Both Companies and Students are Unequal in Ability to Use AI Productively

It is relatively hard to determine whether artificial intelligence is helping some students learn better or stunting their development of learning skills.

But that might not be unexpected. 

In the business world, higher-performing firms seemingly are better at adopting and extracting value from new technologies, including AI. 

A small minority of high-performing companies capture the bulk of AI-driven returns (top five to 20 percent account for the large majority of gains), according to one analysis. Other studies tend to agree.  

Students who are already good at standardized tests may also be better at using AI as a learning technology. 

When educational researchers evaluate AI tools (such as automated writing assistants, generative AI tutors, or adaptive platforms), they frequently report positive short-term gains in student output, task completion speeds, or test scores. 

However, critics and recent consensus reports argue that these metrics may conflate performance with learning. That’s akin to “teaching to the test,” where the objective is to improve student test scores by focusing on improving test performance, rather than other learning objectives. 

Study	Design / students	What it found	Why it matters
Contractor & Reyes, 2026 – Experimental Evidence on the Learning Impact of Generative AI	Randomized experiment with undergraduates; AI vs. no AI; unaided tests immediately and one week later	AI increased immediate knowledge-test scores by 0.27 SD, with gains persisting one week; essay effects depended on whether AI was used for augmentation versus automation	Strong evidence that AI can improve actual unaided performance, but how students use AI matters greatly. (IZA)
Fischer, Rau & Rilke, 2025/26 – AI Tutoring Enhances Student Learning	RCT, 334 university students	AI tutoring raised test performance 0.23 SD; effects were largest among students with lower baseline knowledge and stronger self-regulation	Direct evidence for your hypothesis: baseline ability changes the size of the AI effect. (Scale)
Harvard AI tutor RCT, 2025	194 undergraduates; AI tutor vs. active-learning classroom	AI group learned substantially more in less time; researchers measured pre-test knowledge and post-test learning	Shows why pre-testing matters: the relevant outcome is change from baseline, not simply the final score. (PubMed Central (PMC))
Noy et al./PNAS high-school mathematics study	High-school students using GPT assistance in math practice	Unrestricted/basic GPT assistance could reduce performance on an unassisted exam; students nevertheless believed they had learned more	Particularly important: perceived learning and measured learning diverged. (DOI)
ChatGPT cognitive-crutch RCT, 2025	120 undergraduates; ChatGPT-assisted vs. traditional study	AI group scored 57.5% vs. 68.5% on a surprise retention test 45 days later	Shows why immediate test performance can be misleading: easier performance during learning can come at the cost of retention. (ScienceDirect)
Oreopoulos et al., 2026 – AI tutoring/mastery math	>6,000 middle-school students; randomized AI vs. conventional computer-assisted learning	AI students made fewer attempts but were more accurate conditional on attempting; strongest delayed-learning evidence occurred when AI was embedded in a mastery structure	AI may alter the learning process, not simply raise scores. Structure matters. (National Bureau of Economic Research)
Oreopoulos & Low, 2026 – Khanmigo	Two-year cluster RCT in 18 Tennessee middle schools	AI-tutor assignment raised math achievement about 0.06–0.08 SD per year, with larger effects for full-year active participation	Real-world AI effects can be modest because actual usage is much lower than theoretical availability. (National Bureau of Economic Research)
Saloojee et al., 2026 – medical students	RCT, final-year medical students, ChatGPT available during clinical exams	No significant improvement in clinical performance; prior academic performance predicted scores	A useful reminder that strong students don't automatically benefit more from AI; task and implementation matter. (PubMed)
Wu et al., 2026 – meta-analysis of 35 experimental/quasi-experimental studies	Meta-analysis	Finds substantial variation in ChatGPT effects across subjects, educational levels, instructional approaches and knowledge types	The average AI effect conceals considerable heterogeneity. (Nature)
Deng et al., 2025 – meta-analysis of experimental studies	Systematic review/meta-analysis	Overall positive effects on academic performance and higher-order thinking, but also reduced mental effort	Illustrates the central ambiguity: better outcomes can coexist with less cognitive effort. (DOI)
Nickow, Oreopoulos & Quan, 2020 – tutoring meta-analysis	Meta-analysis of tutoring experiments	Tutoring produced a large average learning effect (~0.37 SD), but effects varied substantially by program/context	Important pre-AI benchmark: individualized assistance has long produced heterogeneous effects, so AI shouldn't be expected to have one uniform effect. (National Bureau of Economic Research)
Kraft, Schueler & Falken, 2024 – tutoring generalizability	Meta-analysis of 265 RCTs	Effects fell to roughly one-third to one-half of the original estimates when studies were restricted to settings resembling large-scale standardized-test interventions	A major warning about extrapolating impressive experimental effects to real-world populations. (ERIC)

Up to a point, AI might help learning by minimizing friction. On the other hand, learning arguably often requires encountering friction and overcoming it. So if AI minimizes cognitive struggle, it might also negatively affect learning. 

The other unavoidable problem is that some students are just better equipped to benefit from AI tool use. Students who are already high performers, highly self-regulated, or technologically fluent tend to extract more value out of AI tools. 

Struggling students might rely too much on “getting the answers” without developing thinking, research or other skills. In other words, “output” might not reflect student learning so much as AI answers. 

And it might be a reasonable assumption that better-equipped learners will also tend to be those that learn most when using AI as well. 

Teachers might agree that an in-class essay exam using blue books is a more-reliable test of what a given student might know, though obviously also favoring better writers, compared to any out-of-class essay, which can be a better test of what a given AI engine knows and expresses. 

The point is that better-performing students are also likely to be better-performing users of AI. 

Study / Report Focus	Authors / Organization	Key Finding Regarding Performance vs. Learning	Source Link
Children’s and Adolescents’ Learning with Educational Technology	American Psychological Association (APA) (2026)	Warns that metrics like time-on-task, completion rates, and immediate output are frequently mistaken for learning. Highlights that AI can improve immediate performance while simultaneously reducing long-term independent knowledge retention.	APA Report on Educational Tech
Measuring Student Trust and Over-Reliance on AI Tutors	Educational Cybernetics and Studies (2025)	Investigates how moderate trust improves confidence, but over-trust triggers automation bias—where students accept AI-generated answers without critical evaluation, inflating performance metrics while depressing deep cognitive skill-building.	ECSE Article
Assessing the Quality of AI-Generated Exams: A Large-Scale Field Study	ArXiv Educational Technology Research (2025)	Evaluates AI-generated assessment items across ~1,200 students using Item Response Theory (IRT), establishing that while AI can efficiently construct valid psychometric tests, student baseline engagement heavily dictates performance outcomes.	arXiv Study
Improving Student Learning with Hybrid Human-AI Tutoring	Educational Research Working Paper Series (2023–2026)	Explores the nuance of baseline abilities, finding that while lower-achieving students show measurable proficiency gains with hybrid support, unmonitored independent AI usage risks widening the gap due to varying digital literacy and self-regulation skills.	arXiv Working Paper

I’m not sure how we compensate for learning prowess in general, anymore than we seem to systematically produce high-performing firms.

How Much "Lost Economic Impact" from 5G Spectrum Licensing Delays?

A study prepared for ComReg suggests delays in awarding additional 5G spectrum have caused as much as €1.06 billion to €2 billion in lost economic benefit to Ireland. That estimate involves extrapolating from other studies of 5G economic benefit. 

source: ComReg 

To be sure, it is reasonable to assume that delayed spectrum awards also mean delayed construction projects, which, by definition, provide some temporary economic boost as the work is done. And if one believes 5G services boost average revenue per unit, then some losses might be expected on that score. 

But it is complicated. Not all the activity occurs locally, or in Ireland, so there is “leakage.” Also, there is “substitution.” Even if one assumes some 5G customers spend a bit more than they did for 4G, the net changes might not be so large. When a “new 5G customer” also was an existing 4G customer, a 4G account is “lost” as the 5G account is added. 

The net change is not as great as the raw numbers might indicate. And that is not a novel issue. One example is the estimated economic impact of new sports stadia. 

Economic impact studies on the impact of 5G, home broadband or artificial intelligence are always based on assumptions various observers will disagree about. 

One big issue is the necessity of qualifying every forecast with the caveat that it implicitly or explicitly assumes “all other things remain equal” or unchanged. Rarely does anything important remain “unchanged” when other big changes happen. 

But without assuming away all those changes, analysis would be impossible. 

For example, if one added up all the economic benefit estimates from all studies in a single nation, in a single year, from every industry and all investments, the total would clearly exceed total economic output by a substantial margin. 

Perhaps each participant in a value chain--such in car production--each adds value to a complete car, but cannot each claim the full economic value of the car produced. 

When many industries contribute to an examined area of economic growth, one ends up “double counting” output when each contributing industry claims the economic boost is entirely from its own efforts. 

In other cases, even claimed “growth” might simply be “substitution.”

Consider the example of economic benefits from municipal funding of sports venues and stadiums. One always sees estimates of revenue generated by such investments:

• Coates, D., & Humphreys, B. R. (2008). The growth effects of sports franchises and events. Journal of Regional Science, 48(4), 697-718.

• Rosentraub, M. S. (1999). Major league losers: The real cost of sports stadiums and arenas. Brookings Institution Press.

• Wenner, L. A. (2000). Sports economics: A survey of the literature. Journal of Sports Economics, 1(1), 1-31.

On the other hand, rival studies suggest there is no net benefit:

• "The Economic Impact of Sports Franchises: A Critical Review of the Literature" by Dennis Coates, Brad Humphreys, and Andrew Zimbalist (2006)

• Baade, R. A., & Matheson, V. A. (2003). The economic impact of sports teams and facilities. Journal of Economic Perspectives, 17(3), 115-132.

• Coates, D., & Humphreys, B. R. (2002). The economic impact of professional sports teams and facilities: A critical review. Journal of Economic Policy Reform, 5(1), 1-24.

• Noll, R. G. (1974). The economic effects of professional sports leagues. Brookings Institution Press.

For all such reasons, it is difficult to say much about what delays in licensing 5G spectrum might actually mean, in terms of economic output.

Next Normal or New Normal?

The post-Covid business environment--for connectivity providers as much as for any other industry--might be “next normal” or “new normal.” The former might indicate bigger changes for some industries but fewer disruptions for others, essentially accelerating trends already present.

The latter might indicate fairly permanent and significant life alterations that were not already in place. “Just different” is one way of describing “next normal,” while “never be the same” characterizes “new normal.”

McKinsey consultants analyzed the change potential across more than 2,000 tasks used in some 800 occupations in the eight focus countries.

“Considering only remote work that can be done without a loss of productivity, we find that about 20 to 25 percent of the work forces in advanced economies could work from home between three and five days a week,” McKinsey says. 

“This represents four to five times more remote work than before the pandemic and could prompt a large change in the geography of work, as individuals and companies shift out of large cities into suburbs and small cities,” McKinsey notes.

The geography of communications also should shift, with some possible ramifications. Pre-Covid, mobile traffic demand was generally concentrated at 30 percent of cell sites. That should lessen, with a greater percentage of traffic at the other 70 percent of sites.

That also should alleviate some capital investment intended to boost radio capacity at the busiest urban sites. Additionally, mobile data demand could grow less rapidly than before the pandemic, if significant percentages of workers stay at home, more of the time, connected to their Wi-Fi networks. 

Less urban traffic also means less demand for all the associated businesses catering to urban workers. That suggests less demand for small business connectivity, for some time, or perhaps even permanently.

Less demand for urban office space will shrink the market for business connectivity supplied to those locations. There eventually should be higher demand for at-home upstream bandwidth as well, as more people need better support for two-way video sessions. 

Use cases for fixed wireless should improve, as the suburban and rural locations more people will be working from are precisely those locations where sustainable business cases for new fiber to home installations are toughest. 

There will unpleasant social implications, as low-wage, lesser-skilled jobs are among those which will be displaced, post-Covid. As has been the case for decades, the growth will be happening in health care and knowledge work. 

source: McKinsey 

“Compared to our pre-Covid-19 estimates, we expect the largest negative impact of the pandemic to fall on workers in food service and customer sales and service roles, as well as less-skilled office support roles,” McKinsey says. 

“Jobs in warehousing and transportation may increase as a result of the growth in e-commerce and the delivery economy, but those increases are unlikely to offset the disruption of many low-wage jobs,” McKinsey adds. 

“In the United States, for instance, customer service and food service jobs could fall by 4.3 million, while transportation jobs could grow by nearly 800,000,” McKinsey notes. “Demand for workers in the healthcare and STEM occupations may grow more than before the pandemic.”

On the other hand, it also is fair to ask whether travel, hospitality and some forms of business travel will be permanently depressed. A year ago, some quipped that “trade shows are dead,” as a permanent trend. To be sure, all large in-person events were temporarily halted, for health reasons. 

But we need to be careful about extrapolating present circumstances into the future. The pandemic will pass. And we can be sure that any linear extrapolation from pandemic behaviors will prove incorrect. 

Gradually, demand for experiences provided by in-person events will return. For business-to-business sales operations, they will be almost necessary. 

“We found that some work that technically can be done remotely is best done in person,” McKinsey says. “Negotiations, critical business decisions, brainstorming sessions, providing sensitive feedback, and onboarding new employees are examples of activities that may lose some effectiveness when done remotely.”

Extrapolating Remote Work Trends from Immediate Circumstances is Likely Not Wise

Some of us have been hearing predictions about the growth of remote work (it used to be called telecommuting) for four decades or so. And while there have been secular changes, it is difficult to make a case that anything really has changed the adoption curve of full remote work, even if lots of people take some work home from the office, routinely. The underlying trends are what they are, and might get something of a boost, but that might be hard to detect.

A Gartner survey of 229 human resources leaders finds execs now believe more remote work will be done by their employees, post pandemic. “While 30 percent of employees surveyed worked remotely at least part of the time before the pandemic, Gartner analysis reveals that post-pandemic, 41 percent of employees are likely to work remotely at least some of the time,” said Brian Kropp, Gartner HR practice chief of research. 

What all that means is not yet clear, as the definitions of remote work vary widely. Some of us might consider remote work to be “employees who are based full time at remote or home locations.” 

Others might include employees who work remotely at least half the time. That is a very small number of people, at the moment, perhaps as few as 3.6 percent of the entire workforce, by some estimates. 

The number of U.S. employees working at home 50 percent of the time or more in 2020 is estimated at five million, representing 3.6 percent of the workforce, according to Global Workplace Analytics. And that is after 40 years of evangelization that some of us are personally aware of. 

But most people likely take a broader view of remote work, including some work from home days each week or month. 

In the past, “telecommuting” has generally been thought of as employees working “at home” sometimes--or full time--instead of at the office, campus or plant. That sort of thing might not differ much from workers occasionally or even routinely bringing some work home from the office. 

One way of setting a reasonable universe of potential remote work is to evaluate the total number of jobs that conceivably could be done entirely remotely. By some estimates, only a third of jobs can be done remotely, according to a study conducted by professors Jonathan Dingel and Brent Neiman of the University of Chicago Booth School of Business. 

The study suggests 34 percent of U.S. jobs can plausibly be performed at home. Assuming all occupations involve the same hours of work, these jobs account for 44 percent of all wages. The converse is that 66 percent of jobs cannot plausibly be shifted to “at home” mode. 

If we assume that most people will consider “working from home” sometimes as a valid case of remote work, the universe of jobs appears to be close to 34 percent, looking at jobs that can be completely remote, full time. Using less stringent definitions would produce a higher number, but the value of such estimates might be questionable. 

It is not clear that the actual requirements of remote work, done on a casual or occasional basis, actually include much more than having a smartphone, a PC and adequate internet access at home, plus the standard cloud computing apps typically used in an office. 

More specific computing tasks, requiring sophisticated equipment (robots or industrial or process machinery) are not the sort to be done at home on a casual basis. 

To be sure, some executives will look to reduce spending on office facilities by shifting some work to full remote status, while allowing others to work substantially from home. But technology is not the only issue. Managers must trust that worker productivity remains substantially the same when work moves remotely. 

But recall that similar predictions were made in 2009 when the HiN1 virus outbreak happened. It is by no means clear that some non-linear acceleration of remote work trends happened after that, and was sustainable. 

Government Broadband Policy Too Often Ignores Moore's Law

Government planners often are too optimistic about what their proposed programs can achieve. In the case of broadband, they have tended to be too modest. The U.K. government launched in 2010 an effort to enable superfast internet access across the country. Keep in mind that a year earlier, the government said it wanted a 2 Mbps minimum speed across the country. 

In 2011 the goal goal was bumped up 24 Mbps per household by about 2015. To be sure, there is a difference between a minimum floor and a maximum aspiration. But past experience with speed increases--even in 2010--should have prompted lawmakers and policymakers to aim higher. 

Speeds increase at Moore's Law rates, one can argue, at least for some suppliers, such as the cable companies. 

Comcast has doubled speed every 18 months, for example. In 2010, typical Comcast speeds already were up to 100 Mbps. Few customers bought the fastest-available service, of course. But the minimum speed of about 12 Mbps grew to about 50 Mbps by 2015. Using the Moore’s Law doubling in 18 months would have produced speeds in excess of 100 Mbps by 2015, which is what happened. 

source: Comcast

This example from the Australian National Broadband Network actually is too conservative. Extrapolating from 1985, it suggests typical internet access speeds “should” have grown from about 10 Mbps in 2009 to perhaps 100 Mbps by 2015. 

source: NBN

When at least some suppliers are doubling speeds every 18 months, most targets and goals set by government are going to be eclipsed very quickly, no matter how ambitious the goals seem at the moment.

The point is that although government goals will tend to focus on minimums, as for universal service, aspirational targets need to incorporate what we know about Moore’s Law and its application to internet access bandwidth. 

With or without any specific government policies (other than staying out of the way), typical and minimum speeds would double about every 18 months to 24 months. So, one might argue, the U.K. government goal quickly was surpassed by commercial supply that did, in fact, increase at Moore’s Law rates, as did computing.

Most rational observers would have argued that physical networks could not improve speed so fast, as labor intensive and capital intensive as outside plant remains. Perhaps few thought Moore’s Law  rates of progress were possible for outside plant. On the other hand, few probably believed Moore’s Law would apply to computing hardware, either. 

The most-startling strategic assumption ever made by Bill Gates was his belief that horrendously-expensive computing hardware would eventually be so low cost that he could build his own business on software for ubiquitous devices. .

How startling was the assumption? Consider that, In constant dollar terms, the computing power of an Apple iPad 2, when Microsoft was founded in 1975, would have cost between US$100 million and $10 billion.

source: Hamilton Project

The point is that the assumption by Gates that computing operations would be so cheap was an astounding leap. But my guess is that Gates understood Moore’s Law in a way that the rest of us did not.

Reed Hastings, Netflix founder, apparently made a similar decision. For Bill Gates, the insight that free computing would be a reality meant he should build his business on software used by computers.

Reed Hastings came to the same conclusion as he looked at bandwidth trends in terms both of capacity and prices. At a time when dial-up modems were running at 56 kbps, Hastings extrapolated from Moore's Law to understand where bandwidth would be in the future, not where it was “right now.”

“We took out our spreadsheets and we figured we’d get 14 megabits per second to the home by 2012, which turns out is about what we will get,” says Reed Hastings, Netflix CEO. “If you drag it out to 2021, we will all have a gigabit to the home." So far, internet access speeds have increased at just about those rates.