Brave New Words: How AI Will Revolutionize Education (and Why That's a Good Thing)

  • Page xvi Large language models such as GPT- 4, short for Generative Pre- trained Transformer, are essentially big, powerful— albeit digital—" word brains" trained on a colossal amount of information from books, articles, websites, and all sorts of written material.
  • Page xvii What it lacks in real- world sensory experiences of the human brain, it compensates for by having exposure to more language than any human might hope to read, watch, or listen to in multiple lifetimes.
  • Page xviii The best way to think of a parameter is a number describing the strength of a connection between two nodes in the neural net that represent the large language model. You can view it as a representation of the strength of a synapse between two neurons in a brain. When it was first launched in 2018, GPT- 1 had more than 100 million parameters. I found myself suddenly exhilarated to be one of the first people on the planet to see the capabilities of GPT- 4.
  • Page xix We've also shown that rather than somehow being a substitute for the teacher, videos can off- load pieces of a lecture, freeing up more time for personalized learning, hands- on activities, or classroom conversation. This arguably makes the teachers more valuable, not less. And now it was time to see if generative AI could do the same— support students and let teachers move up the value chain.
  • Page xxi In 1948, while working for Bell Labs, he started dabbling in the field we now know as artificial intelligence. Shannon decided to play with how an algorithm approximates language. He published a paper in The Bell System Technical Journal called "A Mathematical Theory of Communication." It was the early days of digital computers—well before the advent of the internet—and Shannon's information theory first made the case that a series of probabilistic processes could approximate the English language. ... [In] 1950, [Alan Turing] wrote a foundational paper titled "Computing Machinery and Intelligence," where he introduced the concept of the imitation game, which we now know as the Turing test.
  • Page xxiii convincingly human technology capable of being a great tutor might also be a technology that bad actors could use to defraud or brainwash unsuspecting people.
  • Page xxiv As someone trying his best to use technology for good, I wondered if we might be able to utilize generative AI to achieve the greatest positive effect and edge us closer to that utopian scenario, especially in the realm of education. ... The implications of GPT-4 were nothing short of revelatory. ... What really caught my attention, though, was its ability to write in different tones and styles.
  • Page xxvi The AI was not quite perfect yet. It was getting math incorrect more than I liked, but I could even see improvement as I got better at prompting it.
  • Page xxvii what if the AI was able to help teachers write their lesson plans? What if it could enter into a debate with a student? What if it could create projects? What if it could help a student remove stressors or inspire a student to create new ideas? What if the technology was able to quiz a student or lead a student in a review session? Educators would be able to create novel activities that students could do with the AI. ... The AI might help students compose essays, making them better writers by providing them with immediate feedback. ... Artificial intelligence, we worried, might turn our kids into a generation of cheaters who were not going to learn anything. With the AI taking over, parents who once helped their kids with homework might lose an important point of connection.
  • Page xxviii Nearly two decades earlier, I had seen similar fears around on-demand video in education: Was it going to be a distraction for students? Was it going to lower their attention spans? Would it isolate students instead of promoting connections between them and their teachers? How were students going to be able to know what to watch? Whom would they ask if they got stuck on a subject and had questions? ... might hasten learning globally and even get us closer to realizing a world in which every person on earth had access to affordable world-class learning.
  • Page xxx We had to show AI's real value to students as a Socratic tutor, as a debating partner, as a guidance counselor, as a career coach, and as a driver of better outcomes in their academics. So, alongside OpenAI, we created a rapid prototyping team that began to build an AI-infused education platform we would come to call Khanmigo.
  • Page xxxii Here we were, working with a technology that took writing, and everything we knew about teaching and learning in all domains, to another level.
  • Page xxxiii We are at a turning point in education, one with far-reaching implications that is changing, and will continue to change, everything about learning, work, and human purpose.
  • Page 3 In early 2023, the Los Angeles Unified School District became the first major school system to ban it. ... it said, the tool did not help build critical thinking and problem-solving skills.
  • Page 4 fear large language models will worsen instances of students not doing their own work. We fear that the potential effect on student writing skills will be catastrophic, with generative AI able to generate text quickly and efficiently for them. We fear that because GPT-generated text pulls from millions of online sources full of biased language and viewpoints, the information it lays out in its final form will be biased as well.
  • Page 5 The most successful students will be those who use AI to help make conceptual connections for developing ideas. Students who learn to use AI ethically and productively may learn not only at an exponentially higher rate than others but also in a way that allows them to remain competitive throughout their careers. They will have a deeper understanding of the given subject matter, because they will know how to get their questions answered. Rather than atrophying, their curiosity muscle will be strengthened.
  • Page 6 Workers will need to learn how to use large language models to automate almost any traditional white-collar process, too, from collating information to doing analysis on spreadsheets.
  • Page 8 Could it help close learning gaps or provide access to quality education regardless of geographic constraints, economic limitations, or social circumstances? ... What might it be like if every student on the planet had access to an artificially intelligent personal tutor: an AI capable of writing alongside the student; an AI that students could debate any topic with; an AI that fine-tuned a student's inherent strengths and augmented any gaps in learning; an AI that engaged students in new and powerful ways of understanding science, technology, engineering, and mathematics; an AI that gave students new ways of experiencing art and unlocking their own creativity; an AI that allowed for students to engage with history and literature like never before?
  • Page 10 On March 15, 2023, when we launched our AI assistant, by this point called Khanmigo—
  • Page 11 Among other things, it provides a personalized and patient tutor that focuses on the learner's interests or struggles and empowers educators to better understand how they can fully support their students. ... A caring and student-attuned instructor who presented clear learning objectives, assessments, and specialized feedback until, eventually, that student demonstrated a real grasp of the material.
  • Page 17 Luckily, by the time we launched, educators were starting to come around by March 2023, it seemed that educators craved a tool that used the power of ChatGPT-like technology for education, implemented with learning and teacher support in mind.
  • Page 29 If a teacher's goal is to give students practice and assessment in structured thinking, language, and grammar, simple storytelling, or just forming and backing up an opinion, you don't necessarily need a traditional take-home writing assignment, where there is likely a strong temptation to use ChatGPT. Instead, an in-class, proctored, five-paragraph essay might do the trick, and it would unfold in a context where there is more support from the teacher (and teachers can directly observe students in their process). If the task might be hard to complete in one sitting, students can work on it in multiple classes, always with the teacher around to support students and ensure that the work is their own.
  • Page 32 The people who weren't very good writers in class wrote bad essays outside class, too, he says. The AI helps them catch up, and it gives them an explanation of where they are with their skills; for instructors, the AI helps flag the students who need the most help and attention.
  • Page 33 educators are finding that these generative AI tools make our students far more skilled and efficient writers. ... The future of writing in school will evolve into a more diverse set of activities, depending on the pedagogical goals and comfort of the teacher. ... I'm an educator looking to ensure that my students have strength here, I would do more in-class writing assignments where it is 100 percent the student's own work.
  • Page 34 At its best, an education-based AI platform can be the world's finest assistant and co-collaborator, objective in its assessments and thorough in its analytics, designed to do one thing and one thing only: to sharpen a student's skills. ... Beyond facilitating reading comprehension, AI can allow learners to immerse themselves in the worlds of the characters in ways that would have seemed like science fiction only a few years ago. ... Based on a 2020 Gallup analysis of data from the U.S. Department of Education, 54 percent of Americans between the ages of sixteen and seventy-four read below a sixth-grade level.
  • Page 39 Instead of answering multiple-choice questions about a text or passage, imagine that students write out the author's intent behind a choice of words or explore the main idea of the passage (while highlighting those parts of the passage). Picture the AI then asking follow-up questions based on what a student writes. It could ask students to draft a conclusion for an incomplete essay that forces them to understand everything that came before. ... Imagine if, when reading a book, a student could have a discussion with the AI at the end of every chapter. The AI might ask the student what they think about the book so far, or whether anything was particularly interesting or confusing about the material.
  • Page 42 The first time I saw examples of this, I had the same questions I am sure many of you did: Is generative AI a creativity killer? If the root of creativity is individual agency, what happens when our kids can simply log on to an app leveraging generative AI, type in or speak a request, and then create imaginative works designed by an imagination not their own? How will our kids learn to think creatively for themselves?
  • Page 45 Some would also argue that generative AI's "creativity" is just derivative from all the data it has been exposed to. But isn't that very human as well? Even the large leaps in human creativity have been closely correlated to things that the creator has been exposed to.
  • Page 46 Much like poets hanging out at a café in Paris, humans and AI can augment each other and inspire a mutually creative process.
  • Page 48 beneficial to children's creativity by providing them with a tool to help them generate, play, and get feedback on ideas in a judgment-free zone. ... With the emergence of artificial intelligence, we're also seeing a shift in the barriers to entry that once limited people from learning a variety of crafts. ... Generative AI is the writing tutor that will teach learners, exploring diverse genres, themes, and narrative structures with them. ... learn to play musical instruments, suggesting practice routines and fingering techniques and deciphering initial musical scores based on their preferences. It can help with improvisation by providing melodic ideas and chord progressions aligned with their playing style.
  • Page 49 Will generative AI, with its ability to produce images, music, and stories, eventually make professional creatives obsolete? ... The net effect of the world of generative AI is that we are going to get more expressions of creativity, and creatives with wider and deeper skill sets, somewhat out of necessity but also thanks to the opportunity generative AI provides.
  • Page 50 Mozart, Einstein, and da Vinci weren't just innately gifted. They had access to opportunities and resources that the bulk of humanity didn't have access to. ... Technology has generally lowered the cost of access to world-class tools and learning. .... Not only does it allow students to produce more polished, finished works, but it can model the creative process with them.
  • Page 53 Not only does Khanmigo let users chat with literary characters, they can also chat with historical figures—anyone from Benjamin Franklin to Cleopatra to Rembrandt:
  • Page 55 We designed Khanmigo to provide responses both drawn from accurate history and portrayed through the lens of the character with acuity.
  • Page 56 you no longer have to travel to Colonial Williamsburg to talk to George Washington or Benjamin Franklin.
  • Page 62 If this tool can be used to engage students and classrooms about history in a way that traditional textbooks and movies can't, I think it is healthy as long as there are reasonable guardrails in place (including helping the user know about the limitations).
  • Page 63 Generative AI, with its ability to mix media and content, has the potential to bring history and civics lessons to life. By offering an interactive and immersive learning experience, it empowers students to delve into historical events, engage in meaningful discussions, and develop a deeper comprehension of civic principles. Its personalized explanations, responsive question prompts, and diverse perspectives stimulate critical thinking and encourage students to form their own well- informed opinions. With these types of tools, history and civics lessons transcend conventional boundaries, empowering students to connect with the past and understand the present.
  • Page 69 As impressive as all of this was, it quickly became clear that it wasn't accurate all of the time. If you asked it for links to sources, it might make them up.
  • Page 72 between the safeguards on Khanmigo and ongoing improvements to GPT-4, the platform can work with specific data and do so with minimal hallucinations.
  • Page 88 we still haven't gotten to what is perhaps the most overlooked, yet most important, role of a tutor: providing motivation and accountability. ... the AI can step in to address these needs. ... parents and teachers was that the platform is powerful for students who proactively seek out the AI's help.
  • The Most Important Subject-Matter Domain to Master > Page 91 AI essentially showing students that domains, or distinctions between subjects, no longer matter.
  • Page 96 Would technology further isolate students by allowing them to learn alone on a computer? ... It turns out that not only does this not have to be the case, but thoughtful use of technology can actually increase human-to-human interaction.
  • Page 97 A properly designed AI can take things even further and actually facilitate conversations among human beings.
  • Page 98 With AI and AI-based tutors, the days of static learning are over as collaborative education expands in scope and capability.
  • Page 102 Working in the Artificial Intelligence Laboratory, Joseph Weizenbaum developed ELIZA, one of the world's first natural-language-processing computer programs that was able to simulate a conversation. Weizenbaum set out to show how artificial the communication was between a human and an AI, but instead he was shocked to learn that many people became emotionally attached to the artificial intelligence. Those using ELIZA often forgot that they were conversing with a computer. ... patient spoke to DOCTOR, and DOCTOR responded as a Rogerian therapist might, by using nondirectional questioning and reframing statements.
  • Page 103 We also know that both mental health problems and academic underperformance have a similar cause in students: an innate sense that they have little control and purpose in their lives, he says.
  • Page 107 Because AIs aren't sentient, they can't be truly empathetic. Empathy involves sensing and modeling other's emotions and contexts in your own mind. They can, however, simulate empathy quite well.
  • Page 109 break down challenging concepts into more manageable chunks.
  • Page 110 Still, a gap remained between what you could learn from articles, videos, and exercises and what a great tutor could do. For example, the benefits of rapport, motivational support, and dynamic conversation between the technology and the student remained elusive. For these interactive qualities, learners still needed to turn to a parent, teacher, or tutor. ... That changed, however, with AI technology and the introduction of large language models.
  • Page 111 The AI tutor knows the material and is a subject-matter expert that can provide personalized, adaptive learning exercises based on a student's individual needs and abilities. It can also offer immediate feedback and adjust its teaching methods. Unlike any parent I know, the AI tutor has endless energy and no other job
  • Page 112 Generative AI can provide parents with resources and time that might otherwise require sessions with a therapist or reading parenting self-help books. The AI becomes a parenting coach.
  • Page 115 even if AI becomes a net positive, those moments in which we unplug and turn off our screens become much scarcer.
  • Page 116 Generative AI is neither an abdication of parental responsibility nor simply a tool for keeping an eye on our kids. Rather, like all technology before it, it is a tool that we can use to amplify our intent.
  • Page 117 Large language models can focus the learning time and create more space for other points of productive contact that a child has with parents and other people.
  • Page 135 The internet is a useful but scary place, even for adults. In the late 1990s, we were all blown away by the power to search across billions of pages for answers, products, and services. However, as page views began to drive ad revenue, most websites became less about offering visitors what they actually wanted and more about persuading them to click on ads. ... This includes search. Roughly the first half-dozen links you see are actually ads. ... Because of this, you are likely to find just as much misinformation as information when you search for therapies that might help a loved one with their illness or try to more deeply understand an issue in the news.
  • Page 136 A site like YouTube could have some valuable educational content or even enriching entertainment, but it also has a lot of junk that is unhealthy for young people.
  • Page 138 Everything I've written regarding kids is arguably useful for adults as well. It would feel like browsing the internet with a thoughtful, intelligent friend who's willing to help me get to the information that I want faster. It would also protect me from unhealthy ads or information. Part VI: Teaching in the Age of AI
  • Page 143 In 2017, a slim man took the stage at the British Science Festival, one of the oldest science festivals in the world. Gazing out on a sea of leading researchers from around the globe, Sir Anthony Seldon, a renowned educator and historian, stated that by 2027 teachers will be AI rather than humans. The coming technology, he said, will force teachers to take a classroom assistant role while technology will be the conveyor of knowledge. ... agree with Seldon that personalization in learning is an aspiration that we should strive for and that AI is going to play a big role in getting us there. However, I completely disagree with his prediction that this technology will somehow minimize the importance of the human teacher. If anything, it's going to do the opposite.
  • Page 144 At the end of the day, the biggest fear from educators is the world that Seldon envisions, a world where artificial intelligence reduces demand for teachers.
  • Page 144 Engelbart believed that people were going to use technology to augment their abilities the same way that a tractor augments the work of a farmer to produce food.
  • Page 145 do not believe machines are going to relegate the teacher to the role of the teaching assistant. Rather, the AI is the teaching assistant.
  • Page 146 When students use generative AI to write papers, for instance, their quality is going to go up, akin to the advent of the word processor leading teachers to expect that their students now create beautifully typed, formatted, well-thought-out essays in ways that the typewriter did not allow. ... Students have to do a pre-mortem with their work before they turn it in for their grade.
  • Page 146 "Lectures do not make as much sense when I've got tools like ChatGPT that can do truly amazing training, all remotely,"
  • Page 149 they can do all the grunt work involved in teaching—writing rubrics, giving students feedback on their essays, and drafting student narrative progress reports for parents.
  • Page 150 By freeing educators from the administrative work that so often mires their days, artificial intelligence unlocks time and the resources for teachers.
  • Page 151 "Can you write a script for me for the section on historic repercussions of the Treaty of Versailles after World War I?"
  • Page 153 A teacher can push the AI further still and ask it to help come up with a creative student project based on the lesson.
  • Page 157 Rather than facilitate cheating, the AI can give real-time feedback and support on how to write better.
  • Page 161 having students work on their writing and papers in class allows them to get support from both the professor and other students. It makes class time more active. Longer essays can be done over multiple class periods. ... students should do what used to be homework in class and watch recorded lectures on their own time.
  • Page 161 there are important benefits to having students write essays independently, such as developing their ability to plan and not procrastinate, both of which are skills arguably as useful as learning to write.
  • Page 164 Before generative AI came on the scene, it could take days or weeks before students got feedback on their papers. ... By that point, they may have forgotten much of what they had written, and there wouldn't be a chance for them to refine their work. Contrast this to the vision in which students receive immediate feedback on every dimension of their writing from the AI. They will have the chance to practice, iterate, and improve much faster.
  • Page 171 artificial intelligence continues to play a transformative role in bridging the global education divide and fostering equal-opportunity learning for all.
  • Page 172 a student doesn't keep pace in understanding a foundational concept, the class keeps moving. Limited support exists for personalization or for revisiting gaps, much less for one-on-one tutoring.
  • Page 173 A platform like Khanmigo exists to bridge this gap—offering personalized, accessible, and high-quality education.
  • Page 174 Generative AI adds a new layer of expenses beyond the cost of paying the salaries of engineers, designers, product managers, and content developers to iteratively improve a platform like Khanmigo. ... large language model like GPT-4 are significant. Right now, our best estimate of the computation costs of average usage of Khanmigo is between five and fifteen dollars a month per user.
  • Page 186 I've already argued that generative AI is going to transform schoolwork and grading in the classroom; students will be able to do much richer assignments, and teachers will have more support grading them. I've also discussed how standardized assessment is likely to change. Assessment will be deeper, much more continuous, and indistinct from learning. Over time, either standardized tests like SATs and ACTs will move in this direction, or new assessments will enter the space to take advantage of the opportunity.
  • Page 188 admission directors need to wrestle with whether this entire exercise of writing essays even still provides a credible signal for admissions. ... it's worth questioning why essays and recommendations are part of admissions in the first place.
  • Page 192 Rather than essays or recommendations alone, what if the AI could do extensive text- or voice-based interviews with students, guidance counselors, and teachers? Part IX: Work and What Comes Next
  • Page 197 In 2023, IBM announced that it was suspending or slowing back-office hiring by 30 percent over a five-year period for jobs that could ultimately be done by AI. IBM's revelation suggests that the future of work is going to roll out differently, with back- or middle-office jobs disappearing, together with non-client-facing roles involving tasks such as creating budgets, managing data, and organizing records. Reading the tea leaves, we can see where this might be going. ... Writers and copywriters using AI could potentially be three to five times as productive.
  • Page 198 The copywriters and technical writers who are going to survive are going to be the ones who lean in most on AI to increase their productivity. The other 90 percent are going to have to find something else to do.
  • Page 202 even if the generative AI can write pieces of code, you really need to know how those pieces can fit together."
  • Page 204 "Exactly what the job market of tomorrow looks like is very hard to predict, but the deeper the skill set, whether it's medical consultation, scientific thinking, or customer support, the more value it's going to have, even in a world where productivity will be enhanced by AI," Bill Gates tells me. Not only is there more reason than ever for kids to continue to learn about their fields of interest, he says, but students need to accelerate learning these skills, and to learn them as well as possible. "Entry-level jobs are going to require people to understand how to use large language models and all of the tools they offer. You'll need them to create everything from invoices to business plans. The workplace is going to encourage its workforce to come up with the best product it can. The higher your skill level is, the more your skill will retain a substantial value in the workforce. It's the workforce plus the AI, working together."
  • Page 205 We are entering a world where we are going back to a pre– Industrial Revolution, craftsmanlike experience. A small group of people who understand engineering, sales, marketing, finance, and design are going to be able to manage armies of generative AI and put all of these pieces together. ... From an economics point of view, entrepreneurship is really the creativity of knowing how to put resources together in order to create value. ... I believe all human beings are born highly creative and entrepreneurial. Unfortunately, our Industrial Revolution–designed education system unintentionally suppresses both traits.
  • Page 206 As Bill Gates mentioned, the successful workers of the future will be those with deep and broad skills. ... The "three Rs" of reading, writing, and arithmetic are more important than ever.
  • Page 218 The traditional labor pyramid—with less-skilled manual labor forming the bottom layer, bureaucratic white-collar jobs making up the middle layer, and highly skilled knowledge work and entrepreneurship making up the top—no longer applies.