"Digital Divide" is Closing

Policy advocates and policymakers have worried about a "digital divide" in U.S. Internet usage, as much as global policymakers have worried about the difference between communications use in developed and developing regions.

But a new study by eMarketer suggests that the U.S. digital divide is closing fairly rapidly. By 2014, in four years, Internet usage rates by Americans of black ancestry will just about equal rates of U.S. "whites" today, while Hispanic American use of the Internet will rise to within six percentage points of the current U.S. average usage by "white" Americans.

That is not to say rates will be identical, but the point is that almost nobody thinks "white" Americans generally are victims of a "digital divide" today, though there are more issues in rural or isolated parts of the country. In fact, most of the non-adoption factors now are of a "demand" sort rather than a "supply" sort. In other words, most people who want broadband already buy it.

If by 2015 Americans of "black" or "Hispanic" heritage have those same rates, the significance of the "divide" should be largely moot. That is not to say the issue is completely moot, but closing the last percentage or two of gap in any endeavor always is a matter of effort and reward. And since the primary issue these days is demand, not supply, some circumspection might be in order, in terms of the amount of effort expended, compared to the potential benefits. The markets, and consumers, seem to be doing a relatively good job, unaided, in terms of closing the digital divide.

"AI Divide" Might Not Really Matter

As always seems inevitable, new digital technologies raise concerns about “digital divides.” And so it already is with artificial intelligence. 

To be sure, we might argue the natural outcome of competition is concentration, especially in industries where capital investment is a requirement. That arguably means AI frontier models, built on a foundation of expensive high-performance computing hardware, will favor contestants with lots of capital access. 

And that might apply not only to firms but to countries. But maybe we should not be surprised. In the internet era, we saw the development of “winner take all or most” market structures in most application areas. And there are clear reasons to believe this precedent will continue in the AI era as well. 

And that has implications for the “AI divide” thesis. 

AI Divide Concerns
Source	Publication Date	Key Findings	Implications
UNCTAD Technology and Innovation Report 2025	April 3, 2025	AI market projected to reach $4.8 trillion by 2033, but benefits may concentrate in wealthy nations. 118 countries, mostly from the Global South, are absent from AI governance.	Highlights risk of global inequality and need for equitable access to AI tools and infrastructure.
The Emerging AI Divide in the United States	April 17, 2024	High ChatGPT search volumes in U.S. West Coast; low in Appalachia and Gulf states. Education is the strongest predictor of AI awareness.	Indicates spatial and socioeconomic divides in AI adoption, reinforcing existing digital disparities.
AI Is Deepening the Digital Divide	November 30, 2023	AI exacerbates digital divide, excluding billions. Global North-South gap widens due to lack of access, skills, and participation.	Calls for inclusive education, stakeholder engagement, and ethical AI strategies to mitigate divide.
How to Bridge the AI Divide	August 3, 2023	AI amplifies economic inequality, with high-skill workers gaining while low-skill workers face displacement. Geographic disparities worsen.	Emphasizes need for equitable policies to distribute AI benefits and prevent economic downturns in automation-prone regions.
The Current State of the AI Market: The AI Divide	April 15, 2025	AI market to reach $1.81 trillion by 2030, driven by U.S. and China. Developing nations struggle with infrastructure and talent shortages.	Urges localized AI development and global cooperation to prevent concentration of AI benefits.
AI for the Global Majority	February 21, 2025	2.6 billion people lack internet access, limiting AI benefits. Biased datasets exclude Global Majority communities.	Stresses need for inclusive AI design and digital literacy to prevent deepening socioeconomic divides.
PwC’s Global Artificial Intelligence Study	June 25, 2017	AI to boost global GDP by 14% by 2030, with 70% of gains in China and North America, <6% in developing regions.	Warns of unequal distribution of AI’s economic benefits, necessitating strategic investments.
The Rising Costs of AI	December 7, 2024	AI development costs create a “new rentierism,” locking advanced AI behind paywalls, deepening inequality.	Advocates for EU-led efforts to democratize AI access and foster innovation ecosystems.

Network effects, near-zero costs of content reproduction, global reach, data flywheels and platform roles explain the “winner take all or most” structures. 

In digital platforms, the value of the network increases with each additional user, creating positive feedback loops that favor a few firms with the largest networks. 

Software and digital services can be replicated at low to virtually “no” cost, so a firm deemed to have the “best” product can ride user preferences to leadership. 

The internet dissolves geographic barriers, so leading firms can scale globally without proportional increases in cost.

Products which benefit from network effects (especially social and user-generated-content networks) also find that having more users that generate more data leads to better products which leads to more users. 

Finally, some companies became platforms, enabling third parties to conduct transactions, spurring their growth (Apple's App Store, Amazon's Marketplace).

AI might follow a similar path as well. AI models improve with more and better data, so entities owning or controlling massive datasets can build better-performing models, attracting more users, yielding even more data. It’s a flywheel effect. 

Also, AI model development is capital intensive in a way that internet apps have tended not to be. 

Training state-of-the-art models requires massive compute power and engineering talent. Only a few firms can afford this.

Also, AI frontier models are likely to eventually emerge as something like operating systems, creating whole ecosystems built around them. That, of course, will lead to a “brand preference” as well. 

All of which will inevitably create an unequal market structure and concern about an “AI divide” that could affect individuals, firms, industries and countries. 

On the other hand, we are early in the process, so there typically also is a lag between the time an important new technology, such as general-purpose technology begins to be introduced, and the time when we can clearly see the impact. (GPTs are technologies with pervasive applications, high productivity potential, and continuous improvement that affect most to all of an economy). 

And we are likely to see uneven benefits and outcomes, early on, reinforcing concerns about “divides.”

Economist Robert Solow’s 1987 quip that “you can see the computer age everywhere but in the productivity statistics,” reflects the fact that entities need time to redesign their business processes to take advantage of new GPTs. It wasn’t until the late 1990s when PC-influenced productivity surged, for example. 

Likewise, the steam engine took decades to significantly boost industrial output, and electricity’s full economic impact wasn’t realized until the early 20th century, nearly 50 years after its invention.

AI, often considered a GPT due to its broad applicability across sectors like healthcare, finance, and manufacturing, is in its early stages. Despite rapid advancements, its economic contributions presently might be called  modest compared to projections. 

The point is, we will be hearing lots about AI divides. It’s sort of inevitable. 

On the other hand, there is room to question whether an “AI divide” actually has serious consequences. 

Even when a few firms dominate a technology category, widespread benefits can still accrue to individuals, firms, industries, and nations. The dominance of providers does not inherently preclude others from capturing significant value—as long as access to the technology is available and usable.

Value creation by people, firms, industries and nations is different from “value capture” within a technology segment. In other words, Microsoft productivity tools might be dominant within global enterprises. But that never prevents user firms from wringing value out of the tools they do not create or own. 

The same arguably goes for platforms that many third parties can take advantage of. Taken all together, productivity increases can happen broadly even when dominant suppliers of tools exist. The analogy might be electricity, which often is a monopoly in a given area, but all people, businesses and entities can use it to create value. 

The further point is that “AI divides” might not exist too long, and that supplier leadership, in any case, does not imply general societal or economic inability to wring value out of the tools.

Digital Divide Now is More Subtle

There is no digital divide on inter-city trains, inter-city buses or airplanes, a study of use of personal devices on buses, trains and airplanes suggests.

Internet access remains an issue in rural and lower-income areas, compared to suburban and mid-income urban areas, to be sure, but the issues now are more subtle, having as much to do with people not seeing Internet access as useful as actual physical lack of access.

To be sure, use of fixed high speed access services is lower among households with less income. Access is 70 percent amongst households with $10,000 or less annual income, in the 85-percent range for households with income between $20,000 and $40,000, and above 90 percent for households in higher income ranges.

But age explains non-use of the Internet as much as income. Also, mobile Internet access is substantial among lower-income households, ranging from 50 percent to 60 percent among lower-income groups.

In fact, many younger users use mobile Internet access, rather than fixed network access. In other words, much of the digital divide that remains in U.S. Internet access is explained by age or use of mobile access.

Access speeds in rural areas continue to lag offered speeds in urban and suburban areas, as a rule, though the gap is closing, as cable TV high speed access services tend to be much faster than all-copper digital subscriber line connections. That is true even in India.

That is one good reason why AT&T, among others, is upgrading rural networks with fiber.

In fact, widespread use of connected personal devices on inter-city transportation services suggests the important role ownership of connected devices now plays.

On Greyhound inter-city buses, the use  of personal technology use is now significantly higher than on airplanes and is only marginally below that on Amtrak and discount bus lines, a study by the Chaddick Institute for Metropolitan Development has found.

In fact, for the first time in five years, use of personal devices on at least one  inter-city bus service was higher than on airplanes or Amtrak.

Among the 505 passengers observed on 20 Megabus and Van Galder buses operating from curbside locations in 2013, 59 percent were using technology, compared to 46 percent in 2012.

In large part, that might be because the amount of use of new “connected” bus services--which offer travelers uninterrupted cell phone signals as well as free Wi-Fi and power outlets--grew 30 percent between 2012 and 2013.

On Amtrak, the share of technology users was flat at 52 percent in 2013, the study found.

Availability of power outlets, Wi-Fi and mobile access likely explains the lighter use of personal devices on airplanes, according to the Technology in  Intercity Travel Study.

Technology use on airlines remained virtually flat and continues to lag behind other

modes in 2013, suggesting that lack of communications “for no incremental cost” is an issue.

But the ban on phone calls aboard aircraft, as well as the lack of power outlets, likely also are issues.

The two fastest growing modes of intercity travel over the calendar years 2012 and 2013—intercity trains and discount buses—were also those in which the technology use was observed to be the highest in early 2013.

The amount of discount bus service grew by four percent between 2012 and 2013, while the number of Amtrak seat-miles grew by 1.4 percent, as did airline seats.

Availability of Wi-Fi and mobile Internet connections, the “no incremental cost” access and lawfulness of device app use on trains and buses possibly explains the higher use of personal devices on buses and trains.

Mobile device connections are disabled in the air, on airplanes, in addition to being unlawful. When Wi-Fi is available, usage requires payment, and power outlets often also are not available.

But there seems to be no “digital divide” between passengers on inter-city buses, trains or airplanes.

The Chaddick Institute survey in 2014 consisted of 1,659 airline travelers, 1,608 intercity train (Amtrak) passengers, 505 discount city-to-city bus passengers (Megabus and Coach USA), 270 conventional intercity bus passengers, and 2,992 commuter rail passengers.

No Digital Divide?

One doesn’t hear quite so much about “digital divides” as was common a decade ago. One suspects that is because the supply of communications, both voice and data, is a problem the world seems capable of solving.

Some of us can remember great "handwringing" an concern in international policy circles about how to bring telephone service to two billion people who never had made a phone call. You don't hear such concern anymore, since we rapidly are solving that problem with mobile communications, a solution not envisioned in the 1970s and 1980s.

Two decades ago the question largely had shifted to the problem of how to get computing into the hands of the next three billion people. There was some work around the notion of special devices optimized for rural villagers that would be low cost, perhaps $150 or so.

For many at the time, likely most knowledgeable observers, the prevailing thinking was that it couldn't really be done. And that remained true even as recently as the middle of the 2000 decade.

But as we stumbled upon a solution to the problem of getting communications to people at prices they could afford, we are about to solve the problem of getting computers to people, also at prices they can afford.

The notion, for some time, has been that in many parts of the world, the smart phone would be "the computer" most people used. That might turn out to be largely correct, for at least a time.

But it also now is possible that we know how to create and sell computers to people that cost no more than $150. Consider that the prototype "One Laptop Per Child" device had a screen of 7.5 inches diagonal and flash memory, with no keyboard.and used Wi-Fi for Internet connectivity.

Oh, that's right, we now call that a tablet, and it is made and sold commercially by the likes of Amazon, Barnes & Noble, and soon Google and likely Apple.

One might note a similar process at work even in the area of rural broadband access. Does it make sense to spend up to $50,000 per home to provide broadband access service to less than 200,000 U.S. rural locations, when at least two other approaches are already available?

The Federal Communications Commission, for example, conducted a  gap analysis that suggested $13.4 billion in subsidies would be required to  expand availability to only 250,000 of the highest cost homes (0.19 percent of all U.S. homes).

According to the FCC, those homes would require subsidies of about $53,600 – on top of what service providers would expect to spend to connect a typical home.

Excluding the cost of serving these 250,000 homes, the cost of connecting the remaining 6.75 million homes would entail a subsidy of about $1,500 per home passed.  

Keep in mind that the “broadband gap” affects less than five percent of U.S. homes. The Federal Communications Commission itself estimates that seven million U.S. households do not have access to terrestrial broadband service, representing about 5.4 percent of the 129 million U.S. homes.

Additionally, analysis assumes those households actually are occupied, but they are not. Some percentage is unoccupied, and some are used only partly as vacation homes.

According to the U.S. Census Bureau, about 18 million of the 130 million units are not occupied (about 14 percent).

On average, the gap estimated by the Commission is $3,357 per home passed, note Dr. George Ford, Phoenix Center chief economist, and Lawrence J. Spiwak, Phoenix Center president.

Even the then director of the  National  Broadband Plan, Blair Levin, said it will be too expensive to  provide service to the last two percent of home using terrestrial facilities. Therefore, those homes should be served by satellite broadband.

A more reasonable approach to satellite broadband at the time might have been that if it costs $50,000 to provide a 4:1 Mbps terrestrial broadband service to a household, then is it reasonable to accept a “lower” service level by a network that already reaches those locations?

The situation has also changed since that analysis. ViaSat’s “Exede” satellite broadband service already has been offering speeds up to 12 Mbps downstream and up to 3 Mbps upstream, for $50 per month, since early 2012.

The HughesNet service, which has launched a new satellite of its own, will begin offering faster service beginning this month. Since both the ViaSat and HughesNet services use exactly the same satellites, it would be reasonable to assume that HughesNet will offer speeds comparable to that of Exede.

In fact, the National Broadband Plan explicitly recognized that the cost of ubiquitous coverage of terrestrial broadband could not be justified and furthermore recommended the use of “satellite  broadband” as an alternative, as it is  ubiquitously available, Phoenix Center argues.

The cost picture has changed dramatically since the FCC conducted its gap analysis. Though the original plan called for a 4 Mbps capability, Exede already sells a 12-Mbps service for $50 a month. As of Aug. 13, 2012, Hughesnet has not announced firm pricing and speeds.

But there is no reason to believe HughesNet will offer speeds any less than offered by Exede. The point is that by spending an abundance of money, the government simply does not make sense at the margin.

The point is that one has to be careful about interpreting Internet or broadband usage trends. In past years, some observers have argued there is a significant “digital divide,” using statistics about Internet or broadband usage (people who buy and use that product or capability) rather than broadband or Internet “availability” (the service actually is available to purchase).

A decade ago, access mostly was measured by activity occurring only on the fixed networks, as well.

The point is that the difference between whether a product is available, and whether a consumer chooses to buy it, is highly significant. A “digital divide” argument assumes that a product such as broadband access is statistically “unavailable,” meaning there is an “access to the product” problem. Lower availability in rural or sparsely-populated areas is one form of the argument.

The latest study of mobile broadband behavior shows the relevance of the distinction between “available to buy” and “I want to buy it.”

About 17 percent of mobile phone owners do most of their online browsing on their phone, rather than a computer or other device, the Pew Center Internet & American Life Project reports.

Young adults and non-whites are especially likely to use their mobile phones for the majority of their online activity, the researchers say.

Nearly half of all 18 to 29 year olds (45 percent) who use the Internet on their mobile phones do most of their online browsing on their mobile device, the study found.

Half (51 percent) of African-American mobile Internet users do most of their online browsing on their phone, double the proportion for whites (24 percent). Two in five Latino cell internet users (42 percent) also fall into the “cell-mostly” category.

Digital marketing specialist Troy Brown, president of one50one thinks Hispanics generate 50 percent of their Internet traffic from mobile phones.

The important implications here are that people might prefer to use mobile Internet and mobile broadband, rather than fixed access, in numbers that are significant.

Additionally, those with an annual household income of less than $50,000 per year and those who have not graduated college are more likely than those with higher levels of income and education to use their phones for most of their online browsing, Pew researchers say.

In that sense, the adoption of mobile broadband might be a rational choice, similar to adoption of mobile “as voice,” when people choose to use a mobile exclusively for their voice service. In fact, we might already have reached the point where further growth of fixed network broadband connections is limited or slowed because users have better mobile alternatives.

The Pew data also suggests that mobile broadband has been particularly important to populations that in the past have “under-indexed” for use of fixed network broadband. The point is that, aside from buyer preferences being at play, fixed network broadband purchases increasingly are not suitable measures of “broadband adoption.”

When 17 percent of mobile users report “mostly” using their mobiles for Internet access, any metrics of adoption that ignore mobile access will be misleading.

On a larger level, one might make the argument that the digital divide, and the communications divide, simply are problems we are solving.

AI Will Transform at Least 25% of Jobs, and That's Likely Way Too Conservative

A study by the International Labor Organization suggests artificial intelligence will transform about 25 percent of jobs. That might be a low estimate. 

Consider the impact of personal computers and digital skills. A 2023 report by the National Skills Coalition, in partnership with the Federal Reserve Bank of Atlanta, found that 92 percent of jobs analyzed required digital skills. 

The Brookings Institution analyzed changes in the digital content of 545 occupations (covering 90 percent of the U.S. workforce) between 2002 and 2016. They found a significant shift. 

In 2002, 56 percent of jobs required low digital skills. By 2016, this dropped to 30 percent.

The share of jobs requiring high digital skills jumped from five percent in 2002 to 23 percent in 2016.

Jobs requiring medium digital skills rose from 40 percent to 48 percent.

By 1993, nearly half of all U.S. workers were operating computer keyboards at work, a steady increase from 25 percent in 1984.

So the notion that AI will only affect 25 percent of jobs seems quite low. 

Study	Year(s) of Research/Publication	Key Findings Regarding PC/Computer Transformation of Jobs	Percentage of Jobs Transformed/Affected (where specified)
Autor, Levy, and Murnane	2003 (and later related works)	Introduced the "routine-biased technological change" (RBTC) framework, showing how computers automate routine tasks (cognitive and manual), leading to job polarization (growth at high and low ends of the skill spectrum).	Implied significant transformation across jobs with routine tasks.
Autor, Katz, and Krueger (NBER Working Paper No. 5956)	1997	The computer revolution explains a substantial portion (30-50%) of the increasing wage gap between college-educated workers and those with less education since the 1980s. Industries with high computer use reorganized work to disproportionately employ more educated workers.	Significant impact on the skill premium; "nearly half" of workers used computer keyboards by 1993.
Frey and Osborne (Oxford University)	2013	Analyzed 702 occupations, classifying them by susceptibility to computerization. Found that jobs with tasks requiring perception & manipulation, creative intelligence, and social intelligence are less likely to be automated. Identified "bottlenecks" to automation.	47% of jobs are at "high risk" of being computerized.
Bessen (Scholarly Commons at Boston University School of Law)	2015	Argued that computer use is associated with faster employment growth in occupations that use computers more, even routine and mid-wage jobs. Emphasized that automation often augments labor and leads to job reallocation and skill changes, rather than net job loss.	Computer use associated with ~1.7% increase in employment per year at sample mean.
National Skills Coalition with Federal Reserve Bank of Atlanta	2023	Analyzed 43 million online job postings. Found overwhelming demand for digital skills across nearly all industries and occupations, including entry-level. Highlighted a significant "digital skill divide" where many workers lack foundational digital skills.	92% of all jobs analyzed required digital skills (47% "definitely digital," 45% "likely digital").
Brookings Institution	Analysis up to 2016	Examined the digital content of 545 occupations. Found a significant increase in jobs requiring high digital skills and a decrease in those requiring low digital skills.	Share of jobs requiring high digital skills jumped from 5% (2002) to 23% (2016).<br>- Low digital skill jobs decreased from 56% (2002) to 30% (2016).
Dillender and Forsythe (NBER Working Paper 29866)	2022	Investigated the impact of computerization on office and administrative support jobs. Found a modest positive effect on wages and employment in local labor markets, though overall employment in office support fell. Increased skill levels needed for these positions.	Administrative support share of employment returned to 1950s levels by 2019 after peaking in 1980s.
McKinsey Global Institute	2017 (focusing on impact through 2030)	Explored how automation technologies (including AI and robotics, building on PC foundations) will change or replace jobs. Argued that while some jobs will be displaced, many more will be changed, and new ones created.	60% of occupations have at least 30% of constituent work activities that could be automated.<br>- Estimated 75 million to 375 million workers (3-14% of global workforce) will need to switch occupational categories by 2030.