How Network Effects Make AI Smarter
Network effects have dictated the success of technologies from the telephone to shopping platforms like Etsy, and AI tools such as ChatGPT are no exception. What is different, however, is how those network effects work. Data network effects are a new form. Like the more familiar direct and indirect network effects, the value of the technology increases as it gains users. Here, however, the value comes not from the number of peers (like with the telephone) or the presence of many buyers and sellers (as on platforms like Etsy), but from feedback that helps it make better predictions. More users mean more responses, which further prediction accuracy, creating a virtuous cycle. Companies need to consider three lessons: 1) feedback is crucial, 2) routinize meticulous gathering of information, and 3) consider the data you share, intentionally or not.
Late last year, when OpenAI introduced ChatGPT, industry observers responded with both praise and worry. We heard how the technology can abolish computer programmers, teachers, financial traders and analysts, graphic designersand artists. Fearing that AI will kill the college essayuniversities rushed to revise curricula. Perhaps the most immediate impact, some said, was that ChatGPT could reinvent or even replace the traditional internet search engine. Search and the related ads bring in the vast majority of Google’s revenue. Will chatbots kill Google?
ChatGPT is a remarkable demonstration of machine learning technology, but it is barely viable as a standalone service. To appropriate its technological prowess, OpenAI needed a partner. So we weren’t surprised when the company quickly announced a deal with Microsoft. The union of the AI startup and the legacy tech company may finally pose a credible threat to Google’s dominance, upping the stakes in the “AI arms race.” It also offers a lesson in the forces that will dictate which companies will thrive and which will falter in deploying this technology.
To understand what compelled OpenAI to ally itself with Bing (and why Google may still triumph), we consider how this technology differs from past developments, like the telephone or market platforms like Uber or Airbnb. In each of those examples, network effects — where the value of a product goes up as it gains users — played a major role in shaping how those products grew, and which companies succeeded. Generative AI services like ChatGPT are subject to a similar, but distinct kind of network effects. To choose strategies that work with AI, managers and entrepreneurs must grasp how this new kind of AI network effects work.
Network Effects Work Differently for AI
AI’s value lies in accurate predictions and suggestions. But unlike traditional products and services, which rely on turning supplies (like electricity or human capital) into outputs (like light or tax advice), AI requires large data sets that must be kept fresh through back-and-forth customer interactions. To remain competitive, an AI operator must corral data, analyze it, offer predictions, and then seek feedback to sharpen the suggestions. The value of the system depends on — and increases with — data that arrives from users.
The technology’s performance — its ability to accurately predict and suggest — hinges on an economic principle called data network effects (some prefer data–driven learning). These are distinct from the familiar direct network effect, like those that make a telephone more valuable as subscribers grow, because there are more people you can call. They are also different from indirect or second-order network effects, which describe how a growing number of buyers invites more sellers to a platform and vice versa — shopping on Etsy or booking on Airbnb becomes more attractive when more sellers are present.
Data network effects are a new form: Like the more familiar effects, the more users, the more valuable the technology is. But here, the value comes not from the number of peers (like with the telephone) or the presence of many buyers and sellers (as on platforms like Etsy). Rather, the effects stem from the nature of the technology: AI improves through reinforcement learning, predictions followed by feedback. As its intelligence increases, the system makes better predictions, enhancing its usefulness, attracting new users and retaining existing ones. More users mean more responses, which further prediction accuracy, creating a virtuous cycle.
Take, for example, Google Maps. It uses AI to recommend the fastest route to your destination. This ability hinges on anticipating the traffic patterns in alternative paths, which it does by drawing on data that arrives from many users. (Yes, data users are also the suppliers.) The more people use the app, the more historical and concurrent data it accumulates. With piles of data, Google can compare myriad predictions to actual outcomes: Did you arrive at the time predicted by the app? To perfect the predictions, the app also needs your impressions: How good were the instructions? As objective facts and subjective reviews accumulate, network effects kick in. These effects improve predictions and elevate the app’s value for users — and for Google.
Once we understand how network effects drive AI, we can imagine the new strategies the technology requires.
OpenAI and Microsoft
Let’s start with the marriage of OpenAI and Microsoft. When we beta-tested ChatGPT, we were impressed with its creative, human-like responses, but recognized it was stuck: It relies on a bunch of data last collected in 2021 (so don’t ask about recent events or even the weather). Even worse, it lacks a robust feedback loop: You can’t ring the alarm bell when suggestions are hallucinatory (the company does allow a “thumbs down” response). Yet by linking to Microsoft, OpenAI found a way to test the predictions. What Bing users ask — and how they rate the answers — are crucial to updating and improving ChatGPT. The next step, we imagine, is Microsoft feeding the algorithm with the vast cloud of user data it maintains. As it digests untold numbers of Excel sheets, PowerPoint presentations, Word documents, and LinkedIn resumes, ChatGPT will get better at recreating them, to the joy (or horror) of office dwellers.
There are at least three broad lessons here.
First, feedback is crucial. The benefits of AI intensify with a constant stream of user reactions. To remain intelligent, an algorithm needs a data stream of current user choices and rating of past suggestions. Without feedback, even the best engineering algorithm won’t remain smart for long. As OpenAI realized, even the most sophisticated models need to be linked to ever-flowing data sources. AI entrepreneurs should remember this.
Second, executives should routinize meticulous gathering of information to maximize the benefits of these effects. They ought to traverse the typical financial and operational records. Useful bits of data can be found everywhere, inside and outside the corporation. They may come from interactions with buyers, suppliers, and coworkers. A retailer, for example, could track what consumers looked at, what they placed in their cart, and what they ultimately paid for. Cumulatively, these minute details can vastly improve the predictions of an AI system. Even infrequent data bits, including those outside the company’s control, might be worth collecting. Weather data helps Google Maps predict traffic. Tracking the keywords recruiters use to search resumes can help LinkedIn offer winning tips for job seekers.
Finally, everyone should consider the data they share, intentionally or not. Facts and feedback are essential for building better predictions. But the value of your data can be captured by someone else. Executives should consider whose AI stands to benefit from the data they share (or allow access to). Sometimes, they should limit sharing. For instance, when Uber drivers navigate with the app Waze, they help Google, the owner, to estimate the frequency and length of ridehailing trips. As Google considers operating autonomous taxis, such data could be invaluable. When a brand like Adidas sells on Amazon, it allows the retail behemoth to estimate demand across brands (comparing to Nike) and categories (shoes) plus the price sensitivity of buyers. The results could be fed to a competitor — or benefit Amazon’s private label offerings. To counter that, executives can sidestep platform intermediaries or third parties. They can negotiate data access. They can strive to maintain direct contact with customers. Sometimes, the best solution may be for data owners to band and share in a data exchangelike banks did when establishing ways to share data on creditworthiness.
When you consider AI network effects, you can better understand the technology’s future. You can also see how these effects, like other network effects, tend to make the rich even richer. The dynamics behind AI mean that early movers may be rewarded handsomely, and followers, however quickmay be left on the sidelines. It also implies that when one has access to an AI algorithm and a flow of data, advantages accumulate over time and can’t be easily surmounted. For executives, entrepreneurs, policymakers, and everyone else, the best (and worst) about AI is yet to come.