In March 2017, in a keynote speech at its annual I/O event, Google CEO Sundar Pichai announced that the company was shifting from a “mobile-first world to an AI-first world.” Then a series of announcements followed involving AI in various ways: from the development of specialized chips for optimizing machine learning, to the use of deep learning in new applications including cancer research, to putting Google’s AI-driven assistant on as many devices as possible. Pichai claimed the company was transitioning from “searching and organizing the world’s information to AI and machine learning.”
The announcement was more strategic than a fundamental change in vision. Google’s founder Larry Page outlined this path in 2002:
We don’t always produce what people want. That’s what we work on really hard. It’s really difficult. To do that you have to be smart, you have to understand everything in the world, you have to understand the query. What we’re trying to do is artificial intelligence…. [T]he ultimate search engine would be smart. And so we work to get closer and closer to that.1
In this sense, Google has considered itself on the path to building artificial intelligence for years. Only recently has it openly and outwardly put AI techniques at the heart of everything it does.
Google is not alone in this strategic commitment. That same month, Microsoft announced its “AI-first” intentions, moving away from “mobile-first” and also “cloud-first.”2 But what does the notion of AI-first mean? For both Google and Microsoft, the first part of their change—no longer mobile-first—gives us a clue. To be mobile-first is to drive traffic to your mobile experience and optimize consumers’ interfaces for mobile even at the expense of your full website and other platforms. The last part is what makes it strategic. “Do well on mobile” is something to aim for. But saying you will do so even if it harms other channels is a real commitment.
What does this mean in the context of AI-first? Google’s research director Peter Norvig gives an answer:
With information retrieval, anything over 80% recall and precision is pretty good—not every suggestion has to be perfect, since the user can ignore the bad suggestions. With assistance, there is a much higher barrier. You wouldn’t use a service that booked the wrong reservation 20% of the time, or even 2% of the time. So an assistant needs to be much more accurate, and thus more intelligent, more aware of the situation. That’s what we call “AI-first.”3
That’s a good answer for a computer scientist. It emphasizes technical performance, and accuracy, in particular. But this statement implicitly says something else, too. If AI is first (maximizing predictive accuracy), what becomes second?
The economist’s filter knows that any statement of “we will put our attention into X” means a trade-off. Something will always be given up in exchange. What does it take to emphasize predictive accuracy above all else? Our answer comes from our core economic framework: AI-first means devoting resources to data collection and learning (a longer-term objective) at the expense of important short-term considerations such as immediate customer experience, revenue, and user numbers.
A Whiff of Disruption
Adopting an AI-first strategy is a commitment to prioritize prediction quality and to support the machine learning process, even at the cost of short-term factors such as consumer satisfaction and operational performance. Gathering data might mean deploying AIs whose prediction quality is not yet at optimal levels. The central strategic dilemma is whether to prioritize that learning or instead shield others from the performance sacrifices that entails.
Different businesses will approach this dilemma and make choices differently. But why are Google, Microsoft, and other tech companies going AI-first? Is that something other businesses can follow? Or is there something special about those companies?
One distinguishing feature of these companies is that they are already gathering and generating great swathes of digital data and operating in environments with uncertainty. So, prediction machines are likely to enable tools that they will use extensively throughout products in their business. Internally, tools that involve superior and cheaper prediction are in demand. Alongside this is a supply-side advantage. These companies already house technical talent that they can use to develop machine learning and its applications.
These companies, drawing on the hybrid corn analogy from chapter 17, are like the farmers located in Iowa. But AI-led technologies display another important characteristic. Given that learning takes time and often results in inferior performance (especially for consumers), it shares features of what Clay Christensen termed “disruptive technologies,” meaning that some established companies will find it difficult to adopt such technologies quickly.4
Consider a new AI version of an existing product. To develop the product, it needs users. The first users of the AI product will have a poor customer experience because the AI needs to learn. A company may have a solid customer base and therefore could have those customers use the product and provide training data. However, those customers are happy with the existing product and may not tolerate a switch to a temporarily inferior AI product.
This is the classic “innovator’s dilemma,” whereby established firms do not want to disrupt their existing customer relationships, even if doing so would be better in the long run. The innovator’s dilemma occurs because, when they first appear, innovations might not be good enough to serve the customers of the established companies in an industry, but they may be good enough to provide a new startup with enough customers in some niche area to build a product. Over time, the startup gains experience. Eventually, the startup has learned enough to create a strong product that takes away its larger rival’s customers. By that point, the larger company is too far behind, and the startup eventually dominates. AI requires learning, and startups may be more willing to invest in this learning than their more established rivals.
The innovator’s dilemma is less of a dilemma when the company in question faces tough competition, especially if that competition comes from new entrants that do not face constraints associated with having to satisfy an existing customer base. In that situation, the threat of the competition means that the cost of doing nothing is too high. Such competition tips the equation toward adopting the disruptive technology quickly even if you are an established company. Put differently, for technologies like AI where the long-term potential impact is likely to be enormous, the whiff of disruption may drive early adoption, even by incumbents.
Learning can take a great deal of data and time before a machine’s predictions become reliably accurate. It will be a rare instance indeed when a prediction machine just works off the shelf. Someone selling you an AI-powered piece of software may have already done the hard work of training. But when you want to manage AI for a purpose core to your own business, no off-the-shelf solution is likely. You won’t need a user manual so much as a training manual. This training requires some way for the AI to gather data and improve.5
A Pathway to Learning
Learning-by-using is a term that economic historian Nathan Rosenberg coined to describe the phenomenon whereby firms improve their product design through interactions with users.6 His main applications had to do with the performance of airplanes, whose more conservative initial designs gave way to better designs with larger capacity and greater efficiency as the airplane manufacturers learned through additional use. Manufacturers with an early start had an advantage as they learned more. Of course, such learning curves give strategic advantage in a variety of contexts. They are particularly important for prediction machines, which, after all, rely on machine learning.
Thus far, we have not spent much time distinguishing between the different types of learning that make up machine learning. We have focused mostly on supervised learning. You use this technique when you already have good data on what you are trying to predict; for example, you have millions of images and you already know that they contain a cat or a tumor; you train the AI based on that knowledge. Supervised learning is a key part of what we do as professors; we present new material by showing our students problems and their solutions.
By contrast, what happens when you do not have good data on what you are trying to predict, but you can tell, after the fact, how right you were? In that situation, as we discussed in chapter 2, computer scientists deploy techniques of reinforcement learning. Many young children and animals learn this way. The psychologist Pavlov rang a bell when giving dogs a treat and then found that ringing the bell triggered a saliva response in those dogs. The dogs learned to associate the bell with receiving food and came to know that a bell predicted nearby food and prepared accordingly.
In AI, much progress in reinforcement learning has come in teaching machines to play games. DeepMind gave its AI a set of controls to video games such as Breakout and “rewarded” the AI for getting a higher score without any other instructions. The AI learned to play a host of Atari games better than the best human players. This is learning-by-using. The AIs played the game thousands of times and learned to play better, just as a human would, except the AI could play more games, more quickly, than any human ever could.7
Learning occurs by having the machine make certain moves and then using the move data along with past experience (of moves and resulting scores) to predict which moves will lead to the biggest increases in score. The only way to learn is to actually play. Without a pathway to learning, the machine will neither play well nor improve over time. Such pathways to learning are costly.
When to Deploy
Those familiar with software development know that code needs extensive testing to locate bugs. In some situations, companies release the software to users to help find the bugs that might emerge in ordinary use. Whether by “dog fooding” (forcing early versions of software to be used internally) or “beta testing” (inviting early adopters to test the software), these forms of learning-by-using involve a short-term investment in learning to enable the product to improve over time.
This short-term cost of training for a longer-term benefit is similar to the way humans learn to do their jobs better. While it does not take a tremendous amount of training to begin a job as a crew member at McDonald’s, new employees are slower and make more mistakes than their more experienced peers. They improve as they serve more customers.
Commercial airline pilots also continue to improve from on-the-job experience. On January 15, 2009, when US Airways Flight 1549 was struck by a flock of Canada geese, shutting down all engine power, Captain Chesley “Sully” Sullenberger miraculously landed the plane on the Hudson River, saving the lives of all 155 passengers. Most reporters attributed his performance to experience. He had recorded 19,663 total flight hours, including 4,765 flying an Airbus A320. Sully himself reflected: “One way of looking at this might be that for 42 years, I’ve been making small, regular deposits in this bank of experience, education, and training. And on January 15, the balance was sufficient so that I could make a very large withdrawal.”8 Sully and all his passengers benefited from the thousands of people he’d flown before.
The difference between the skills of new cashiers and pilots in what constitutes “good enough to get started” is based on tolerance for error. Obviously, our tolerance is much lower for pilots. We take comfort that pilot certification is regulated by the US Department of Transportation’s Federal Aviation Administration and requires a minimum experience of fifteen hundred hours of flight time, five hundred hours of cross-country flight time, one hundred hours of night flight time, and seventy-five hours of instrument operations time, even though pilots continue to learn from on-the-job experience. We have different definitions for good enough when it comes to how much training humans require in different jobs. The same is true of machines that learn.
Companies design systems to train new employees until they are good enough and then deploy them into service, knowing they will improve as they learn from experience doing their job. But determining what constitutes good enough is a critical decision. In the case of prediction machines, it can be a major strategic decision regarding timing: when to shift from in-house training to on-the-job learning.
There are no ready answers for what constitutes good enough for prediction machines, only trade-offs. Success with prediction machines will require taking these trade-offs seriously and approaching them strategically.
First, what tolerance do people have for error? We have high tolerance for error with some prediction machines and low tolerance for others. For example, Google’s Gmail app reads our email, uses AI to predict how we may want to respond, and generates three short responses to choose from. Many users report enjoying using the app even though it has a 70 percent failure rate (at the time of writing, the AI-generated response is only useful for us about 30 percent of the time). The reason for this high tolerance for error is that the benefit of reduced composing and typing outweighs the cost of providing suggestions and wasting screen real estate when the predicted short response is wrong.
In contrast, we have low tolerance for error in the realm of autonomous driving. The first generation of autonomous vehicles, which Google largely pioneered, was trained using specialist human drivers who took a limited number of vehicles and drove them hundreds of thousands of kilometers, much like a parent supervising a teenager on driving experiences.
Such human specialist drivers provide a safe training environment, but they are also extremely limited. The machine only learns about a few situations. Someone may take many millions of miles in varying environments and situations before they have learned how to deal with the rare incidents that lead to accidents. For autonomous vehicles, real roads are nasty and unforgiving precisely because nasty or unforgiving human-caused situations can occur on them.
Second, how important is capturing user data in the real world? Understanding that training might take a prohibitively long time, Tesla rolled out autonomous vehicle capabilities to all of its recent models. These capabilities included a set of sensors that collect environmental data as well as driving data, which is uploaded to Tesla’s machine learning servers. In a very short time, Tesla can obtain training data just by observing how the drivers of its cars drive. The more Tesla vehicles are on the road, the more Tesla’s machines can learn.
However, in addition to passively collecting data as humans drive their Teslas, the company needs autonomous driving data to understand how its autonomous systems are operating. For that, it needs to have cars drive autonomously so that it can assess performance, but also analyze when a human driver, whose presence and attention are required, chooses to intervene. Tesla’s ultimate goal is not to produce a copilot or a teenager who drives under supervision, but a fully autonomous vehicle. That requires getting to the point where real people feel comfortable in a self-driving car.
Herein lies a tricky trade-off. To get better, Tesla needs its machines to learn in real situations. But putting its current cars in real situations means giving customers a relatively young and inexperienced driver, although perhaps as good as or better than many young human drivers. Still, this is far riskier than beta testing whether Siri or Alexa understood what you said or if Gmail correctly predicts your response to an email. In the case of Siri, Alexa, or Gmail, a mistake means a lower-quality user experience. In the case of autonomous vehicles, a mistake means putting lives at risk.
That experience can be scary.9 Cars can exit freeways without notice or press the brakes when mistaking an underpass for an obstruction. Nervous drivers may opt not to use the autonomous features and, in the process, hinder Tesla’s ability to learn. Even if the company can persuade some people to become beta testers, are those the people it wants? After all, a beta tester for autonomous driving may be someone with a taste for more risk than the average driver. In that case, who is the company training its machines to be like?
Machines learn faster with more data, and when machines are deployed in the wild, they generate more data. However, bad things can happen in the real world and damage the company brand. Putting products in the wild earlier accelerates learning but risks harming the brand (and perhaps the customer); putting them out later slows learning but allows for more time to improve the product in-house and protect the brand (and, again, perhaps the customer).
For some products, like Gmail, the answer to the trade-off seems clear because the cost of poor performance is low and the benefits from learning from customer usage are high. It makes sense to deploy this type of product in the real world early. For other products, like cars, the answer is murkier. As more companies across all industries seek to take advantage of machine learning, strategies associated with choosing how to handle this trade-off will become increasingly salient.
Learning by Simulation
One intermediate step to soften this trade-off is to use simulated environments. When human pilots are training, before they get their hands on a real plane in flight, they spend hundreds of hours in what are very sophisticated and realistic simulators. A similar approach is available for AI. Google trained DeepMind’s AlphaGo AI to defeat the best Go players in the world not just by looking at thousands of games played between humans but also by playing against another version of itself.10
One form of this approach is called adversarial machine learning, which pits the main AI and its objective against another AI that tries to foil that objective. For example, Google researchers had one AI send messages to another using an encryption process. The two AIs shared a key to encoding and decoding the message. A third AI (the adversary) had the messages but not the key and tried to decode them. With many simulations, the adversary trained the main AI to communicate in ways that are hard to decode without the key.11
Perhaps the most exciting developments in simulated learning are coming alongside developments in quantum computing. Quantum computers use the properties of quantum mechanics to enable complex calculation. In certain areas, including factoring large numbers and simulating certain physical and chemical processes, they have an advantage over classical computing. Such computers are being developed by Google, IBM, and Microsoft and quantum-focused firms like D-Wave and Xanadu.12 In 2022, Toronto startup, OTI Lumionics and researchers at the University of British Columbia demonstrated that quantum machine learning could perform in a superior manner to classical computing in simulating organic light-emitting diode (OLED) materials for use in displays.13 This is critical in developing the ability to display colors more accurately but also has potential applications for other material development and drug discovery.
Such simulated learning approaches cannot take place on the ground; they require something akin to a laboratory approach that produces a new machine learning algorithm that is then copied and pushed out to users. The advantage is that the machine is not trained in the wild, so the risk to the user experience, or even to the users themselves, is mitigated. The disadvantage is that simulations may not provide sufficiently rich feedback, reducing, but not eliminating, the need to release the AI early. Eventually, you have to let the AI loose in the real world.
Learning in the Cloud versus on the Ground
Learning in the wild improves the AI. The company can then use real-world outcomes that the prediction machine experiences to improve the predictions for next time. Often, a company collects data in the real world, which refines the machine before it releases an updated prediction model.
Tesla’s Autopilot never learns on the job with actual consumers. When it is out in the field, it sends the data back to Tesla’s computing cloud. Tesla then aggregates and uses that data to upgrade Autopilot. Only then does it roll out a new version of Autopilot. Learning takes place in the cloud.
This standard approach has the advantage of shielding users from undertrained versions. The downside, however, is that the common AI that resides on devices cannot take into account rapidly changing local conditions or, at the very least, can only do so when that data is built into a new generation. Thus, from the perspective of a user, improvements come in jumps.
By contrast, imagine if the AI could learn on the device and improve in that environment. It could then respond more readily to local conditions and optimize itself for different environments. In environments where things change rapidly, it is beneficial to improve the prediction machines on the devices themselves. For example, on apps like Tinder (the popular dating app where users make selections by swiping left for no or right for yes), users make many decisions rapidly. This can feed into the predictions immediately to determine which potential dates to show next. Tastes are user-specific and change over time, both over the course of a year and by time of day. To the extent that people are similar and have stable preferences, sending to the cloud and updating will work well. To the extent that an individual’s tastes are idiosyncratic and rapidly changing, then the ability to adjust predictions at the level of the device is useful.
Companies must trade off how quickly they should use a prediction machine’s experience in the real world to generate new predictions. Use that experience immediately and the AI adapts more quickly to changes in local conditions, but at the cost of quality assurance.
Permission to Learn
Learning often requires customers who are willing to provide data. If strategy involves doing something at the expense of something else, then in the AI space, few companies made a stronger, earlier commitment than Apple. Tim Cook wrote, in a special section devoted to privacy on Apple’s home page: “At Apple, your trust means everything to us. That’s why we respect your privacy and protect it with strong encryption, plus strict policies that govern how all data is handled.”14
He went on:
A few years ago, users of Internet services began to realize that when an online service is free, you’re not the customer. You’re the product. But at Apple, we believe a great customer experience shouldn’t come at the expense of your privacy.
Our business model is very straightforward: We sell great products. We don’t build a profile based on your email content or web browsing habits to sell to advertisers. We don’t “monetize” the information you store on your iPhone or in iCloud. And we don’t read your email or your messages to get information to market to you. Our software and services are designed to make our devices better. Plain and simple.15
Apple did not make this decision due to a government regulation. Some claimed Apple made the decision because it was purportedly lagging behind Google and Facebook in developing AI. No company, certainly not Apple, could eschew AI. This commitment would make its job harder. It plans to do AI in a way that respects privacy. It is making a big strategic bet that consumers will want control over their own data. Whether for security or privacy, Apple has bet that its commitment will make consumers more, not less, likely to allow AI onto their devices.16 Apple isn’t alone in betting that protecting privacy will pay off. Salesforce, Adobe, Uber, Dropbox, and many others have invested heavily in privacy.
This bet is strategic. Many other companies, including Google, Meta, and Amazon, have chosen a different path, telling users that they will use data to provide better products. Apple’s focus on privacy limits the products it can offer. For instance, both Apple and Google have face recognition built into their photo services. To be useful to consumers, the faces have to be tagged. Google does this, preserving the tags, regardless of device, since the recognition runs on Google servers. Apple, however, because of privacy concerns, has opted to have that recognition occur at the device level. That means if you tag faces of people you know on your Mac, the tags will not carry over to your iPhone or iPad. Not surprisingly, this creates a situation where privacy concerns and consumer usability hit a roadblock. (How Apple will deal with these issues is unknown at the time of writing.)
We do not know what will emerge in practice. In any case, our economist filter makes it clear that the relative payoffs associated with trading people’s privacy concerns for predictive accuracy will guide the ultimate strategic choice. Enhanced privacy might give companies permission to learn about consumers but may also mean the learning is not particularly useful.
Experience Is the New Scarce Resource
Navigation app Waze collects data from other Waze users to predict the location of traffic problems. It can find the fastest route for you personally. If that were all it was doing, there would be no issue. However, prediction alters human behavior, which is what Waze is designed to do. When the machine receives information from a crowd, its predictions may be distorted by that fact.
For Waze, the problem is that its users will follow its guidance to avoid traffic problems, perhaps through side streets. Unless Waze adjusts for this, it will never be alerted that a traffic problem is alleviated and the normal route is once again the fastest. To overcome this obstacle, the app must therefore send some human drivers back toward the traffic jam to see if it is still there. Doing so presents the obvious issue—humans so directed might be sacrificial lambs for the greater good of the crowd. Not surprisingly, this degrades the quality of the product for them.
There are no easy ways to overcome the trade-off that arises when prediction alters crowd behavior, thereby denying AI of the very information it needs to form the correct prediction. In this instance, the needs of the many outweigh the needs of the few or the one. But this is certainly not a comfortable way of thinking about managing customer relationships.
Sometimes, to improve products, especially when they involve learning-by-using, it is important to jolt the system so that consumers actually experience something new that the machine can learn from. Customers who are forced into that new environment often have a worse experience, but everyone else benefits from those experiences. For beta testing, the trade-off is voluntary, as customers opt into the early versions. But beta testing may attract customers who do not use the product the same way as your general customers would. To gain experience about all your customers, you may sometimes need to degrade the product for those customers in order to get feedback that will benefit everyone.
Humans Also Need Experience
The scarcity of experience becomes even more salient when you consider the experience of your human resources. If the machines get the experience, then the humans might not. Recently, some expressed concern that automation could result in the deskilling of humans.
Air France Flight 447 crashed into the Atlantic on route from Rio de Janeiro to Paris in 2009. The crisis began with bad weather but escalated when the plane’s autopilot disengaged. At the helm during that time, unlike Sully in the US Airways plane, a relatively inexperienced pilot poorly handled the situation, according to reports. When a more experienced pilot took over (he had been asleep), he was unable to properly assess the situation.17 The experienced pilot had slept little the night before. The bottom line: the junior pilot may have had almost three thousand hours in the air, but it was not quality experience. Most of the time, he had been flying the plane on autopilot.
Automation of flying has become commonplace, a reaction to evidence that showed that most airplane accidents after the 1970s were the result of human error. So, humans have since been removed from the control loop. However, the ironic unintended consequence is that human pilots garner less experience and become even worse.
For economist Tim Harford, the solution is obvious: automation has to be scaled back. What is being automated, he argues, are more routine situations, so you require human interventions for more extreme situations. If the way you learn to deal with the extreme is by having a great feel for the ordinary, therein lies a problem. The Air France plane faced an extreme situation without the proper attention of an experienced hand.
Harford stresses that automation does not always lead to this conundrum:
There are plenty of situations in which automation creates no such paradox. A customer service webpage may be able to handle routine complaints and requests, so that staff are spared repetitive work and may do a better job for customers with more complex questions. Not so with an aeroplane. Autopilots and the more subtle assistance of fly-by-wire do not free up the crew to concentrate on the interesting stuff. Instead, they free up the crew to fall asleep at the controls, figuratively or even literally. One notorious incident occurred late in 2009, when two pilots let their autopilot overshoot Minneapolis airport by more than 100 miles. They had been looking at their laptops.18
Not surprisingly, other examples we’ve discussed in this book tend to fall into the category of airplanes rather than customer service complaints, including the whole domain of self-driving cars. What will we do when we don’t drive most of the time but have a car that hands control to us during an extreme event? What will our children do?
The solutions involve ensuring that humans gain and retain skills, reducing the amount of automation to provide time for human learning. In effect, experience is a scarce resource, some of which you need to allocate to humans to avoid deskilling.
The reverse logic is also true. To train prediction machines, having them learn through the experience of potentially catastrophic events is surely valuable. But if you put a human in the loop, how will that machine’s experience emerge? And so another trade-off in generating a pathway to learning is between human and machine experience.
These trade-offs reveal the implications of the AI-first declarations of the leadership of Google, Microsoft, and others. The companies are willing to invest in data to help their machines learn. Improving prediction machines takes priority, even if that requires degrading the quality of the immediate customer experience or employee training. Data strategy is key to AI strategy.
KEY POINTS
- Shifting to an AI-first strategy means downgrading the previous top priority. In other words, AI-first is not a buzzword—it represents a real trade-off. An AI-first strategy places maximizing prediction accuracy as the central goal of the organization, even if that means compromising on other goals such as maximizing revenue, user numbers, or user experience.
- AI can lead to disruption because incumbent firms often have weaker economic incentives than startups to adopt the technology. AI-enabled products are often inferior at first because it takes time to train a prediction machine to perform as well as a hard-coded device that follows human instructions rather than learning on its own. However, once deployed, an AI can continue to learn and improve, leaving its unintelligent competitors’ products behind. It is tempting for established companies to take a wait-and-see approach, standing on the sidelines and observing the progress in AI applied to their industry. That may work for some companies, but others will find it difficult to catch up once their competitors get ahead in the training and deployment of their AI tools.
- Another strategic decision concerns timing—when to release AI tools into the wild. AI tools are, initially, trained in-house, away from customers. However, they learn faster when they are deployed into commercial use because they are exposed to real operating conditions and often to greater volumes of data. The benefit to deploying earlier is faster learning, and the cost is greater risk (risk to the brand or customer safety by exposing customers to immature AIs that are not properly trained). In some cases, the trade-off is clear, such as with Gmail, where the benefits of faster learning outweigh the cost of poor performance. In other cases, such as autonomous driving, the trade-off is more ambiguous given the size of the prize for being early with a commercial product weighed against the high cost of an error if the product is released before it is ready.