Introduction
Like billions of people around the world, you may have suddenly become familiar with the following words straight out of Silicon Valley or a futuristic movie: Machine learning. GenAI. Large Language Model. Generative. Training. GPT. Explainability. Neural Network. Deep Learning. Hallucination.
These words are synonymous with Artificial Intelligence (AI), the computer systems we can teach to do "thinking" work. Sometimes we teach machines to find patterns of information that we can't given the flood of data in our digital lives. Sometimes we teach machines to do a specific task in a way that augments our life, or supports the work that we already do. The words above are casually thrown around by everyone from tech bros to journalists to advertisers to the AI "assistants" that are starting to pop up on every app and website. But they can distract from what is really going on.
You've picked up this book because "AI" seemingly comes up in every job interview, work meeting, or classroom discussion, shows up in every search you do online, invades your social media, and is splashed across every advertisement you see. And this all seems to have happened overnight. What changed? Everything. And nothing at all. Everything because we as a digital society reached a tipping point. Nothing at all because for twenty years AI has been with us, becoming more and more capable behind the scenes. Is this scary? Exciting? What exactly is AI, anyway? You are likely thinking: How is "the machine" already in my life without me being aware? How do I make the most of the most important innovation since we humans first wrote things down 5000 years ago and then, 575 years later, figured out how to print copies? How do I protect myself, the people I care about, my education, and my job?
The truth is that you—and me, our parents, our kids—have been "teaching the machine" for years. We are all simultaneously consumers and producers in an AI economy that has been around for decades. This book is titled You Teach the Machines because it is fundamentally that simple. AI depends on you. The machine learns from the data you create, just as an infant learns the basics of language from the words you speak. The machine learns to do things that matter to you when you tell it right from wrong, just as a toddler learns grammar when you correct them.
Many corporations have much to gain from AI appearing to be Oz the Great and Powerful. Something magical, an otherworldly black box. It's not. It's just a machine, often with a cynical man behind the curtain. A machine taught with your data and your feedback. The more we all understand AI for what it is, the more we can maximize benefit and minimize harm. You've been teaching AI for as long as you've been using Google, Amazon, social media, and navigation apps, for as long as you've been going to the doctor and swiping your credit card. It's time to take charge and put AI to work for you. You teach the machines.
I started my tech career in 1996 during another period of rapid change. The web was brand new and data, for the first time, was considered an asset with a dollar value (Kerr, 1991). A few years on I started working in AI at Ask Jeeves, the first natural language internet search engine. Twenty-five years later I created a comprehensive data asset and AI program as the Chief Research Informatics Officer of a leading pediatric academic medical center. I left that job, one I loved, to write this book. To help all of us navigate the change of AI.
If you, like most people, have a lot of questions and reservations, and even fears about AI, this book aims to demystify this groundbreaking technology and put your mind at ease. The chapters here will answer your questions about AI, including:
- How did AI seemingly show up everywhere overnight?
- What could change in my life because of AI?
- Can I trust AI?
- How do I use AI to make my life and my family's life better?
Cuneiform To ChatGPT
Your questions are, for the most part, about change. Change from an old normal is always accompanied by uncertainty, and we as humans are hardwired to fear what we don't know. Artificial intelligence is a relatively recent arrival in human history. We don't yet share a widespread understanding of AI or a new normal of daily use. Our uncertainty and fear are completely natural and understandable. This is a revolutionary technology!
Fundamentally, AI is a completely new way that we humans capture our knowledge. That hasn't happened since the Sumerians invented writing in ancient Mesopotamia around 3200 BC and people no longer had to simply remember everything. Historians think the first writing, called cuneiform, was invented to give a customer a grocery receipt. Before cuneiform writing, the only way to capture information was for one person to remember what another person told them. That meant there was no written history. No written receipts. No written recipes. Imagine you figured out that leaving open jars of barley out in a rainstorm made beer, another Mesopotamian innovation from roughly the same time. You tell the recipe to your friend over a few beers, but after a few too many, you both forget what it was. Fast forward, to a time when cuneiform writing has been invented, and you happen upon the barley-in-the-rain trick again. This time, you write it down, and it gets passed along to generations thereafter. From that point forward, it was possible to record human knowledge, leading to massive cultural changes and advancements. Writing made it possible for individual humans to record important knowledge and share it with a relatively few other humans (Finkel & Taylor, 2015). AI makes it possible for us to collect massive amounts of digital knowledge to share with others worldwide, like ChatGPT, which incredibly incorporates information scraped from a public archive of all the websites ever made called Common Crawl.
Artificial intelligence is also an entirely novel way to broadly share, disseminate, and use human knowledge. This hasn't happened with such seismic consequences since serial entrepreneur Johannes Gutenberg figured out how to scale up book production with moveable type in Mainz, Germany in 1450. Gutenberg borrowed heavily to do the R&D and engineering required to invent the printing press. His first book? Copies of the twenty-eight-page learn-to-read Latin schoolbook Ars minor, the first part of an ancient text called Ars grammatica. He's more famous for the next book he printed, the Bible, but it's notable that he started with an educational book, whether he did it intentionally or otherwise. In those days, you had to know Latin to get what might today be called an office job. Before Gutenberg, getting your kid started on Ars grammatica meant paying a scribe much of a laborer's yearly wage to hand-write a copy with pen and ink. That meant only better-off kids learned Latin and went to school. Whatever Johannes Gutenberg intended, printing relatively cheap copies of the equivalent of Dick and Jane for Latin was an early example of doing right by doing well. The cheaper it was to learn to read, the more books he could sell!
Eventually, a diaspora of trained printers who stole Gutenberg's technology set off an explosion of printing across Europe. A mere fifty years later, more than twelve and a half million books had been printed! Printing made it possible for a single human to record important knowledge, then share it with millions of others so they could do stuff with the knowledge. This invention, more than any other, launched Europe, and then the rest of the world, into the modern era. This shifting of the Earth is known as the Gutenberg Effect.
It's important to note that at the time, there was pushback against the printing press. Turns out the intellectual and economic classes felt more secure when knowledge was captured and made available at great expense by scribes copying out books by hand. Secure in their stations, the ruling class viewed the work of a scribe as morally superior to the ink-stained labor of setting type and cranking a press. Classist snobbery was also fueled by the fact that the labor of printing was taken up by the lower classes. But the expansion of literacy in these same craftsmen, indentured apprentices, and servants opened new markets for popular books very different from highbrow manuscripts (Houston, 2016). For example Desiderius Erasmus' widely read and sometimes-banned books advocated the then-radical idea that the church and monarchs should serve the people first (Erasmus, 1515; Erasmus, 1516).
Newspapers, magazines, radio, television, and the internet all extended the innovation of the printed book. All were initially disparaged. If you're of a similar vintage to mine, you may remember some professors prohibiting the use of the internet to do research for your term paper. Be thoughtful about similar criticisms by today's intellectual elite as AI emerges and evolves. Know that some criticism of AI as somehow inauthentic may be defense of the established cultural clout of experts. But also don't completely disregard these same experts, who are justifiably nervous. Some of their criticisms are valid, such as the potential for erosion of critical thinking and writing skills by overuse of language AI.
How does AI relate to writing in ancient Mesopotamia and printing in medieval Germany? Historical, disruptive—and ultimately constructive—precedents of writing and printing help us understand change in our lives caused by the emergence of AI. For now, consider that writing, and then printing, made it possible for the expertise, knowledge, and thoughts of a single human to spread through literacy and education to many other humans. At its worst, this can lead to Adolf Hitler's Mein Kampf. At its best, it can lead to Henry Gray's groundbreaking textbook Gray's Anatomy. Artificial intelligence can make it possible for human knowledge and expertise to spread even further, not just through other humans, but through machines we teach to augment our lives. Artificial intelligence makes it possible for us to capture the knowledge of many, many (all?) humans and share it so it can be used by many, many more (all?). But just as the printing press amplified whatever you printed with it, AI can amplify whatever we teach it. Sometimes for the worse, such as in 2023, when hackers who couldn't write computer code themselves used AI to generate code that they then used to attack an email company. Sometimes for the better when, the same year, biomedical researchers who couldn't write computer code themselves used the same AI to generate computer code at a prestigious hospital to get important data out of a database, greatly speeding up their important work. For better or worse, you teach the machines.
We started with a discussion of writing and printing, historical precedents of innovation that sparked massive global change, very deliberately because AI burst into our world through the exact same door: language.
Our Black Swan Moment
On November 30, 2022, a startup company called OpenAI released ChatGPT, an AI language tool that quickly gained widespread notoriety. The public reaction and significant notoriety of this tool in our public consciousness was an outlier, impossible to predict beforehand. There was a huge impact from this notoriety: its release and the subsequent freak-out in almost every corner of our society changed our conversations and expectations in a very big way, whether or not we actually took the time to play around with ChatGPT. Since that moment, we've been furiously working after the fact to explain how this all happened. We want to go back and figure out how we could have predicted how ChatGPT has upended our lives if we'd only paid attention. This trifecta of an outlier event with huge impact we desperately try to explain afterward is characteristic of what's become known as a black swan event.
Black swan is a theory for understanding the outsized impact of rare events on human society. Events like the September 11 attacks and the subsequent public, political, and military response. Highly improbable but changed everything (Taleb, 2007). A black swan event is also a matter of perspective. A number of expert analysts were not surprised by the attacks on September 11; to them an attack was certain, even if they did not know exactly when or where it would happen. The general public, the military, and much of the Federal government not only did not anticipate the attacks but were completely unprepared to handle such an improbable but consequential event. My goal in writing this book is to help you be informed so you may better handle change we can anticipate, but especially the unlikely but consequential disruption AI is certain to cause. The more you know about AI, the more you use AI, the better off you will be in coping with unexpected events AI will bring in the future. Artificial intelligence depends on you. You teach the machines.
Sam Altman, CEO of OpenAI, may or may not have predicted the public reaction to his launch-and-see-what-happens "creative disruption." It's a typical path in the technology industry, both useful (gather new information) and intellectually lazy (skip the deliberation). Sam did it with his one and only startup before realizing investors (not entrepreneurs) made the big bucks and he made enough money investing to convince him he was a Master of the Universe. Except his only prior startup was a failed social network app called Loopt. Lower stakes. He was successful enough as an investor to co-found OpenAI as a nonprofit artificial intelligence research organization that quickly evolved to have multiple for-profit subsidiaries. It's complicated. After launching ChatGPT, Sam claimed the release was a benevolent attempt to help us all face the reality of the new power of AI. That may or may not have been a reason. More likely it was to grab market share with a product built using an algorithm developed at Google and a few billion dollars he needed to show a return on. Meanwhile people at Google, including Geoffrey Hinton, the originator of modern AI, were doing the more difficult "just because we can, should we?" deliberation, carefully considering the implications of technology certain to cause major disruption. Ultimately, it doesn't matter what Sam claimed as his reason. The immediate impact on public discourse and industry was huge and will likely be viewed as historically significant.
The black swan of AI bursting into our consciousness made us feel like the world shifted underneath us. Professionally, I found myself in the middle of a storm of uncertainty and fear. My colleagues at the Children's Hospital of Philadelphia and I were stunned that ChatGPT had been put into the wild. We were hit by the black swan and found ourselves in the role of helping people who hadn't seen it coming to process and understand what had happened. All of a sudden, much of my job became explaining the emergence of AI after the fact. Before ChatGPT, our society didn't know that AI was around the corner. Now we feel like we don't know what AI will bring, good or bad. The reality is that AI will almost certainly bring more black swan events in the future. But you also have more agency than you realize. You teach the machines.
Why is it worth considering our reaction to AI in the black swan context? Step back from AI for a moment. The black swan way of thinking is useful because it's about admitting we don't know what we don't know. The black swan theory says that randomness is more prevalent than we as humans are wired to accept. Many, if not most, of the truly important events in our lives or our society are unpredictable. So maybe—just maybe—we should balance the amount of time we try to predict and plan with more time spent preparing for change and uncertainty.
Please don't take this suggestion as a forecast for doom and gloom. We humans seem to be hardwired to see change and uncertainty as inherently bad, probably because we don't live very long, and we have evolved to be almost entirely concerned with self-preservation, and to see anything new as a threat. The emergence of AI was jarring in the sensational way that it quickly went from tech company back office to dining room table conversation. We consider the idea of the black swan theory not so much because it explains the emergence of AI, but because it's a good way to prepare for the uncertainty that lies ahead—the certain-to-happen but impossible to predict unintended outcomes of AI. The black swan theory tells us it's up to us to work together to demand and take an active role in this change. So let's go back to the moment when ChatGPT burst onto the scene. Why did it seem like such a big deal?
ChatGPT delivered to everyday users a compelling language-based conversation experience on just about any topic. Note the use of the word "compelling" instead of "accurate" or "correct." This is an important point that we'll get back to. Over the next few months, social media, TV news, and traditional media exploded with examples of ChatGPT answering questions, writing poems, drafting letters, and summarizing information. We were seeing a credible artificial general intelligence application for the first time. ChatGPT is one among many examples of a "chatbot," a computer system taught to respond to text or voice. Chatbots had been mostly annoying and useless up until this point. Boomers, Gen Xers, and Millennials: Remember Clippy? Most recent chatbots weren't much better. They were rarely welcome, mostly unhelpful interjections to already painful customer service experiences. If it seemed like every chatbot you interacted with before ChatGPT was narrow-minded and knew nothing about your specific problem, you're right. Chatbots prior to the current generation were not much more than a fancy search interface to whatever narrow catalog of information their owner wanted you to go through before passing you on to a more expensive human. What was different this time?
The release of ChatGPT gave anyone who signed up for a free account an interface to incredibly powerful new language-based AI technology. Millions upon millions of people did just that and had their first experience with AI in a message-based conversational format that seemed to respond somewhat credibly on any topic. ChatGPT differed from our experience with Clippy and its chatbot descendants in part due to the many billions of dollars invested in its underlying "brain," a type of AI known as a "large language model" that we'll learn more about later. Suffice it to say, ChatGPT could reply to you more authentically based on its having read hundreds of millions of publicly available texts multiple times.
ChatGPT also used a new type of AI (more on this in just a bit) that was much better at figuring out the important words and phrases in your message. Prior chatbots were clumsier and not as insightful. People got super excited because ChatGPT seemed more responsive to what was important in their message. In a way, ChatGPT was the first chatbot that seemed like a good listener. And it could respond to questions on just about any topic because it had read just about everything on the internet. Did you know that ChatGPT was taught with the contents of every web page published all the way back to 2008? That includes but is not limited to every word written by copywriters in online clothing catalogs, every public website, and every online news article not behind a paywall. This was the closest we, the general public, had come to the artificial intelligence of popular science fiction.
Amara's Law
ChatGPT was a big deal. It was creepily relatable and everyone could access it, so we quickly worked ourselves into a frenzy. What does this witchcraft mean!? Companies cynically jumped on our frenzy and slapped "Now with AI!" all over their products, further whipping us up. But there's a great adage known as Amara's Law, after scientist and futurist Roy Amara, who coined it: We tend to overestimate the effect of a particular technology in the short run and underestimate the effect in the long run.
Let's look at communications: My grandmother was born in 1913 and lived to be over 100, passing away in 2014. She remembered her excitement at the first telephone installed in her house when she was a girl. But the Dick Tracy communication watch popularized by the comic book character when my grandmother was in her early twenties didn't arrive until the advent of the voice-enabled smartwatch when she was in her nineties. The gizmo still didn't gain widespread use until Apple released the third version of their smartwatch the year my grandmother passed away. On the other hand, in her last decade, she was an avid user of the internet, the impact of which she couldn't have envisioned when she made her first telephone call or read Dick Tracy comics. Impact overestimated in the near term (Dick Tracy watch), underestimated in the long term (internet).
Is Amara's Law true for AI? Are we overestimating in the short run (Terminator, killer robots) and underestimating in the long run (democratization of expertise, workforce disruption)? Economists describe production as the result of a combination of land (natural resources), labor, capital, and entrepreneurship. If you consider AI to be a form of automation (it certainly can be), then it will change the relationship between labor and the other factors. Does that mean AI will put writers out of work? Yes and no. I'm glad I'm writing this book and not brief descriptions of menswear in an online catalog. Generation of that kind of text is a layup for modern AI. "The machine" has been taught by the digitized text written by generations of copywriters who came before. AI is already making it harder, likely soon impossible, to make a living as a writer of new descriptive catalog copy. In that sense, the answer is yes, AI will put some writers out of work. On the other hand, AI is helping me by improving autocorrect and providing helpful suggestions on brevity, arguably making it less intimidating to sit here and write my first book, so in that sense the answer is no. For me. For now.
Let's take another look at how AI changes the relationship between labor and production. After a long day, I sometimes relax with a TV show called How It's Made. Each episode is a compendium of narrated video shorts documenting the making of things. Think hockey stick manufacturing, snack food production, and lumber milling. A narrator voices soothing descriptions of the "worker" and the "machine" acting in concert to mass produce wooden hockey sticks or stamp out cheese crackers from dough. Generally, the worker does less-frequent tasks like dumping ingredients for cheese crackers into a giant mechanical mixer. The machine does the high-frequency, repetitive tasks like rolling dough, stamping out crackers, and packaging, while industrious instrumental music plays at a low volume in the background. In one memorable episode on milling dimensional lumber out of raw logs, the worker is a saw operator. The machine is a computer system that's been taught to look at the end of each log to determine the ideal cutting pattern to get the most valuable lumber with the least number of cuts. A camera feeds the machine an image of the end of the log. The machine calculates and projects a grid-like pattern of laser light on the end of the log to show the worker where to cut. The worker lines up the saw according to the machine-generated cut pattern, and the saw buzzes through the log. After each cut, the machine adjusts the pattern of laser light to indicate the next cut. The worker flicks joysticks and taps pedals at high speed to move the log and the saw blade to match the laser-projected cut instructions of the machine. It is mesmerizing—a melding of worker and machine much more compelling than the spellchecker that saved me from spelling "laser" as "lazer."
In these cases, AI affects how land, labor, capital, and entrepreneurship combine to produce books and lumber. In both cases, labor for the most part loses out while capital and entrepreneurship win. This is the age-old story of increased productivity driven by automation. But consider the difference between the two anecdotes. The AI-enabled automation of cut planning or mechanical production is familiar. A previous, slower manual manufacturing step (lumber-cut planning) performed by expert sawyers was in part taken over by a faster machine, allowing one saw operator to process many more logs, almost certainly putting saw operators out of work. But the automation of intellectual skills like editing, copywriting, and revising for brevity? That's new. Those jobs have never been under threat. The capability of modern AI to perform tasks that previously required human cognition threatens entirely new sectors of labor. We're going to delve deeply into this later on, but for now consider these examples and Amara's Law as context for the rest of the book.
The Making Of Modern AI
ChatGPT blew up our feeds, birthed a black swan, and got you to the point that you picked up this book. But how did we get to ChatGPT in the first place? By travelling a long road leading to either a rocket-assisted jump into the shiny future or the edge of a cliff above raging rapids, depending on your point of view. (Joking. It's neither.) Wasn't artificial intelligence just an academic science experiment? For a long time, it seemed that way.
Artificial intelligence has been around and in the works since the 1950s. We're not going to delve into the details of the last seventy-five years of AI evolution here. We will look at AI through a theory of evolutionary biology called "punctuated equilibrium," where a species stays mostly the same for maybe ninety-nine percent of its time on Earth, then changes dramatically in a short period of time (Eldridge & Gould, 1972). Sound familiar? We're also going to look at the major jumps in evolution as changes in a silly but illustrative pretend recipe for AI.
Trust that we'll go into more detail in the next chapter, but for now we're going to oversimplify the heck out of things. Artificial intelligence is made from three ingredients: data, computing power, and algorithms. Data is digital information such as web pages, banking transactions, and video from the camera in your new car. Computing power is the ability to perform math calculations on a computer chip. Algorithms are rules for performing math calculations.
Recipe for AI
100,000,000 cups of data
1,000,000 cups of computing power
1 tbsp of the best machine learning algorithm you can get
Separate data into two equal parts, setting one aside. Combine computing power and algorithm in a large bowl. Transfer to your stand mixer and add the first half of the data. Mix completely, then pour into a pan and bake in an electric oven. Check the AI repeatedly until it reaches an accuracy level equal or better than a human. Remove to cool. Now, a really good recipe can be repeated. Repeat the whole process with the 50,000,000 cups of data you set aside, go to the market for another million cups of computing power and another tablespoon of algorithm. Run through the recipe with the second half of the data. If you get the same result, congratulations!
The recipe for AI has always been the same. Lots of data, lots of computing power, and a great algorithm. We could assemble lots of data. Smart people came up with new algorithms. But for most of the last seventy-five years, there wasn't nearly enough computing power to bake useful AI.
Research kicked off in the mid-1950s but stayed largely static for thirty years. This was our equilibrium. We needed 1,000,000 cups of computing power but only had two, so brilliant mathematicians and early computer scientists dreamed up beautiful algorithms partly inspired by the interconnected neurons of our brains, called "neural network" algorithms. In the 1980s, better computer hardware developed by among others IBM, Sun Microsystems, and Thinking Machines Corporation finally caught up to these algorithms. This was the first jump in evolution, a punctuation of the equilibrium. The recipe still required 1,000,000 cups of computing power, which wasn't available, but we could still make a decent AI machine by fiddling with the algorithm and using the 10,000 cups of computing power in the computer we had. The AI of this era are sometimes known as expert systems. They were capable of performing highly specific tasks like calculating mortgage rates or monitoring the safety of hydroelectric dams. Think of a 1980's expert system as a healthy unleavened cake made with buckwheat and molasses. It's a cake, but not very good. Still, it was a major evolutionary jump. A big change in the equilibrium.
By the 2000s, neural network algorithms and the available computing power could be reliably used for specific tasks like handwriting recognition on bank checks. Given my own chicken-scratch handwriting, the fact that a computer could tell the difference between a 2 and a 7 on my checks was pretty cool. Another big jump in evolution.
Artificial intelligence research marched along for another fifteen years until the mid-2010s, when computer hardware again caught up with the newest "deep learning" algorithms. Deep learning algorithms are like stacks of multiple neural networks that mimic the multiple layers (deep) of interconnected neurons (neural network) in the human brain. We now had 100,000 cups of computing power in one computer, enough power to use a deep learning algorithm to teach a machine. The next step in evolution! Deep learning improved the ability of AI to deal with human language, enabling massive improvement in tasks like machine translation and voice recognition. Remember 2015, when many of us started saying, "Hey, Alexa" to oh-so-helpful (but sometimes creepy) devices we brought into our homes? They were surprisingly cheap for what they did but were always listening. We already know the next evolutionary leap happened around 2022. What sparked it this time?
Was it once again more computing power? Actually, no. Turns out that not long after 2015, we got to the required 1,000,000 cups of computing power in one computer. Was it more data? No, we had more of that than ever before. Ironically, for the first time, we had more of both ingredients than we could use! But we were still stuck because, even with all that computing power, it would still take decades for the deep neural network algorithms to process all the available data. This was because you had to do everything on one computer, a bottleneck that limited how much a machine could learn. The key ingredient was a fresh new algorithm that could be mixed with data and as much computing power on as many computers as you could throw at it. This new algorithm could be used to teach AI in weeks or months instead of years or decades.
In 2017, a group of researchers at Google published a research paper describing a new kind of deep learning algorithm called a "Transformer." The paper is titled "Attention Is All You Need" (Vaswani et al., 2017). In it, they show how the Transformer algorithm figures out how much attention to pay to relationships between words. For example, the Transformer can figure out that the relationship between "ice" and "cream" is worth paying attention to, while "ice" and "pizza" don't have anything to do with each other. That way when you ask it to complete the sentence, "I like to eat ice…" it'll give you "I like to eat ice cream." Prior to this approach, algorithms would "re-read" the same whole sentence over and over to figure out what a sentence meant. Relying on attention turned out to be a better way to teach a machine to "understand" complex language, and it turns out that it was a more efficient algorithm for a computer to run, too. This meant that in our recipe the Transformer needed only 900,000 cups instead of 1,000,000 cups of computing power. And it got better results on standard language translation tasks (English to German and English to French). Now we could use the Transformer algorithm in our AI recipe to bake the best AI ever and still have 100,000 cups of computing power left over. But the earth-shattering thing about the Transformer was that it could take one big job (understand a long paragraph) and divide it up into lots of little jobs (understand a word, phrase, or sentence). In the world of AI, that meant you could divide up the job of teaching a machine among hundreds of thousands of computers instead of just one. Computers that all talk to each other over super-fast network connections to arrive at the answer. This divide and conquer approach is called "parallel computing."
Ten percent less expensive (in computing power), with a better result on a longstanding benchmark and the ability to do things in parallel was a very big deal. The paper set off a race to develop the next AI among the big "tech" companies who many years before had actually become big "data" companies. For decades, Google/Alphabet, Facebook/Meta, Amazon, and Microsoft harvested every scrap of our (and our kids') data they could get away with. They invested their massive profits into mind-numbingly huge numbers of computers, which they made yet more money on by renting out as "cloud" computing. All hoped to one day, someday, maybe win the AI race. It's ironic that Google invented the algorithm but still lost the AI race, even though its hybrid corporate/academic "Brain" division hosted and funded the research that produced the Transformer algorithm. Google Brain had attracted leading researchers to work at Google with the promise of intellectual freedom (work on what you find interesting) and academic credit (publish your discoveries). These were scientists who chose to be professors at universities because they didn't want corporate bosses telling them what to work on, and who wanted to share the new knowledge they created with humanity. Brain attempted a win-win where researchers had freedom of inquiry and could share their discoveries while Google could support and benefit from their work. Google Brain researchers published the "Attention Is All You Need" paper in an open online journal and their code on an open-source website for anyone in the world to use—including engineers at a two-year-old startup called OpenAI.
OpenAI took the Transformer and ran with it, ultimately creating ChatGPT. They had billions to spend on data acquisition and cloud computing power from their investors, and they incorporated as a nonprofit, so they were able to attract the often mission-oriented best and brightest scientists and engineers. The good people at OpenAI scraped all the words on all the web pages ever published on the internet and engineered their own version of the Transformer: The Generative Pretrained Transformer, better known as GPT which quickly led to ChatGPT and Sam Altman's decision to release ChatGPT in 2022 whether we were ready or not. In chapter 1 we'll return to how AI works, but now we'll tackle the most basic question: If AI is so great, why am I and everyone I know freaked out by it?
Agency Is What You Need
Artificial intelligence learns from the data we create in our digital lives, and artificial intelligence is taught by what we deem "intelligent" (or not) when interacting with just about every digital tool we use.
Today, you are teaching AI with every "like" button you click, every star rating you give, every purchase in one of the many AI-enabled convenience stores, every notification on your phone you respond to (or don't), every automated text message you reply to, every click on a search page, every drive you take with a navigation app, and on and on. If you feel like the digital world is constantly clamoring, vying, demanding your attention, you're right. Artificial intelligence companies have adapted the Transformer algorithm to learn from your behavior, not just your words. Your likes, ratings, and views are used to teach Transformer-based AI what attracts attention. Attention is the new gold to be harvested from humans, now that we've handed over our data. Do you get sucked into social media? Do your parents? Your kids? Go easy on them and on yourself. Social media companies spend billions of dollars on AI that knows what will grab your attention and when to show it to you so you spend more time in the app. It's almost like they learned from the cigarette companies, who figured out how to deliver nicotine so their product was more addictive. Oh, wait! Do an internet search for "meta addictive" and read more.
We are active participants in a system that completely depends on us while it also influences, changes, and sometimes manipulates our lives. Yet we have zero control, influence, or say in how AI is developed or injected into our world. It's as if Gutenberg built the first printing press by putting the twelve million people who read printed books to work in the shop helping him cast the moveable type. That's why we feel helpless. Why we feel resigned. Why we hunker down and hope for the best. That is, when we're not afraid.
Because AI is scary. Multigenerational scary. Leaving aside apocalyptic science fiction, people of all ages see that, whether we like it or not, we're teaching machines that have the potential to do real and lasting harm. Just about every technological innovation from our past has been embraced by younger generations and feared by older generations. When it comes to AI, teens and twenty-somethings, as well as the middle-aged, and the "most experienced" of us share unease at best. This book was inspired by a similarly titled lecture I started giving when I saw family, friends, people in my community, and professional colleagues wrestling with their fears. People of all generations came to these talks, which means that all generations care and are engaged. Young and "more experienced" alike shared anecdotes and asked variations on the same questions rooted in fear of the road ahead.
Survival Signals
Our fears and anxieties around AI may be broken down and grouped into what I call the "five D's" of Destruction, Deception, Dumbing Down, Disconnection, and Displacement.
- Destruction is the fear of runaway AI causing us physical harm: killer robots, homicidal computers with disembodied voices, autonomous cars running down pedestrians.
- Deception is the fear of bad actors using AI to scam us: a fake voice or video of a trusted person in our lives used to steal, fake images used to cause embarrassment, complex misinformation campaigns carried out automatically by AI.
- Dumbing Down is the fear that our kids, our doctors, even we ourselves will take the easy way out with AI and we'll be worse off: AI does my homework (and I don't learn), AI makes the diagnosis (and perpetuates some bias).
- Disconnection is the fear that AI will further reduce the real-life social interactions that make human life worth living.
- Displacement is the fear that AI will make our work less valuable by automating some or all of it: AI writes my book better/faster/cheaper, AI analyzes financial data better than I can.
But here's the good news: Fear and our "gut" instinct are a big part of how we humans thrive as a species. Our gut instinct is actually a very real and powerful intuition that allows us to quickly see possible, even likely, outcomes without going through a more time-consuming stepwise logical analysis. When these outcomes might be bad, our subconscious intuition gives us what security specialist and author Gavin de Becker called the "gift" of the emotion fear. Fear, in turn, amplifies our intuition into an almost magical state of threat assessment. We are mostly conditioned by evolution to respond to physical threats, but as I contemplated my own fear of AI, I found an interesting parallel in the same signals we use to detect and avoid bodily harm.
In The Gift of Fear (1998), author Gavin de Becker helps us understand seven "survival signals" that can reveal a bad actor. Forced teaming, charm, too many details, typecasting, loan sharking, unsolicited promises, and discounting the word "no" are all tactics de Becker describes as the tools of manipulators, people who put their own interests above yours to your detriment. We can all use our intuition to watch out for these tactics and use them to spot people trying to manipulate us at the dawn of AI, in particular the companies and commercial interests almost exclusively driving AI in the U.S. at the time this book was being written. You'll notice that a lot of advertising and social media apps draw on these tactics to subconsciously influence our behavior.
Forced teaming is when AI or an AI company gets into our lives by "the projection of a shared purpose or experience where none exists." Think about the advertisements you were seeing for AI at the time at the time of this book's publication. Google depicts AI as this happy helper on your phone that makes your photos better. Apple depicts AI as a reassuring, benevolent force in all its products that is there for you to make your life better in innumerable ways. Amazon's Alexa and "Hey Google" make the actors smile while they warmly weave AI into their happy home lives. The corporations behind these tools are singularly invested in your purchase of their products. They might not be evil, but they're not going to show up for you when you're broke and can't pay the subscription fee.
Charm is when an AI or AI company does everything it can "to compel, to control by allure or attraction. Becker advises "think of charm as a verb, not a trait. If you consciously tell yourself, 'this person is trying to charm me,' as opposed to 'this person is charming,' you'll be able to see around it." The financial interests pushing AI on us do everything they can to charm us into using their products. In the cases where we are allowed to know AI is at work, interfaces pour out buckets of niceness. Alexa's voice and "her" choice of words are sweet and reassuring. ChatGPT is unfailingly polite when it is saddened that it cannot address our prompt. All the while, the AI and the business behind it are charming us into dependence to get us to part with our money.
Too many details is when AI or bad actors in the AI industry cloud the context of what's actually happening by deliberately distracting or overwhelming us. "Context" means what AI is doing, how it's doing it, and what the effects are. De Becker writes that "context is always apparent at the start of an interaction…but too many details can make us lose sight of it." When it first came out, ChatGPT had a pretty obvious context. It was presented as a research-grade tool, standalone, separate from your interactions with the rest of the digital world. It answered questions, wrote letters. The context was clear. A year later, everything around us was "powered by AI," with no meaningful explanation, only a fire hose of acronyms like LLM, GPT, and the word "generative" in every sentence. At the time of this writing, Google presented AI-powered search results at the top of their search page as authoritative with only a faint generative AI is experimental in a smaller, lighter font after you clicked "show more" and after all the content. Click yet another very faint "learn more" and you were taken to a separate page with a lot of words and generalizations about how helpful AI may be but absolutely no real explanation of what it was doing, what "generative" meant or what the effect of using AI might be. Distracting and overwhelming!
Typecasting is when an AI company's message "…involves a slight insult, and usually one that is easy to refute." Look to the trend in advertising where we're made to feel "less than" if we don't invite the latest whiz-bang AI and technology into our lives. Apple's first advertisements for "Apple Intelligence" depicted people being dumb and lazy, until Apple Intelligence saved the office idiot by drafting a pithy email to his boss or made the forgetful husband look good by creating a last-minute heartwarming video for his wife's birthday. The unsubtle message is that you're a mouth-breathing dummy and need Apple Intelligence to save you from yourself. The German software company SAP ran advertisements in airports all over the world telling us, "You'll never have all the right answers. But your AI can." They're hoping you leap to your feet, thinking, "I can have the right answers! I'll tell my IT department to buy SAP AI!"
Loan sharking is when an AI company "generously offers assistance but is always calculating the debt." The day I wrote this paragraph, you could buy Amazon's newest AI Echo Dot for delivery in Germany for €64 ($67). The same day, the same hour, you could buy the exact same device in the U.S. for $22! Germans are protected by modern comprehensive data use laws. Americans are not. Alexa is offering to assist Americans for one third the cost, but as we'll see later, you may be surprised by what she can do with your data.
The unsolicited promise is made by AI and AI companies when they appear to solve your problem but do not in any way provide a guarantee. It's always a good idea to take a minute to figure out who is left holding the bag when things go wrong. For example, if you ask Google's AI a question about legal matters, it's answer is presented at the top of the page, in a way that looks visually authoritative. You are welcome to take the answer and run with it. But if you click "show more" and scroll to the fine print, you may see another subdued disclaimer, this time Generative AI is experimental. For legal advice, consult a professional. When you buy a car with an AI Advanced Driver-Assistance System, you may have responded to advertising showing a relaxed person traveling down the road with the promise that the car is taking care of everything. But that is not a guarantee! Every time you tap the warning message on the screen in the car at startup, you've signed a contract that you and your insurance policy guarantee the safety of the car. You hold the liability bag if the car's AI decides to follow faint lines on the road and swerve into the car next to you.
Discounting the word "no" is when the fine print governing just about every AI tool Americans use starts with the words, "By using the service you agree…" and somewhere in the middle includes "…to grant us an unrestrictive, free, worldwide license to your data.…" But if the choice is to use AI-driven Microsoft Word and give up control of your data or refuse and become irrelevant without the ability to use the world's main word processing program do you really have a choice? You may say "no," but then what?
You're right to be intuitively wary of AI because it's being served to us with all of these tactics. It's natural to be afraid. I am. But I try to use that fear to motivate myself to take charge, learn, and use AI to maximum benefit in my life and the lives of people I care about. I try to resist letting fear become paranoia or justification to barricade the door and simply hope for the best. Fear helps us realize there will be unintended consequences from AI. Fear also helps us motivate to put ourselves in a less vulnerable position.
While big corporations are doing what they're legally required to do, maximize shareholder return, there's nothing intrinsically bad about AI. It's just a machine. It can even do amazing good. For example, one of the most important projects I worked on during my time at the Children's Hospital of Philadelphia was figuring out how to use AI to scale up and share the expertise of the very, very few pediatric radiologists who can reliably differentiate a rib fracture sustained in a bike accident from a rib fracture resulting from child abuse. Physicians and researchers are teaching machines to make the distinction so children in towns without an expert radiologist can be protected. As scared of AI as I am, I can't just throw it away because I know the good it is capable of.
If you realize how much agency you have as a consumer, parent, employee, and voter you can choose how AI does and doesn't impact your life. How AI can be used for good. AI, by its very nature, must learn from you. Without you, AI has no way to learn what's relevant, what's right or wrong, and therefore has no value. Your data and the context of your life are what make AI work. You teach the machines.
Our goal is to be drivers, not passengers. Be the windshield, not the bug. You create the data to teach AI. You teach AI the right answers, what's funny, if music is good, and how to do the most important tasks. This book will help you understand AI, warts and all. It will help you try AI out, see what it can do to make your life and your work better and more valuable. Remember, it's up to you. You teach the machines!
Exercises: Try It Out
Throughout this book, I'll share some of the things I've done to learn about AI by using AI. Just as there were no experts when the internet and World Wide Web came to be in 1993, there are no experts at the use of AI. It's all new and changing fast. There is no "they" as in "they are smarter, more experienced at using AI." The field is divided only between the people who start to use AI and the people who don't.
This means that whether you have used AI or not, you are not behind. My goal with this book is to help you get comfortable with AI if you have not done so already. Some of these exercises may seem basic, others may challenge you to stretch. I encourage you to try them all
- Spell Check—example of pattern-recognition behavior that augments our spelling
- Use your phone to type a new text message, or use your computer to type a new document. Misspell "their" as "thier," and see what happens.
- Autocomplete—example of "generative" behavior that augments our spelling and writing, and helps us search things more quickly
- Do a web search using your phone or computer. Type in "what is" and look at the list of suggestions. These are generated from the first two words you typed in.
- Add autocomplete in email. It will finish your sentence for you based on what you've typed in the past or what it thinks you should say. Can be annoying sometimes!
- Autofocus—example of a feedback and optimization behavior that augments our ability to quickly take good pictures
- Open the camera app on your phone. Sit at a table and put a drinking glass or other similar object on the table in front of you. Hold your phone so the glass is in the foreground of the screen. Tap on the image of the glass, then tap on the image of something farther away.
References
de Becker, Gavin, 1998. The Gift of Fear: And Other Survival Signals That Protect Us from Violence. Dell.
Eldridge, Niles, & Stephen Jay Gould, 1972. Punctuated Equilibria: An Alternative to Phyletic Gradualism. Models in Paleobiology. (Retrieved on April 28, 2025, from Punctuated equilibria: an alternative to phyletic gradualism // (1972) pp 82-115 in "Models in paleobiology", edited by Schopf, TJM Freeman, Cooper & Co, San Francisco. : Eldredge, N. & Gould, S.J. : Free Download, Borrow, and Streaming : Internet Archive)
Erasmus, Desiderius, 1515. Sileni Alcibiadis. Froben.
Erasmus, Desiderius, 1516. The Education of a Christian Prince. in [a]edibus Ascensianis.
Finkel, Irving, & Jonathan Taylor, 2015. Cuneiform: Ancient Scripts. J. Paul Getty Museum.
Houston, Keith, 2016. The Book: A Cover-to-Cover Exploration of the Most Powerful Object of Our Time. W. W. Norton & Company.
Kerr, James M, 1991. The IRM Imperative: Strategies for Managing Information Resources. Wiley.
Taleb, Nassim Nicholas, 2010. The Black Swan: The Impact of the Highly Improbable. (2nd ed.). Random House.
Vaswani, Ashish, Noam Shazeer, Niki Parmar, et al., 2017. Attention Is All You Need. ACM Digital Library. (Retrieved on May 8, 2025, from Attention is all you need | Proceedings of the 31st International Conference on Neural Information Processing Systems)
