Is Generative AI Overhyped? Lessons from Steel Mills and Electric Dynamos

by Ryan Frazier

The opinions on the impact of Generative AI–now, in the near future, and distantly–are varied.  To be more direct, the viewpoints are all over the map.  The New York Times has published opinions describing it as both “mid” tech (or what my kids might describe as “meh”) and as a technology whose impact is under-appreciated.  There are papers from prominent tech leaders who are worried enough to propose new national security frameworks and others saying that this is just another example of overhyped technology which never quite lives up to the promise–think VR, the metaverse, and crypto.  I have friends and colleagues posting online opinions across the full range of these viewpoints, and most of them aren’t what I would consider shallow thinkers.

I’m not going to be bold enough to proclaim who is “right” in all of this, but I’d offer that there are likely elements of truth in all of these points, depending upon what, where, and when you are focusing.  Personally, I think  AI is overhyped in some contexts, but not in others. I also feel that its impact will become more substantive over time, though the timing and exact nature of that impact is difficult to predict.  The world is a complicated place, and AI is a complicated technology.  

To help us think through these questions, there are two interesting research frameworks which can explain first why AI can look like revolutionary technology and a failed product at the same time, and secondly how the impact of new technologies can take time to emerge because other systems need to evolve as well.  And despite these research ideas being built around old, industrial technology–steel mills and factory electrification–I think they can both help structure our thinking in the very different world of today.

Disruptive Innovation

The first notion is one of the more popular concepts in recent business research–disruptive innovation.  First formulated by Clay Christensen at Harvard Business School in an article published in 1995, his research showed how a particular type of innovation works to disrupt existing markets.  While the popular notion often reduces this to any new idea that comes along unexpectedly, the theory is more nuanced.  In his book “The Innovators Delima” Christensen looked at competition in the steel production industry.  The innovation he explored was the “mini mill” for steel production.  It initially produced a low quality steel only suitable for rebar, which had poor margins for traditional steel manufacturers.  Because of this, the existing companies were willing to cede ground, which the new companies then captured.  The new process had a significantly lower cost structure for these innovative companies, so they could earn higher profits and also target nonconsumers–customers who couldn’t afford the traditional products at higher prices, but could make use of the cheaper “inferior” products in new ways.  The result was a virtuous cycle, where the increasing volume and revenues for the new companies allowed them to learn how to improve their process and products and steadily increase their market share.  They continuously gained ground on the traditional competitors, who had dismissed the new processes as inferior, and eventually overtook them–disrupting the industry. There are many other examples, including such popular ones as Netflix disrupting the video market and overtaking Blockbuster. (If you are interested in learning more, I highly recommend the HBS Online Disruptive Strategy course developed with Professor Christensen).  

Applying this to Generative AI, we see how the arguments that it isn’t very good–its writing is OK, the images it generates are slightly weird, it doesn’t always give the right answers–line up with the theory of disruptive innovation.  Considering AI in the context of marketing and online content, while it may not be as good as a professional art director or writer, it may be good enough at the lower end of the market, where quality may not be great to start with and any cost saving is valuable.  It also appeals to people who aren’t consumers of those professional services now, like small businesses.  AI also exhibits the classic evolution of rapidly improving in quality.  Two years ago AI-generated images of people often had the wrong number of fingers, strange distortions in the background, and were very unrealistic.  Now I’ve generated pictures of people and had colleagues ask “who is that?”   Hallucinations, while still present, are far less frequent and generative AI can successfully solve challenging science, math and logic problems, to the point where some models are being described as having “PhD level” skills.  Because of its broad use and application, not all aspects of Generative AI behave like disruptive innovation, but it is relevant in many contexts, and we should expect more and more disruption in a variety of industries to continue occurring.

General Purpose Technologies and the Productivity Paradox

The impact of AI in a variety of industries leads us to our second area of exploration.  Throughout modern history there have been a number of key innovations which are referred to as “General Purpose Technologies.”  Think of things like the steam engine in the late 1700s, electrification in the late 1800s and early 1900s, and computer and internet technologies of the late 20th and early 21st century.  They are general purpose because they can be leveraged and used in many different contexts.  Steam engines revolutionized transportation (trains), farming (mechanical tractors), and production (steam-powered industrial machinery).  The internet has touched almost every industry from communications, to banking and finance, education, entertainment and many more.  However, when these technologies first emerge, though they have many applications they often fall into what is called the productivity paradox–while they seem like they can change everything, their impact doesn’t seem to be as significant as expected at first.  

One of my favorite papers on this subject, “The Dynamo and the Computer,” was published in 1990 and drew parallels in the adoption of computer systems in the context of the early years of electrification.  The paper explains how it took many years from the initial application of electric motors (known then as a dynamo) in factories until they were the dominant source of power.  The issue wasn’t that electricity wasn’t a better solution than the major sources of power at that time–water wheels and stream generators–but that the entire ecosystem of factory production, from the physical design and layout of factories to the people needed to implement the new innovations, all needed to adapt or emerge. Factories had to move from central drive and pulley systems powering groups of machines to a distributed electric wiring layout for powering individual machines.  The capital investments required to make this change weren’t immediately worthwhile for many.  Finding experts to perform electrification and manage the new systems was challenging.  At the same time, this change also led to changes in factory siting and design, as there were no advantages of expensive multi-story buildings for central drive systems.  Instead, single-story buildings which were cheaper to build and easier to reconfigure internally became the ideal, but it took time to discover this, apply the learnings and realize the benefits.  Because of these challenges, it was often in industries that were experiencing significant growth (and so building new factories) where electrification adoption was strongest.

Though there are many differences between physical dynamos and a digital product like AI, without doubt we will see similar types of challenges as AI is adopted, which is why the potential benefits often feel so much greater than what we actually see.  Some will use this to argue that AI is “just hype” and so believe the excitement will quickly fade.  However, if you think of this in the context of a larger system, which must change and adapt to leverage the new technology, the uneven adoption and impact isn’t necessarily a sign of it being overhyped, but of the challenges of adaptation.

To make the best use of AI requires thinking about the distribution of work tasks differently than in the past.  Data and information is best leveraged by AI when it is easily accessible, but in many companies it is highly distributed and siloed.  The tacit knowledge of staff–their past learnings, understanding of corporate structure, networks of peers they can collaborate with–is largely unavailable for use by AI technologies.  Employees will need to learn a new set of skills and experiment with how best to leverage them to get work done.  All of this takes time to achieve, and the best practices for achieving it are still being developed and studied.  This is also why adoption will be uneven across companies and industries.  But if you look at organizations accustomed to rapid change and evolution–in particular technology companies like Google, Amazon, and Salesforce–you see rapid adoption.  Startups are another place where adoption is easy, as there is nothing to replace, and indeed there is evidence of new “10x founders” who are significantly more capital efficient than in the past because of AI

Hype vs. Reality?

While it’s true that the excitement may fade as some of these reconfiguration challenges get worked through and the early successes may feel uncompelling, they don’t undermine my belief that AI is indeed a disruptive, transformative, general purpose technology that will broadly impact the world.  While I’ve not included much specific evidence that supports this side of the argument, I’ve been collecting examples which I will share in the future.  These examples, taken together with the models I’ve discussed here, explain why AI can be both overhyped in the moment while still holding the makings of a massive transformation which will have major impacts on the world, impacts we are still in a position to influence.

Content Note:  This article was written entirely by the Author, with light copy editing provided by Google Gemini. Editorial review was conducted by ChatGPT to critique the summarization and application of the presented theories. General background research support provided by ChatGPT.