built since 2022 sit in high-water-stress regions, where the water footprint is large and, by researchers’ own account, badly under-measured. The IEA’s April 2025 Energy and AI report opens from the premise that there is no AI without energy and models a decade of demand growth independent of any vendor’s self-report. A national average and a Memphis neighborhood tell different stories.

Key ideas

Smartphone dependence

The Pew numbers locate the 2026 divide: 91 percent of US adults own a smartphone, about 78 percent have home broadband, and 16 percent have a smartphone and nothing else. The sharpest split is by income, with 54 percent home-broadband subscription in households under $30,000 against 94 percent at $100,000 and up, and income is the most durable predictor, so this is not a generation gap that ages out. Among Hispanic adults, smartphone-dependence rose from 20 percent in 2023 to 28 percent now.

The worst-served places in the country are tribal lands. An Urban Institute study documents persistently lower broadband access there than in the rest of the country and, drawing on interviews with 15 people who work on tribal broadband, catalogs why federal money is hard for tribal governments to reach: complex applications, high project costs, and administrative burden.

The three levels of the divide

Access is only the first divide. Jan van Dijk and Kenneth Hacker separate the physical divide, who has a device and a connection, from the divides that open once people are online. Eszter Hargittai named the second level in 2002: the gap in skills and in the range of uses among people who all have access. Alexander van Deursen and Ellen Helsper named a third: the gap in the tangible outcomes people get from being online, the jobs, benefits, and services that use actually delivers. The smartphone-dependence numbers above are a first-level problem. Access without outcomes is not equity, so the third level is where this week’s question really sits.

Whether internet use narrows or widens that outcome gap is an open question, and two separate research debates bear on it. One, revisited from Week 1, runs between the social diversification hypothesis, that marginalized groups use the internet to widen their networks and reach past offline barriers, and the social stratification hypothesis, that digital tools reproduce the inequalities already in place. When Tsetsi and Rains tested the two on network data, they found internet use associated with wider gaps, not narrower ones.

The other debate comes from social psychology and concerns individual differences. The social enhancement hypothesis, the “rich get richer” claim, holds that people with strong offline networks extend them online; the social compensation hypothesis, the “poor get richer” claim, holds that people with weaker offline ties build new ones online. Zywica and Danowski gave the pair these names.

The federal retreat

The DEA put $2.75 billion into digital inclusion across three programs: $60 million for state planning, $1.44 billion for state capacity, and $1.25 billion in competitive grants, aimed at eight named covered populations: low-income households, seniors, incarcerated people, veterans, people with disabilities, people with language barriers, racial and ethnic minorities, and rural residents. Every state’s programs were planned or launched when termination letters went out in May 2025. More than 20 states sued, the National Digital Inclusion Alliance filed a separate federal suit in October 2025, and the case remains unresolved.

What the ACP subsidy bought while it ran was surveyed directly. In the FCC’s December 2023 survey of enrolled households, 68 percent said they had inconsistent or no connectivity before enrolling, and 80 percent of that group named affordability as the reason; 72 percent were using the subsidized connection to schedule or attend health care appointments, 48 percent for work or job applications, and 77 percent said losing the benefit would force them to change plans or drop internet service entirely. The benefit ended six months later.

Research on the subsidy programs adds two findings: uptake varied with community-level factors, and welfare stigma measurably suppressed enrollment even among eligible households, so a benefit can be funded and advertised and still miss people who qualify for it. The digital-navigator model that replaced much of this work has one real multi-site evaluation: a $10 million Google.org-funded pilot made 18 grants totaling about $6.6 million to rural and tribal navigator programs, and over 31 months the evaluation found the navigators effective but chronically underfunded, with lack of funding the primary barrier to keeping them and role overload putting them at risk of burnout.

What a social worker specifically does about any of this has a written answer. A policy brief in Families in Society argues that helping professionals should engage digital equity through policy advocacy, program implementation, and serving as information conveyors for clients navigating broadband access and literacy programs.

Reading environmental numbers

Google’s disclosure claims a 33x per-prompt energy improvement in one year; the numbers are also self-reported, exclude training, and draw the measurement boundary favorably. The correctives are external baselines (the IEA’s Energy and AI report, the LBNL national accounting) and peer-reviewed measurement (generative models cost roughly 30 times more energy than task-specific ones for the same job, even controlling for parameter count).

The footprint also includes hardware. Projections of generative-AI e-waste run from 1.2 to 5.0 million tonnes cumulatively over 2020 to 2030, and the same study estimates that circular-economy strategies, reuse, refurbishment, and longer hardware life, could cut that total by 16 to 86 percent. That range is the empirical case for device refurbishers like Human-I-T. Doom and greenwashing are both live failure modes in this literature.

Water and siting

Where a data center stands determines most of what its water use means. Ren and Luers calculate that one GPT-3 text output consumes about 16.9 mL of water, 14.7 mL of it through power-plant water use rather than on-site cooling, so more than 80 percent of AI’s water footprint arrives indirectly through electricity generation. Their prescription is region-specific: low-water cooling where water is scarce, decarbonization where it is not.

Siting decisions are running the other way. A SourceMaterial and Guardian investigation counted 38 Amazon, Microsoft, and Google data centers already operating in water-scarce regions worldwide, with 24 more planned, naming sites in Aragon, Spain, and Maricopa County, Arizona. Google’s single largest Arizona facility is permitted for about as much water as 23,000 residents use, and a former Amazon water-sustainability manager called the company’s water-positive-by-2030 offset pledge “not ethical.”

Infrastructure ownership

Detroit went from about 40 percent digitally included in 2019 to over 70 percent today, largely through a corporate-philanthropic coalition (Connect 313 has served 29,000+ residents since 2020). Detroit is also home to a Black-led, community-owned mesh network, DCTP’s Equitable Internet Initiative, running since 2016, which no executive order can switch off. The initiative drew national attention in a 2017 Motherboard documentary filmed when roughly 40 percent of Detroit residents lacked internet access. DCTP’s own critique names Quicken Loans and Comcast and argues that digital-inclusion work turns extractive when residents are consulted but never empowered to design the solutions.

The ownership question has a theory canon. Ostrom’s Governing the Commons, cited more than 16,000 times, documented communities governing shared resources long-term without market or state, and her polycentric-governance argument, delivered as her Nobel lecture, is the strongest case for overlapping local institutions over a single provider.

The financial evidence is contested on both sides. ITIF’s survey of 20 municipal broadband projects found 11 cash-flow negative and five with payback periods over a century; Provo sold its iProvo network to Google for $1 with the city still responsible for roughly $39 million. The Institute for Local Self-Reliance answers with a Harvard Berkman Klein Center study that found community-owned networks charging lower prices than private competitors in 23 of 27 communities examined, and counts nine of the country’s ten fastest broadband networks as municipal.

The community-ownership side has city-level outcomes with numbers attached. By ILSR’s accounting, Chattanooga’s public utility EPB returned almost $2.7 billion to the community on a $220 million fiber build over its first decade, and used the same city-owned fiber to give 17,000 households of students facing income challenges free service for ten years. The same fact sheet reports Wilson, North Carolina’s municipal network selling public-housing residents 40-megabit symmetrical service at $10 a month, and Bristol, Tennessee’s fiber saving city electric customers $6 million a year through the smart-grid hardware that rode in with it.

The xAI Colossus case

Colossus is the supercomputer that runs Grok. Its first site sits in Boxtown, a historically Black neighborhood in South Memphis with a documented history of redlining and industrial pollution, where the metro already fails federal smog standards. xAI installed roughly 35 gas turbines there; in July 2025 the Shelby County Health Department permitted 15 of them over community objection.

The second site, in Southaven, Mississippi, ran gas turbines with no air permit at all, half a mile from homes. The NAACP’s April 2026 complaint counted at least 27 of them; Earthjustice’s June 2026 release on the government’s intervention describes 59 unpermitted turbines at the same facility. Thermal drone footage published in February 2026 showed turbines still running there nearly two weeks after an EPA ruling reaffirmed permitting requirements; a former EPA enforcement chief called it a clear violation, while Mississippi regulators disputed that portable units need permits at all. Estimated annual emissions included about 1,700 tons of nitrogen oxides and 19 tons of formaldehyde, and MLK50 reports the facility is projected to emit more than 6 million tons of greenhouse gases a year, which would make it one of Mississippi’s largest fossil-fuel power plants.

In April 2026 the NAACP, represented by Earthjustice, filed a Clean Air Act suit, and in June 2026 the Department of Justice intervened on xAI’s side, asking the court to dismiss. The federal retreat from digital equity and the federal defense of the AI build-out came from one administration.

How the siting was smoothed locally is also documented. ProPublica and MLK50 reported that the Memphis Chamber of Commerce ran what it called round-the-clock concierge service for xAI and mailed residents a flyer listing nine agencies with supposed oversight authority over the facility, two of which told reporters they had no such authority. Residents have answered on the record: KeShaun Pearson of Memphis Community Against Pollution compares xAI’s promises to Electrolux’s unfulfilled jobs promises to the city, and Young, Gifted & Green founder LaTricea Adams told Capital B that “living in this ZIP code has been a death sentence for Black Memphians.” In class we run the protocol on the case: who defined the problem Colossus solves, who benefits from Grok, who breathes the exhaust, and what recourse Boxtown has.

Before class

In class

Seminar

Vote, then the divide in 2026 with the phone-only exercise (a SNAP application, a telehealth intake, or a resume upload, on a phone with a data cap), the federal retreat as a policy case, and the environmental numbers with the greenwashing critique. Track Policy’s map locates the physical data centers, Colossus among them, with each site’s power and water draw, and Hannah Ritchie’s energy comparison sets one chatbot query beside a shower, a mile driven, and an hour of streaming. Revote before the lab.