5

Artificial Intelligence and Education

Improvements in access to education and rising educational attainment across successive birth cohorts have played a key role in U.S. economic growth over the past 150 years.¹ Rapid expansions in education from the high school movement in the early 20th century and from increased college access in the mid-20th century helped foster shared prosperity in the face of rapid technological change and automation from electrification and then computerization that increased the demand for more educated workers. A slowdown in educational advances and large socioeconomic status gaps in access to high-quality educational opportunities in the late 20th and early 21st centuries have contributed to rising wage and income inequality, increased college and post-college wage premiums, and stagnating earnings for noncollege workers.² Growing income inequality has been associated with rising achievement and educational attainment gaps between high- and low-income children in the United States in recent decades.³ The COVID-19 pandemic generated new educational challenges with school closures and large differences across schools, neighborhoods, and families in resources and preparedness for remote learning. The result has been a substantial slowdown in learning, decline in reading and math test scores, and expanded achievement gaps in K–12 schooling by socioeconomic status and race.⁴

___________________

¹ C. Goldin and L.F. Katz, 2008, The Race Between Education and Technology, Harvard University Press.

² D. Autor, C. Goldin, and L.F. Katz, 2020, “Extending the Race Between Education and Technology,” AEA Papers and Proceedings 110:347–351.

³ G.J. Duncan, A. Kalil, and K.M. Ziol-Guest, 2017, “Increasing Inequality in Parent Incomes and Children’s Schooling,” Demography 54(5):1603–1626.

⁴ The Nation’s Report Card, 2022 “Reading and Mathematics Scores Decline During COVID-19 Pandemic,” https://www.nationsreportcard.gov/highlights/ltt/2022; R. Jack, C. Halloran, J. Okun, and E. Oster, 2023, “Pandemic Schooling Mode and Student Test Scores: Evidence from US School Districts,” American Economic Review: Insights 5(2):173–190; and D. Goldhaber, T.J. Kane, A. McEachin, E. Morton, T. Patterson, and D.O. Staiger, 2023, “The Consequences of Hybrid and Remote Instruction During the Pandemic,” American Economic Review: Insights 5(3):377–392.

Technological change impacts the labor market demand for different types of expertise and tasks, thereby changing the demand for education and training. But new technologies can also impact how education and training are provided. The recent, rapid rise of artificial intelligence (AI) and large language models (LLMs), especially chatbots such as ChatGPT, has led to widespread fears among educators and parents that AI may negatively disrupt education by creating increased cheating opportunities, providing students directly with the answers so that they do not learn key concepts, and spreading misinformation.⁵

At the same time, AI has the potential to transform the education system and improve learning outcomes for K–12, post-secondary schooling, and workforce training by making education more personalized, engaging, and cost-effective. To reach this promise, public and private investments likely will be needed to increase access to high-speed Internet connections and online learning opportunities, incorporate safeguards into AI-enhanced education technology, test the effectiveness of specific uses of AI in education, and train teachers to be more comfortable with and take advantage of generative AI tools and other computer-assisted learning (CAL) technologies. Under a scenario of appropriate investments and research, AI tutors and AI-enhanced technologies could operate as collaborators with students and teachers to facilitate enhanced learning rather than AI just providing a shortcut to the answers without deeper student engagement in learning.⁶ Broad access to high-quality AI tutors for students and AI teaching assistants for teachers also could help reduce the educational inequalities from differences in family and school resources that have become even starker in the wake of the COVID-19 pandemic. In the most optimistic case, all students would have a world-class and engaging virtual AI tutor with them both at school and at home; in-class assignments and homework would be adaptive and self-grading; active learning would be enhanced, and student assessments could focus more on what students can do rather what they retain in the short run; teachers would be empowered to be more creative and effective instructors and mentors; and AI could be leveraged to enhance post-secondary and lifelong learning opportunities.

This chapter will first explore what is currently known about how AI potentially can be deployed to improve learning outcomes for K–12, post-secondary schooling, and adult workforce training. It will then report on how the education system may need to adapt to current and expected changes in the training and continuing education needs of the workforce from the possible impacts of AI on skill demand and career opportunities.

___________________

⁵ See, for example, K. Huang, 2023, “Alarmed by A.I. Chatbots, Universities Start Revamping How They Teach,” New York Times, January 16, https://www.nytimes.com/2023/01/16/technology/chatgpt-artificial-intelligence-universities.html.

⁶ For a vivid and optimistic assessment of AI’s promise to positively transform education, see S. Khan, 2023, “How AI Could Save (Not Destroy) Education,” TED2023, April, https://www.ted.com/talks/sal_khan_how_ai_could_save_not_destroy_education.

ARTIFICIAL INTELLIGENCE AS AN INPUT FOR EDUCATION

AI has the potential to ameliorate key problems in the education production process. Student outcomes are likely to be enhanced when students are highly motivated and actively engaged in learning, feedback is timely, students can relate to and see value in the material, lessons are immersive, and teaching is sufficiently individualized to be at the right level to push students toward their learning frontiers.⁷ But heterogeneity in student backgrounds, preparation, and learning styles makes maintaining such conditions difficult in traditional classroom settings.

Education technologies incorporating AI promise to help surmount barriers in traditional education delivery in at least three ways: personalization, enhanced motivation and engagement, and more timely and broad access to frontier learning opportunities. First, CALs trained using machine learning methods on large data sets of student experiences and performances should be able to identify groups of students facing similar difficulties as well as those students able to handle more advanced material, which would permit more cost-effective personalized instruction potentially enlivened using virtual tutors (or tutoring bots) enhanced by LLMs. Such technologies can also provide teachers with dashboards to monitor individual student learning progress and virtual teaching assistants to help teachers leverage data-driven tools to target group instruction at a more appropriate level.⁸

Second, AI and virtual reality technologies can provide effective, immersive, and engaging learning experiences (including more realistic simulation environments for practical occupational training) to increase student motivation and engagement with material and in learning. Education technologies can further provide salient and timely nudges to parents, teachers, and students to try to offset behavioral biases that often lead to procrastination and disengagement from learning and create barriers to applying to and matriculating in college or training programs.⁹

Third, broader access to AI-enhanced educational technologies might particularly benefit less advantaged students and those in lower-resourced school districts and post-secondary institutions by providing them with similar learning opportunities and

___________________

⁷ For a discussion of the promise of AI to improve education along such dimensions see Khan Academy, n.d., “A New Chapter in Education,” https://www.khanacademy.org/college-careers-more/ai-for-education/x68ea37461197a514:unit-teaching-with-ai/x68ea37461197a514:getting-started-with-ai-in-the-classroom/a/ai-a-new-chapter-in-education-begins, accessed June 29, 2024.

⁸ Evidence on learning gains from data-driven instruction and appropriate targeting of teaching range from high-performing U.S. charter schools to the Teaching at the Right Level approach developed by the Indian nongovernmental organization Pratham. See, for example, W. Dobbie and R.G. Fryer, Jr., 2013, “Getting Beneath the Veil of Effective Schools: Evidence from New York City,” American Economic Journal: Applied Economics 5(4):28–60; and J-PAL, 2022, “Teaching at the Right Level to Improve Learning,” https://www.povertyactionlab.org/case-study/teaching-right-level-improve-learning.

⁹ On the promise and limitation of behavioral science and current nudging interventions for education see P. Oreopoulos, 2021, “Nudging and Shoving Students Toward Success,” Education Next 21(2).

assistance to those already available to more advantaged students. Sufficient investments in AI-based CALs, broadband access, and teacher training and professional development to take advantage of the new technologies in less advantaged areas could serve to narrow large and persistent educational inequalities by family socioeconomic status and community resources. But without appropriate investments in improved access for those who are less advantaged, richer households and schools as well as more advanced learners are likely to take greater advantage of new AI-based learning tools, potentially serving to expand educational inequalities as has seemed to be the case for other recent educational technologies such as massive open online courses (MOOCs) and as experienced with remote learning for K–12 and community colleges during the COVID-19 pandemic. AI also has the potential to provide virtual tutors and better online course and program information to support online post-secondary degree programs, online credential programs (such as those offered by Coursera), and MOOCs. Still, similar complementary investments in human teaching assistants, human advisors, and more universal access to reliable broadband and appropriate digital devices will be needed to make such opportunities feasible for many adult learners and less advantaged post-secondary students.

Much research documents substantial learning gains from more personalized teaching and from access to adaptive CALs.¹⁰ In particular, Benjamin Bloom in the 1980s highlighted large (2 standard deviations) gains of personalized one-on-one instruction by a human tutor over conventional classroom instruction in small-scale randomized controlled trials (RCTs).¹¹ Bloom characterized this huge learning gap between personalized and traditional classroom instruction as the “2 sigma problem” because individual instruction is sufficiently more expensive than traditional group classroom instruction, making it infeasible within the budgetary constraints of public education systems. Bloom’s clarion call motivated experimentation and research to assess the effectiveness of one-on-one or small group tutoring as a supplement to classroom instruction for part of the school day or after school. A recent meta-analysis by Andrew Nickow, Phillip Oreopoulos, and Vincent Quan of well-crafted evaluations (using RCTs or regression discontinuity designs) of interventions providing human tutors to individual students or small groups of students found “consistent and substantial positive impacts on learning outcomes” typically of around 0.37 standard deviation. But these impacts are certainly not close to as dramatic as the 2 standard deviations gain of Bloom’s full move to adaptive one-on-one instruction. Tutoring programs tend to have larger learning impacts

___________________

¹⁰ See, for example, A.V. Banerjee, S. Cole, E. Duflo, and L. Linden, 2007, “Remedying Education: Evidence from Two Randomized Experiments in India,” Quarterly Journal of Economics 122(3):1235–1264; and K. Muralidharan, A. Singh, and A.J. Ganimian, 2019, “Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India,” American Economic Review 109(4):1426–1460.

¹¹ B.S. Bloom, 1984, “The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring,” Educational Researcher 13(6):4–16.

in earlier grades, when the tutors are teachers or paid paraprofessionals as opposed to volunteers and parents, and when done during the school day.¹²

Thus, more personalized instruction can greatly improve learning outcomes, but small group instruction and one-on-one instruction with human teachers and tutors are quite resource-intensive. Attempts to scale tutoring and make it more cost-effective by going online became urgent in response to school closures and learning losses from the COVID-19 pandemic in 2020–2022. Such efforts have included one-on-one virtual tutoring for disadvantaged middle school students using volunteer college students, group virtual peer tutoring (such as Schoolhouse World), and even low-touch text messages to support parents in educating their children (especially in developing countries).¹³ But such online tutoring programs and low-touch attempts at helping parents personalize learning have faced challenges ranging from low uptake by children and parents to difficulties in recruiting and retaining effective human tutors.

CALs have played an increasing role in recent decades in trying to supplement and personalize learning opportunities. CALs have ranged from personal computers and tablets in the classroom to an expanding array of online educational resources. Recent RCTs evaluating educational interventions using CALs have found substantial learning gains of 0.2 to 0.4 standard deviations for well-implemented interventions but also some with little or no impacts.¹⁴ Educational software designed to help students learn specific skills working at their own pace has shown great promise for improving learning outcomes, particularly in math. But CALs have often been clunky and unengaging for students; teachers have not been adequately trained in the use or sufficiently motivated to help students take full advantage of CALs; and the digital divide by socioeconomic status in access to computers and high-speed Internet across schools, neighborhoods, and households has meant CALs can expand educational inequality when many less advantaged students have difficulty accessing CALs and other online learning opportunities.

AI-enhanced CALs, especially those based on LLMs such as the Khan Academy’s Khanmigo AI virtual tutor, show promise for replicating some of the advantages of high-quality personalized instruction and potentially being able to do so at much lower cost and at scale.¹⁵ Students can work with LLMs and their virtual tutoring bots as collaborators and thought partners. For example, LLMs can lead students through steps

___________________

¹² A. Nickow, P. Oreopoulos, and V. Quan, 2020, “The Impressive Effects of Tutoring on PreK–12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence,” NBER Working Paper No. 27476.

¹³ On one-on-one volunteer virtual tutoring, see M. Carlana and E. La Ferrara, 2021, “Apart But Connected: Online Tutoring and Student Outcomes During the Pandemic,” Working Paper, Harvard University. On free online peer-to-peer tutoring, see https://schoolhouse.world. On SMS text message tutoring assistance for parents, see N. Angrist, P. Bergman, and M. Matsheng, 2022, “Experimental Evidence on Learning Using Low-Tech When School Is Out,” Nature Human Behavior 6(July):941–950.

¹⁴ M. Escueta, A.J. Nickow, P. Oreopoulos, and V. Quan, 2020, “Upgrading Education with Technology: Insights from Experimental Research,” Journal of Economic Literature 58(4):897–996.

¹⁵ N. Singer, 2023, “Not Just Math Quizzes: Khan Academy’s Tutoring Bot Offers Playful Features,” New York Times, June 8, https://www.nytimes.com/2023/06/08/business/khanmigo-tutor-chat.html.

of writing essays or solving math problems at a comfortable pace; help students check their own understanding and provide immediate feedback; try to engage students in deeper material; and connect the curricular material to students’ own lives, interests, and goals. Students can debate their virtual tutoring bot to improve their arguments in a nonjudgmental setting. Teachers, parents, and tutors can observe student progress and provide further help if a student gets stuck. An AI tutor can also provide access to information on colleges and career options, helping to supplement the work of often highly stressed guidance counselors facing huge student loads in many U.S. public high schools.¹⁶

New education technologies, such as those based on generative AI, also show promise in combating the perennial problems of student boredom and apathy by making learning more immersive, interactive, and even fun. An AI assistant like Khanmigo can work alongside a student as a digital coach not only to personalize learning but also to make the activity more engaging—for example, through questions and feedback via the Socratic method or by allowing students to have discussions and debates with historical figures, pop culture icons, or characters from a literary or cinematic work. Math word problems, for example, can be rewritten by LLMs for each student, using their favorite hobby or sports team as the motivating topic.

However, inaccurate information and tutoring advice remain a concern with LLMs and AI virtual tutors. Furthermore, feedback from a virtual tutor may not be as rewarding and motivating as that from human teachers and peers. An open practical and research question is the extent to which interactions with an AI virtual tutor can be arranged to complement those with human instructors and classmates.

Augmented and virtual reality devices also may soon become more widely available to make learning more like an immersive simulation exercise or video game, increasing both realism (connection to practical real-world applications) and student engagement. Such approaches are currently used for some professional training such as flight simulation training in aviation, commercial truck driving simulated training, some community college programs, and simulation-based training to deal with high-pressure or dangerous situations for the military and the police.¹⁷ Simulation-based training might be more attractive and effective for nontraditional students who fail to thrive in traditional classrooms and for adult learners. An important hypothesis meriting study is the extent to which the twin promises of AI-based CALs to personalize instruction (teach at the right level) and to make education more immersive reinforce each other in spurring learning gains.

___________________

¹⁶ On the longer-term educational benefits of access to more effective guidance counselors, see C. Mulhern, 2023, “Beyond Teachers: Estimating Individual Guidance Counselors’ Effects on Educational Attainment,” Working Paper, August 23, http://papers.cmulhern.com/Counselors_Mulhern.pdf.

¹⁷ On simulation training for the police and military, see, for example, RAND, n.d., “Simulation-Based Training: Use of Virtual Environments in Diverse Contexts,” https://www.rand.org/well-being/justice-policy/projects/simulation-based-training.html, accessed June 29, 2024.

Potential downsides for educational progress from access to LLMs include students using LLMs for cheating or plagiarism, or students directly getting the answers from ChatGPT without going through the intermediate steps and failing to engage with crucial material more deeply. It is unclear at present the extent to which AI-based shortcuts for artistic and design assignments could serve to dull student creativity or free students to be more creative. Hallucinations and misinformation remain an issue for ChatGPT and other transformer-based LLMs. CALs using AI virtual tutors, such as Khanmigo, can be designed with safeguards to try to avoid such pitfalls—for example, pushing students to work out answers on their own by checking their comprehension and providing nudges and subtle hints at each stage of the problem-solving process instead of giving them the answers directly. Furthermore, students working in groups could try to critique answers provided by an LLM such as ChatGPT to help build critical decision-making and judgment skills and ferret out misinformation. In fact, Finland has integrated a program of media literacy to help students recognize misinformation online as part of its core curriculum.¹⁸ Prompt engineering training overall and in specific domains to get more accurate information from LLMs likely will need to be integrated into educational curricula at all levels.

A further worry is that increased reliance on LLMs and virtual tutors in education could crowd out time spent interacting with human teachers and with a diverse group of other students (human peers). Less time spent in high-quality classroom social interactions with peers could have adverse impacts on the development of valuable soft skills (e.g., social interaction, communication, and teamwork skills). A key issue going forward is to find ways to integrate AI-based CALs into interactive group learning and problem-solving activities among students.

Universities, professional schools, and individual faculty are also grappling in real time with whether and how to incorporate ChatGPT and other generative AI tools into classes given the now widespread use by many students of these readily available online resources. Current policies of individual instructors range from banning any use of ChatGPT in a course to active engagement with ChatGPT in classroom discussions including critiques of the answers provided by ChatGPT. In fact, Tyler Cowen and Alex Tabarrok have already put together an initial guide for how best to use LLMs to supplement college-level instruction in economics.¹⁹ AI-enhanced online tools have the potential to provide students with immediate feedback and guidance on projects (such as coding exercises in a computer science class), provide hints on how to improve work, automate more standardized aspects of grading and assessments, and alert faculty and

___________________

¹⁸ J. Gross, 2023, “How Finland Teaches Even Youngest Pupils to Spot Misinformation,” New York Times, January 11, p. A10.

¹⁹ T. Cowen and A. Tabarrok, 2023, “How to Learn and Teach Economics with Large Language Models, including ChatGPT,” GMU Working Paper in Economics, No. 23-18, May 12.

teaching fellows to which students need further help. Thus, generative AI tools might take over some of the more routine aspects of instruction and grading, allowing for a more efficient allocation of faculty and teaching fellow instructional time and better targeting of one-on-one or small group human tutoring. AI also threatens the viability of take-home exams and assessments and might motivate a rethinking of assessments to evaluate student learning-by-doing and the application of knowledge in novel situations rather than just demonstrating retention and mastery of course concepts.²⁰

Important levers for realizing the potential of AI-enhanced CALs are getting buy-in from teachers, providing teachers with training and coaching on the new CALs and how to use them to better personalize instruction and teach at the right level, and making sure teachers have time to integrate such learning aids into their lesson plans. Inadequate teacher engagement and support could be a major bottleneck limiting the impact of AI on educational outcomes. Studies of education technology interventions often show little (or even negative) impact on student learning when technology is just plopped into a classroom with little guidance provided to teachers.²¹ A recent evaluation of the Khoaching with Khan Academy program, which provides coaches (“Khoaches”) to teachers in grades 3 to 8 to better use Khan instructional videos as a CAL to help offset COVID-19 pandemic learning losses, finds a central role for teacher buy-in and that training, supporting, and guiding teachers to adapt a CAL can be an effective way of generating meaningful test score gains.²²

One potentially successful route toward teacher buy-in involves employing CALs for homework assignments rather than as alternatives to existing in-class teaching methods. For example, the CK-12 Foundation’s website²³ provides a range of K–12 lessons and associated questions to test student understanding. Many teachers assign specific CK-12 lessons, and the associated questions, as homework to review and enhance the material covered in class. Teachers appreciate the fact that this CAL reduces their workload by automatically grading the homework assignments and producing a summary for the teacher regarding which topics create the most difficulty for the students. This adoption of CALs for homework thus reduces overall teacher workload without requiring changes to their current in-class teaching methods. Furthermore, once the homework is being performed online, it is easy to introduce AI technology to create more intelligent and personally customized homework. Each homework

___________________

²⁰ See D. Deming, 2023, “Generative AI Is a Black Mirror for Educators,” Forked Lightning, September 5, https://forklightning.substack.com/p/generative-ai-is-a-black-mirror-for.

²¹ See, for example, S. Beg, W. Halim, A.M. Lucas, and U. Saif, 2022, “Engaging Teachers with Technology Increased Achievement, Bypassing Teachers Did Not,” American Economic Journal: Economic Policy 14(2):61–90.

²² P. Oreopoulos, C. Gibbs, M. Jensen, and J. Price, 2024, “Teaching Teachers to Use Computer Assisted Learning Effectively: Experimental and Quasi-Experimental Evidence,” NBER Working Paper No. w32388, National Bureau of Economic Research.

²³ The website for the CK-12 Foundation is ck12.org, accessed June 29, 2024.

question presented to the student by CK-12 depends on the earlier questions the student answered correctly or incorrectly, providing a customized sequence of questions that adapts to each student’s level of understanding. Furthermore, if a student answers a question incorrectly, a hint and an opportunity to reanswer the question are provided. As of January 2023, the CK-12 system chooses which of several hints to provide to the student, based on which hints have led to the best learning outcomes for previous students who made the same mistake.²⁴ In this way, the more students use CK-12, the better it learns to teach future students. The CK-12 experience illustrates an important potential for CALs, namely their ability to learn to teach better based on observed outcomes of previously taught students. Such increasing returns are especially attractive given the potential scale of CALs; for example, CK-12 has already taught more than 200 million students over the past 15 years, far more than a human teacher could ever reach (or learn from) in a lifetime.

The recent COVID-19 pandemic experience with remote instruction raises worries of large disparities among students (and among educational providers) in access to high-speed Internet connections and digital devices, state-of-the-art learning technologies, and appropriate spaces for learning. A concern is that more advantaged educational institutions and students are likely to be better positioned to take advantage of new AI-based CALs. To try to prevent new educational technologies from further widening educational inequalities by socioeconomic status, a major policy issue going forward will be to make sure all schools, teachers, and students are provided broadband access, appropriate digital hardware, and proper training and support to benefit from AI-enhanced CALs and online learning opportunities.

More rigorous research also is needed to assess the actual educational efficacy and cost-effectiveness of AI-enhanced CALs. For example, there is not yet much clear evidence on the impacts on student learning of AI virtual tutors (such as Khanmigo) alone versus either no tutor or a human tutor. Educational practice would further benefit from research on whether hybrid models combining lower-intensity human tutors plus AI virtual tutors can generate learning gains equivalent to more intensive human one-on-one instruction at lower cost in K–12 schools, post-secondary education settings, and worker and professional training programs. Other important practical and research questions include (1) understanding the impacts of greater reliance on AI-based tools on student critical thinking and creativity, and (2) developing approaches to help teachers incorporate LLMs and AI-based CALs into interactive student group activities to better develop teamwork skills, improve respect for others, and foster innovation.

___________________

²⁴ R. Schmucker, N. Pachapurkar, S. Bala, M. Shah, and T. Mitchell, 2023, “Learning to Give Useful Hints: Assistance Action Evaluation and Policy Improvements,” pp. 383–398 in European Conference on Technology Enhanced Learning, O. Viberg, I. Jivet, P.J. Muñoz-Merino, M. Perifanou, and T. Papathoma, eds., Springer Nature Switzerland.

IMPLICATIONS OF LIKELY ARTIFICIAL INTELLIGENCE IMPACTS ON THE LABOR MARKET FOR EDUCATION

Past technological revolutions have substantially altered production processes and the demand for different tasks, types of expertise, and skills, thus altering the educational and training needs of the workforce as discussed in the previous chapter. The industrial era from the late 19th to mid-20th centuries was associated with an increase in the demand for mass expertise (numeracy and literacy) to perform precise rules-based tasks in production and office jobs. Workforce education needs in the industrial era were primarily met through mass access to secondary schooling fostered by the high school movement, and supplemented by work experience and employer-provided on-the-job training.²⁵ The subsequent information era of advances in classical computing in the late 20th century and early 21st century was associated with the automation of many routine manual and cognitive tasks, reducing demand for the types of middle-skill production and office jobs that provided pathways to middle-class incomes in the industrial era. The automation and reduced cost of routine tasks complemented highly educated workers with college and graduate training by increasing the demand for elite expertise combining expert knowledge with acquired judgment to make high-stakes decisions in nonstandard cases. Classical computing was less able to substitute for the nonroutine manual tasks and social and interactive skills required in many in-person service jobs (e.g., home health aides, janitors, and baristas). The information era was associated with a polarization of the labor market, increased returns to college and (especially) professional and graduate degrees as well as to social and communication skills, and challenges related to resources for college preparation and access to quality colleges for students from less advantaged households.

The open question for the education system raised in Chapter 4 is that of the likely impacts of advances in AI (and the emerging “artificial intelligence era”) on the labor market demand for human tasks, skills, and expertise. One scenario is that AI and robotics are largely implemented in the same manner of past classical computing gains to automate a broader range of routine cognitive and manual tasks, taking over the “last mile” of many mass expertise tasks from expense accounting to producing standardized legal documents to physical tasks in controlled and predictable settings. Under this scenario, AI would further reduce the demand for mass expertise (and some middle-skill jobs) and possibly increase the demand for elite expertise and for nonroutine in-person services emphasizing social and communication skills. Some initial evidence indicates that the surge in AI activity since 2015 has been associated with increased vacancies

___________________

²⁵ C. Goldin and L.F. Katz, 2008, The Race Between Education and Technology, Harvard University Press.

for AI skills themselves and reduced vacancies for previously sought skills in more AI-exposed establishments, suggesting some AI-based automation of previous tasks and increased skill requirements in new job openings.²⁶ An implication of this scenario for education practice is to increase preparation and access to higher education for professional positions requiring elite expertise and specialized knowledge and to further emphasize problem solving, abstract thinking, and social and communication skills.

AI-based software and products increasingly can take over certain (often routine) tasks in professional positions relying on elite expertise, from note taking and medical record inputting for physicians to standard writing tasks for lawyers and other professionals.²⁷ Reduced time on “paperwork” and standardized tasks can free up professionals to use their judgment and decision-making expertise on more complicated cases and to communicate and be more engaged with patients and clients. AI and large databases also can provide professionals with improved decision aides and effective copilots for handling complicated cases and decisions. Machine learning algorithms trained on large samples have the capacity to notice patterns that people might not and thereby offer promise to improve diagnoses by medical professionals and to help scientific researchers and other practicing professionals to develop and explore new hypotheses.²⁸ The shift to a data-rich environment with more information available to both professionals and their clients and patients may further change elite expertise jobs toward a focus on synthesizing information and making tough judgment calls based on tacit knowledge and experience as opposed to being a font of information not available to others. Such shifts in professional jobs based on elite expertise should motivate changes in curriculum in professional schools and encourage more frequent ongoing professional training to help professionals take advantage of AI-based tools to streamline more routine activities and to improve their communication and information-synthesizing skills (e.g., improving “bedside” manner for health care professionals).

A second possible scenario sketched in the previous chapter is that if AI increasingly substitutes for elite expertise, insights from elite expert knowledge will become accessible to a broader range of workers, making the expertise less scarce. AI advances and diffusion in this scenario may increase the demand for translational expertise combining foundational technical knowledge with supporting tools (likely to increasingly

___________________

²⁶ D. Acemoglu, D. Autor, J. Hazell, and P. Restrepo, 2022, “Artificial Intelligence and Jobs: Evidence from Online Vacancies,” Journal of Labor Economics 40(S1):S293–S340.

²⁷ S. Lohr, 2023, “A.I. Outshines in Health Care. At Paperwork,” New York Times, June 26, p. A1.

²⁸ J. Ludwig and S. Mullainathan, 2024, “Machine Learning as a Tool for Hypothesis Generation,” The Quarterly Journal of Economics 139(2):751–827, https://doi.org/10.1093/qje/qjad055; and G. Charness, B. Jabarian, and J. List, 2023, “Generation Next: Experimentation with AI,” Artefactual Field Experiments 00777, The Field Experiments Website, https://ideas.repec.org/p/feb/artefa/00777.html. On the implications of AI for medical practice see D.J. Lamas, 2023, “There’s One Hard Question My Fellow Doctors and I Will Need to Answer Soon: Guest Essay,” New York Times, July 6, https://www.nytimes.com/2023/07/06/opinion/artificial-intelligence-medicine-healthcare.html.

use AI) to accomplish high-value and complex tasks that may previously have required professional workers with elite expertise in positions such as nurse practitioners and physician assistants. AI could then, for example, assist electricians and auto mechanics in dealing with difficult and unfamiliar situations, speed up coding by software developers and allow more attention to overall software architecture, and broaden the cuisine range of chefs and caterers.

A rise in the demand for translational expertise could be met through shifts in education and training emphasis toward strong basic skills (numeracy and literacy) and social/communication skills, foundational training in specific subject expertise (e.g., law, accounting, medicine, plumbing, carpentry), and experiential learning that might end up looking more like a combination of post-secondary vocational training and a liberal arts education. The cooperative education model for higher education long used at Northeastern University that alternates semesters of academic study with full-time work to gain hands-on learning and foster translation expertise is a promising approach that could be adept at expanding post-secondary education opportunities to a broad group of students in the AI era.²⁹ The New York City P-TECH Grades 9–14 high school model, which combines accelerated high school course work, early college, and work-based learning experiences, could play a similar role in providing valuable labor market skills and associate’s degrees.³⁰ Sectoral employment training programs and registered apprenticeship programs that provide occupational and soft-skills training and wraparound services with strong connections to employers in a growing sector and input from unions and/or community groups represent another promising and flexible training model for jobs requiring translational expertise. Recent evaluations find that sectoral employment programs providing training, internships, and support service lasting from several months to 2 years have generated substantial and persistent earnings increases of 12 percent to 34 percent following training.³¹

Community college partnerships with consortia of local employers to provide training for emerging jobs with upward mobility potential also remain a crucial part of the post-secondary education system both for new high school graduates and for returning students attempting to augment or change careers.³² And much evidence indicates that programs of wraparound support for community college students, such as the Accelerated Study in Associate Programs (ASAP) of the City University of New York,

___________________

²⁹ The website for Northeastern’s cooperative education model is https://careers.northeastern.edu/cooperative-education, accessed June 29, 2024.

³⁰ M. Dixon and R. Rosen, 2022, On Ramp to College: Dual Enrollment Impacts from the Evaluation of New York City’s P-TECH 9–14 Schools, MDRC.

³¹ L.F. Katz, J. Roth, R. Hendra, and K. Schaberg, 2022, “Why Do Sectoral Employment Programs Work? Lessons from WorkAdvance,” Journal of Labor Economics 40(S1):S249–S291.

³² R.B. Schwartz and R. Lipson, 2023, America’s Hidden Economic Engines: How Community Colleges Can Drive Prosperity, Harvard Education Press; H.J. Holzer, R. Lipson, and G. Wright, 2023, “Community Colleges and Workforce Development: Are They Achieving Their Potential?” American Enterprise Institute.

can be effective in increasing degree completion and substantially increasing earnings, although strapped budgets have limited the expansion of such effective supplemental programs at many community colleges.³³

Partnerships between community colleges and industry may also play a role in allowing a wider range of students and workers to gain AI-related skills and take advantage of AI educational tools. For example, Amazon Web Services (AWS) has established a “machine learning university” with a new free program to help community colleges and universities teach database as well as AI and machine learning concepts.³⁴ It is providing an educator-enablement bootcamp along with curriculum and course delivery support. AWS is also partnering with the California community college system, the largest in the world, to identify the barriers to student completion of their educational goals. The colleges, in turn, are redesigning their programs and providing personalized advice to students to address these barriers and increase completion rates.

It has proven to be quite difficult to predict the specific skills and training needed for emerging employment opportunities beyond the very short run. A substantial share of employment ends up in new job titles and detailed occupations that did not even exist a decade or two before and often require new expertise.³⁵ Thus, one needs to be humble in forecasting the likely impacts of continued advances in AI on jobs and workforce training needs. But it seems likely that AI will be augmenting job activities and required expertise over a typical worker’s career, reinforcing the need for an effective lifelong learning system for the workforce. Critical thinking, general adaptability, social and communication skills, and capacity for future learning are likely to remain valued worker attributes, and developing these capacities in K–12 and post-secondary schooling and ongoing adult education and training opportunities will remain crucial in the AI era. Training in more effective prompt engineering to take advantage of LLMs is likely to become essential in a broad range of occupations as well.

Progress in algorithmic predictive tools should also be helpful in providing job seekers with improved, more personalized, and better targeted information on relevant job opportunities that match their qualifications and education and training options for appropriate opportunities offering upward mobility.³⁶ AI-powered tools are already starting to affect recruiting processes—the paths by which students find the jobs for which

___________________

³³ C. Miller and M.J. Weiss, 2021, Increasing Community College Graduation Rates: A Synthesis of Findings on the ASAP Model from Six Colleges Across Two States, MDRC; C. Hill, C. Sommo, and K. Warner, 2023, From Degrees to Dollars: Six-Year Findings from the ASAP Ohio Demonstration, MDRC.

³⁴ The website for this “machine learning university” is https://aws.amazon.com/machine-learning/mlu, accessed June 29, 2024.

³⁵ D. Autor, C.M. Chin, A. Salomons, and B. Seegmiller, 2022, “New Frontiers: The Origins and Content of New Work, 1940–2018,” NBER Working Paper No. 30389, August.

³⁶ A. Bartik and B. Stuart, 2022, “Search and Matching in Modern Labor Markets: A Landscape Report,” WorkRise report, Urban Institute, https://www.workrisenetwork.org/publications/search-and-matching-modern-labor-markets-landscape-report.

they are qualified. AI tools potentially can help employers identify qualified candidates, can match job requirements to candidate qualifications, and may be able to help predict which candidates are most likely to succeed in a particular position. AI chatboxes can help job candidates with questions about job requirements and can connect them to recruiters. But systematic biases in such algorithmic prediction tools remain a concern. LinkedIn plans to roll out a new generative AI platform to help both recruiters and job seekers and has worked to try to reduce the biases in its existing AI platforms. Hired, part of the Adecco Group, offers an AI-enhanced platform to match technology and sales talent with job opportunities in major global companies.³⁷ AI-based approaches may also be able to assist with self-employment and entrepreneurial training.

FUTURE OPPORTUNITIES AND RESEARCH NEEDS

The rise of AI and student access to LLMs, such as ChatGPT, is already disrupting education at the K–12 and post-secondary levels, presenting risks of increased cheating and undermining the efficacy of traditional homework assignments and many approaches to student assessment (such as take-home exams). But AI also has the potential to greatly improve education by making it more personalized (to enhance teaching at the right level), engaging, and accessible to both traditional and nontraditional students. Much research is needed to develop new AI-based CALs and virtual tutors, to assess their impacts on learning, and to understand how to use AI more effectively to complement human teachers and enhance classroom peer interactions. And investments in more universal access to broadband and appropriate digital devices and teacher training to take advantage of AI-based educational tools will be essential to increase access to high-quality education for the less advantaged and not to further expand educational disparities by socioeconomic status. Beyond primary, secondary, and continuing education, research is also needed to understand how AI-based tools can help specifically with just-in-time job retraining.

The impacts of AI on the labor market remain highly uncertain. Nevertheless, AI is likely to impact the demand for different occupations and types of expertise greatly, as sketched in Chapter 4. Changes in the curricula of professional education programs to help new professionals work with AI-based tools will be essential. And increased access to evidence-based adult education and retraining programs will be needed to help workers avoid costly displacement and take advantage of new labor market opportunities.

___________________

³⁷ J. Kelly, 2023, “How AI-Powered Tech Can Help Recruiters and Hiring Managers Find Candidates Quicker and More Efficiently,” Forbes, March 15, https://www.forbes.com/sites/jackkelly/2023/03/15/how-ai-powered-tech-can-help-recruiters-and-hiring-managers-find-candidates-quicker-and-more-efficiently/?sh=22f155ba3a3f.