Two Language Model Cases Decided This Week Bear on "Fair Use" of Copyrighted Material for Training

In the second of two “fair use” court cases involving language model training this week, a court has ruled that Meta did not infringe copyright when it used copyrighted book material to train its models. The latest decision, though, does turn on procedural grounds, not the core issue of whether copyrighted material is fair use under all circumstances, perhaps most notably when permission to use the materials for this purpose is absent. 

 

The prior decision upholding fair use of copyrighted material for model training was won by Anthropic. 

That copyright case also tested fair use law as it applies to reading books, which is essentially what model training involves. For humans, reading books is considered fair use, and not a copyright violation. It has been unclear whether an artificial intelligence model, reading books, also is fair use, or not. 

The case partially validates the idea of fair use for training purposes, without compensation or permission, even if in this case the court upheld Anthropic’s use of purchased books for training purposes. 

In other words, training AI on lawfully acquired copyrighted works was deemed fair use. 

Diversity of Thought is What Matters for Decision Making Advantages

Lots of studies of intellectual diversity suggest it helps with decision making. Conversely, one might argue, if all the members of a team, a company, a community or a group have the same points of view, then there really is not much diversity of thought, and the advantage of diversity is lost. That likely is true no matter the amount of gender, race, culture or other forms of physical diversity within groups that think the same. 

It’s the thinking; the intellectual diversity, that really matters. 

Study / Source	Researchers	Key Insight on “Diversity of Thought” for Problem-Solving	Year	Source
“Groups of diverse problem solvers can outperform groups of high-ability problem solvers”	Lu Hong & Scott E. Page (U. Michigan)	Cognitive-heuristic variety in a group beats sheer individual ability; “diversity trumps ability.”	2004	pnas.org
The Difference: How the Power of Diversity Creates Better Groups, Firms, Schools, and Societies	Scott E. Page (U. Michigan / Santa Fe Institute)	Formal models and real-world cases show perspective/heuristic diversity improves prediction, innovation, and collective accuracy.	2007	muse.jhu.edu
“How Diversity Makes Us Smarter” (Scientific American)	Katherine W. Phillips (Columbia Business School)	Exposure to socially or cognitively different others makes groups more diligent, creative, and accurate.	2014	scientificamerican.com
“Likes Attract: The Sociopolitical Groupthink of (Social) Psychologists”	Richard E. Redding (Chapman U.)	Lack of ideological diversity in psychology skews research agendas and peer review; calls for broader viewpoint inclusion.	2012	journals.sagepub.com
Team of Rivals: The Political Genius of Abraham Lincoln	Doris Kearns Goodwin (historian)	Lincoln’s intentionally ideologically mixed cabinet generated vigorous debate and better wartime strategy.	2005	ft.com
“Diversity in Teams: A Two-Edged Sword”	Katherine Y. Williams & Charles O’Reilly (Stanford)	Functional/cognitive diversity boosts creativity on non-routine tasks but can raise conflict—management practices determine net benefit.	1998 – 1999	psychologicalscience.org
“Political Diversity Will Improve Social Psychological Science”	Jonathan Haidt, José Duarte & Lee Jussim	Reviews evidence that viewpoint homogeneity fosters confirmation bias; argues political diversity strengthens methodology and theory.	2015	pubmed.ncbi.nlm.nih.gov
“Getting Unusual Suspects to Solve R&D Puzzles” (Harvard Business Review)	Karim R. Lakhani & Lars B. Jeppesen	Crowdsourcing to outsiders with varied backgrounds solves R&D problems stumping in-house experts—illustrates value of cognitive distance.	2007	hbr.org

Stablecoins for Retail Payments Face Consumer Value Challenges

The view that stablecoins might be a challenge to the Visa payment network if Amazon or Walmart create their own stablecoin payment systems is logical enough. Such payment systems would benefit the retailers by potentially lowering their payment transaction costs, currently represented by the interchange fees Visa earns when consumers pay using its affiliated bank cards. 

Stablecoins offer the potential advantage of lower transaction fees (typically under one percent compared to two percent to three percent for card processing. 

But what is not clear are consumer incentives to switch payment methods. There are, in other words, consumer “lock in" issues; incentives (cash back) for using the Visa system; switching costs; habits and complexity issues that will work against stablecoin usage. 

Consumers likely will not want to give up the “cash back” feature of using Visa-branded credit cards for payments. After all, that is a tangible benefit of using the cards, at no real cost, assuming all balances are paid off regularly, so that no finance charges are accrued. 

From a consumer's perspective, using a Visa card effectively means getting paid to spend money, while switching to stablecoins often means giving up that immediate financial benefit, unless merchants pass along some of their savings to consumers. But that weakens the business case for switching to stablecoin payment. 

The issue then is asymmetry in the value proposition: stablecoins benefit merchants but not consumers. 

Of course, retailers include the cost of payment system fees in the general pricing of all goods they sell, so the consumer benefit of “cash back” is offset by higher retail product costs. But those costs are hidden; the cash back benefit is obvious. 

So the stablecoin adoption process might initially focus on use cases where card rewards matter less, such as  peer-to-peer transfers, international remittances, or purchases where cards aren't accepted.

Fixed Wireless for Home Broadband is the Biggest New 5G Revenue Source

The Ericsson Mobility Report for the first quarter of 2025 is the 11th consecutive quarter in which fixed wireless has accounted for nearly all broadband net additions in the U.S. market. 

During the quarter, AT&T, Verizon, and T-Mobile collectively added 913,000 new connections, bringing the total number of 5G FWA connections to 12.5 million. The number of 4G connections boosts the total further. 

Globally, most fixed wireless connections still rely on 4G. 

source: Ericsson 

And even if fiber to the home is the dominant home broadband trend, fixed wireless continues to be an important platform, as digital subscriber line and hybrid fiber coax connections decrease. 

source: Ericsson 

Despite all the hoped-for advances 5G would bring in terms of new services, so far it is fixed wireless for home broadband which seems to be the biggest new revenue source for mobile service providers, aside from faster mobile internet access. And some of us would say 5G for mobile broadband is not a “new” service but simply the latest version of mobile broadband, as 4G displaced 3G, for example. 

In fact, there is an argument to be made that fixed wireless for home broadband is bigger than all the other “new” 5G services put together, even using arguably optimistic estimates of new revenue. 

5G Service Type	Estimated 2025 U.S. Annual Revenue (USD)	Notes	Sources
Enhanced Mobile Broadband (eMBB)	$10–12 billion	Largest segment; includes premium mobile plans, high-speed data, streaming, gaming	1,2,3
Fixed Wireless Access (FWA)	$5–7 billion	Rapidly growing; over 10 million U.S. homes expected on FWA by end of 2024	4,5
Internet of Things (IoT/mMTC)	$2–3 billion	Includes smart cities, industrial IoT, logistics, and connected devices	1,2,3
5G Entertainment & Gaming	$1–2 billion	Cloud gaming, AR/VR, immersive media	2
5G Advertising	$0.5–1 billion	Targeted, high-speed, interactive ads	2
Private 5G Networks/Enterprise	$1–2 billion	Dedicated enterprise networks for manufacturing, healthcare, logistics, etc.	3

AI Being Adopted Faster than Most Other Important New Technologies

Some might note that businesses and consumers seem to be adopting the use of artificial intelligence chatbots faster than they have adopted the internet, smartphones, social media, e-commerce, smartphones or search. 

That partly results from the fact that physical infrastructure often takes longer to deploy than a new software app; new habits have to be built and content richness also often takes time to create. 

There was only so much users could do with 56 kbps internet connections, compared to broadband. Smartphones once were primarily useful to use email “on the go.” But that’s a far more limited value proposition than “internet in your pocket or purse.” 

Also, content-based apps including social media or search take some time to create vast libraries of content as well as network effects that have “most of the people I care to interact with on this site.” 

AI chatbots have some advantages, already. “No incremental cost” for casual use means there is not a price barrier. Also, the ability to provide “answers” in better ways than the results of search means the use case is obvious and immediate. 

Also, little technical expertise is required for a person to use an AI chatbot. 

And new habits can be adopted on the existing infrastructure (both network and device). That might be seen in the early adoption and mainstream use of important new technologies, where it often took a decade or more for “most people” to use the new technologies. 

Technology	Early Adoption	Mainstream Adoption
Internet	1970s–1990s	Late 1990s–2000s
Smartphones	1990s–2007	2008–2013
Search Engines	Late 1990s	Early 2000s
Social Media	Early 2000s	2010s
E-commerce	Late 1990s	2000s–2010s
AI (Generative)	2022–present	2023–2025

It might also be hard to say for certain, but perhaps there also is some significant element of business leaders' concern about failing to adopt AI quickly enough, given prior experience with earlier waves of technology. 

At the moment, the concern seems most logical in content-based (video, audio, text, education and learning, consulting, research, “advice” roles) or software development roles, as AI tools seem poised to displace some amount of human activity and roles.