The Deskilling of Teaching and the Case for Intelligent Tutoring Systems

: This essay describes trends in the organization of work that have laid the groundwork for the adoption of interactive AI-driven instruction tools, and the technological innovations that will make intelligent tutoring systems truly competitive with human teachers. Since the origin of occupational specialization, the collection and transmission of knowledge have been tied to individual careers and job roles, specifically doctors, teachers, clergy, and lawyers, the paradigmatic knowledge professionals. But these roles have also been tied to texts and organizations that can disseminate knowledge independently from professionals. Professionals and organizations turn knowledge into texts and tools that enable lay people to access knowledge without the intermediation of professionals or organizations. In the 21 st century, one emerging tool for transmitting knowledge is the intelligent tutoring system. This paper examines how technological, epistemic, and economic trends in education are supporting the routinization, proletarianization, and automation of the occupation of teaching, leading to the increasing substitution of intelligent tutoring systems for human instruction. Some trends, such as standardized curricula and testing, both restrict teachers’ professional autonomy and facilitate the creation of pedagogical tools. Other trends reduce teachers’ ability to resist automation. The growth of adjunct teaching and paraprofessional roles in higher education allows organizations to take over and rationalize parts of the traditional teacher role. Faculty evaluations and learning outcomes assessment weaken professional claims to be the sole arbiters of instructional quality and student learning. The widespread use of intelligent tutoring systems also depends on the sophistication of software capable of performing the social-emotional and cognitive roles that educators perform. Eventually, pedagogical software will be able to interactively to the needs and interests of every learner, more cheaply, quickly, and 27 accurately than any human teacher. Assessment of learning will be continuous, and certification of 28 learning will be for specific skills instead of broad area competencies. Intelligent tutoring systems 29 will help transition education from its medieval and industrial-era model to more accessible and 30 flexible continuing education for employment and life enrichment.


35
Covid has forced an abrupt and radical reassessment of how we work, by both workers 36 and employers. Many expect that one impact of Covid will be an acceleration of 37 automation (Anderson et al., 2021;Autor & Reynolds, 2020;Lund et al., 2021). Studying the 38 impacts of epidemics from 2003's SARS to 2014's Ebola outbreak, Sedik and Yoo found that 39 these previous epidemics increased the rate of work automation for at least four years, 40 across 18 industries and 40 countries and controlling for GDP, trade, and demographics 41 (Sedik & Yoo, 2021). Of course, Covid has had a much more abrupt and profound impact 42 on work than these previous epidemics. Covid accommodations made labor more 43 expensive, and workers more reluctant to return to work, creating an acute labor shortage 44 in 2021. Employers have spent a year and a half experimenting with replacing face-to-face 45 work with remote work and automation. Surveys suggest that firms may now allow a 46 quarter of employees to continue working from home, and will shrink their office spaces 47 accordingly (Lund et al., 2021). In a 2020 survey, 43% of employers were expecting to reduce 48 their workforce (WEF, 2020). 49 50 While most economists have seen the little prospect of automation impacting teaching, 51 Covid has had a dramatic impact on education. Parents and teachers have been concerned 52 that their children have learned less online, while others have embraced online education. 53 Teachers have chafed at the insistence that they risk their lives with unvaccinated students, 54 but have struggled to keep students engaged online. In 2021 American K-12 schools faced a 55 more acute teacher shortage than they did before Covid, with fewer college students 56 choosing to enter education and the aging K-12 teacher population dropping out more 57 quickly due to Covid burnout (Zamarro et al., 2021). There are more than half a million 58 fewer public sector teachers in the United States in 2021 than there were in 2019, and the 59 number of unfilled teacher jobs is now at a 20-year high (Hoff, 2021). What if, between Covid 60 shocks, advances in artificial intelligence, and the deepening teacher shortage, automation 61 is now poised to begin re-shaping education the way that it has begun impacting other 62 white-collar occupations? 63 64 In his 1995 novel The Diamond Age: Or, A Young Lady 's Illustrated Primer Neal 65 Stephenson describes a dystopian future with persistent inequality. His protagonist is a poor 66 Chinese girl, Nell, who comes to possess an experimental, interactive, AI-driven tutoring 67 tool, the eponymous "illustrated primer." Through ordinary conversation, the AI assesses 68 Nell's interests and knowledge and begins instructing her using personalized content 69 appropriate to her context. The teaching software developed for the short-lived One-70 Laptop-Per-Child (OLPC) experiment in the 2000s was directly inspired by this device in 71 The Diamond Age (Dodd, 2012).

73
The first examples of "intelligent tutoring systems" that interactively adapt to students 74 began to be introduced back in the 1970s, however, well before recent breakthroughs in 75 machine learning and natural language processing. Meta-analyses of the effectiveness of 76 these crude algorithmic tools found that students learned as much from them as from one-77 on-one tutoring, and more than from classroom instruction (Kulik & Fletcher, 2017;Ma et al., 78 2014;Xu et al., 2019) These tools are certain to become more sophisticated and be applied to 79 more educational domains. What is less certain is their impact on the profession of teaching 80 and the organization of education. Technology futurism has a poor track record predicting 81 the speed or form of technology adoption. Will AI-guided instruction supplant human 82 instruction, or be used as a labor-extending tool by teachers who then focus on the other, 83 less automatable parts of their work? Will personalized, self-paced instruction replace age-84 based classrooms and the four-year college degree, or will school gradually incorporate the 85 new tools without major reform? 86 87 Much of the debate assumes that once a tool is available that it will be widely adopted, 88 destroying jobs and whole sectors in its wake. Many technologies with disruptive potentials 89 are never widely adopted and the literature on work has long noted how firms and workers, 90 especially professionals, resist automation. This essay will combine the sociology of work 91 with a discussion of the technical hurdles to developing teaching software to explore how 92 and when new teaching technologies will be disruptive. The essay attempts to show how 93 the systematization of knowledge necessary for the automation of occupational roles 94 supports and is accelerated by capitalist rationalization, deskilling, and proletarianization. The two oldest and strongest professions, teaching and healing, are based on the need 97 for ordinary people to turn to specialists who have mastered esoteric bodies of knowledge. 98 These fields of knowledge require a lifetime of investment, and in turn, they assert that only 99 other experts can gauge the quality of their service. Both healing and teaching emphasize 100 the importance of personal relationships -doctor-patient and teacher-student -and 101 discourage attempts to measure and systematize their work. Mandated clinical pathways 102 based on medical outcomes data are derided as "cookbook medicine," and curricular 103 guidelines measured by standardized exams are condemned for encouraging "teaching to 104 the test." The slow pace of automation in health care and education is also explained by 105 the complexity of medical and educational knowledge, and the skills needed to deploy 106 knowledge effectively. The work process in a hospital or a university is an order of 107 magnitude more complex than in most factories. Determining the most cost-effective way 108 to produce a widget is far simpler than identifying the best way to produce a cancer 109 remission or an English major. 110 111 One result of the slow pace of automation in medicine and education is that their costs 112 have inflated faster than other products and services, leading to growing pressures from the 113 market and state to ensure that consumers are receiving valid, quality services. The 114 uncertain relationship of medical services to health outcomes and higher education to actual 115 learning and careers has led to growing skepticism of both professions. Now the rapid 116 improvement in machine learning, predictive analytics, and natural language processing 117 may accelerate a century of efforts at routinization and automation. In The System of Professions, Andrew Abbott (Abbott, 1988(Abbott, , 1991 proposes that there are 122 three ways to organize expert knowledge, into professions, organizations, or commodities. 123 Each way of accessing expert knowledge has its advantages and disadvantages. Professions 124 offer strong financial and prestige rewards for those who endure the necessary training and 125 challenging work. Organizations can divide up knowledge among specialized, 126 interchangeable parts of a bureaucracy, lowering costs and improving quality control. The 127 most routine forms of knowledge can become inexpensive commodities, e.g. textbooks. The 128 medical texts of the 2 nd -century Greek physician Galen -three million words of which have 129 survived -were relied on as canonical medical knowledge for more than a thousand years. 130

131
In 19 th century America, where doctors were difficult to find, books purporting to allow 132 self-diagnosis and self-treatment were popular. As physicians professionalized in the 19 th -133 century pressure grew to standardize medical education, suppressing heterodox theories 134 like homeopathy, and codifying medicine into certification exams overseen by state-135 appointed licensure boards and administered by medical schools. As physicians adopted 136 the legitimacy of scientific empiricism consumers increasingly turned to them as 137 independent contractors. In the 20 th century, however, the proliferation of medical knowledge drove the growth 140 of hospitals, medical specialties, paraprofessionals and larger systems of care. Today 141 consumers can seek a self-diagnosis online, or turn to nurses using standardized diagnostic 142 protocols provided by their healthcare organization, or make an appointment for a 143 personalized assessment by a physician. In other words, medical knowledge is increasingly 144 created, codified, and transmitted by organizations and software. While many consumers 145 may prefer medical knowledge from their trusted professional, medical advice from an 146 organization or algorithm allows much wider, faster, cheaper, and often more accurate, 147 access (Walker, 2009). Formal education for children and adolescents in Europe emerged in the 11 th century 160 out of the tutoring of elite boys preparing for the university or clergy. Before the emergence 161 of universities in medieval Europe education was only accessible to those who could buy 162 hand-copied books and hire tutors. The university structure allowed wider access to 163 standardized knowledge and the validation of professional credentials. After the 164 Renaissance, and especially after the invention of the printing press, secondary education 165 for rich European boys expanded until states began organizing and providing universal 166 public education in the 19 th century. Even in the most egalitarian countries, however, 167 schooling reflects the class of their students, with private schools for elites modeled on 168 universities, and the children of laborers receiving standardized, age-based instruction in 169 disciplined classrooms that prepare them for the factory. While K-12 education reflects the industrial era in which it emerged, higher education 172 is still shaped by the medieval system -departments focusing on specific disciplines, 173 granting degrees, taught by relatively autonomous faculty who are also expected to 174 generate new knowledge. With these ancient origins, the organized transmission of 175 knowledge through teachers and schools has enormous cultural and political inertia. If 176 teachers as a profession and schools as an organizational form face challenges today it will 177 not be just from new teaching technologies, but their convergence with sociological trends 178 that are finally re-shaping education a century after they transformed much of the rest of 179 the economy. "The bourgeoisie has stripped of its halo every occupation hitherto honoured and looked up to 184 with reverent awe. It has converted the physician, the lawyer, the priest, the poet, the man of science, 185 into its paid wage labourers." (Marx & Engels, 1847) 186 187 In the 1970s labor economist Harry Braverman (Braverman, 1998) proposed that all 188 firms under capitalism are incentivized to deskill tasks that require expensive, skilled labor. 189 This theory suggested that competition and profits will push most firms towards "scientific 190 management" (Taylor, 1915) dividing and re-organizing discrete tasks into assembly lines 191 to maximize efficiency. This theory fit the Marxian prediction that capitalism would 192 rationalize the work of the petit bourgeoisie, making them more like wage laborers. Having 193 a group of skilled workers in a firm is not only more expensive but a challenge to the 194 rational, hierarchical management of the firm. In healthcare, this dual authority structure 195 can be seen in the tension between medical staff and hospital administration, and in 196 universities between tenured faculty and university administration. Unlike factories, 197 neither hospitals nor universities can as effectively divide up, reorganize and deskill the 198 central skills of their professionals, diagnostic and treatment decision-making in the case of 199 medicine, and instruction and research for teachers. But with rationalized accountability 200 structures and productivity pressures, such as larger classrooms, student course 201 evaluations, and standardized student learning assessment in the case of education, 202 professionals do begin to feel more like proletarian industrial workers.

204
Braverman's thesis fell out of favor when the evidence for deskilling was found to be 205 mixed, and studies showed industries creating new skilled occupations. Another problem 206 for the application of Braverman's deskilling thesis in education is that much of the sector 207 is public and not private, and the competitive and fiscal pressures forcing work 208 rationalization in the public sector are more diffuse. Nonetheless, private sector work 209 organization and managerial methods bleed over into the public sector when state and local 210 budgets squeeze schools and universities. Politicians demand the measurement of 211 educational productivity on the behalf of taxpayers. Schools are pressured to gear education 212 more to the labor market, and less to education for its own sake or to prepare critical citizens. 213 Fiscal pressures in higher education encourage hiring cheap adjunct instructors and fewer 214 tenured faculty. 215 216 While the proletarianization of teaching reduces labor costs, teachers have been able to 217 resist work rationalization, intensification, and automation because schools cannot yet 218 substitute technology for their core skills. For schools and pedagogical software to seriously 219 challenge teaching teachers' core skills and tasks need to be decomposed into standardized, 220 interchangeable parts. Most occupations require multiple skills, but the specific skills and their importance 226 vary within occupations. The core skill of primary care physicians is diagnosis and 227 treatment decision-making, but they are also expected to have social-emotional skills to 228 understand and educate their patients. If physicians employ physician assistants and nurse 229 practitioners, then they can outsource some of the preliminary diagnostics, history-taking, 230 and patient education, and spend more time on a larger patient load using just their core 231 skills. Social workers, therapists, and dozens of other specialties now cooperate in providing 232 patient care, organized around electronic medical records, with the physicians as the central 233 decision-makers. Physicians can maintain their dominant role so long as their core skill is 234 immune to industrial de-skilling and automation. 235 236 A decade ago Frey and Osborne (Frey & Osborne, 2013) used a classification of the 237 tasks and skills involved in 702 occupations to estimate each occupations' vulnerability to 238 automation. Their study suggested that about half of American jobs were vulnerable to 239 automation in the next decade or two. Replication of this analysis in many industrialized 240 countries found roughly the same results. They argued that highly educated occupations in 241 general, and K-12 and postsecondary teachers in particular, were among the least vulnerable 242 to automation. Nonetheless "Even education, one of the most labor-intensive sectors, will 243 most likely be significantly impacted by improved user interfaces and algorithms building 244 upon big data" (Frey & Osborne, 2013). As with Braverman's deskilling theory, labor economists have pushed back on the Frey 247 and Osborne analysis by pointing to the flexibility within occupations to re-focus on un-248 automatable skills and to use automation to extend an occupation's productivity instead of 249 replacing their jobs (Arntz et al., 2016). Assuming that workers can adapt to automation by 250 reallocating their time, critics estimate that far fewer occupations are vulnerable to 251 automation. Workers with more power and prestige, like professionals, are more able to 252 redefine their role and prevent their occupation from being eroded by technology. 253 Occupations that require higher education are not only less likely to be replaced by 254 automation because of the difficulty in automating their tasks but also because they can use 255 their professional authority to defend their core skills and use new technologies as labor 256 extenders. 257 258 In 2020 the McKinsey consulting firm published a study of the impact that technology 259 can have on the reallocation of K-12 teachers' work time (Bryant et al., 2020). They asked K-260 12 teachers in Canada, Singapore, the United Kingdom, and the United States how much 261 time they spent on 37 different activities. They found that, on average, teachers spent half 262 (49%) of their 50-hour workweek in direct interactions with students, including instruction, 263 coaching, and "behavioral, social and emotional skill development." The other half of their 264 week was spent on preparation, grading, professional development, and administration. 265 McKinsey then estimated the impacts of new technologies on the time demands in these 266 various tasks, concluding that the biggest time savers would be on tasks like curricular 267 preparation, grading, evaluation, feedback, and administration, with the smallest impacts 268 on student interaction. For instance, software "packages to help teachers assess the current 269 level of their students' understanding, group students according to learning needs, and 270 suggest lesson plans, materials, and problem sets for each group" could cut the time spent 271 on these tasks in half. The roughly 13 hours saved could then be spent on more personalized, 272 individual instruction (or having a better work-life balance).

274
McKinsey's analysis is presented as an optimistic vision that K-12 teachers can 275 embrace, rather than a rationale for fewer teachers and larger classes. But K-12 teachers are 276 less powerful than university faculty, and their skills, which require less education, are more 277 amenable to standardization and rationalization. One way K-12 teaching has been prepared 278 for deprofessionalization is through the standardization of K-12 curricula.  (Susskind & Susskind, 2016b) argue that the 291 standardization and systematization of knowledge is the key prerequisite for the 292 deconstruction of the professions. The degree to which a body of knowledge has been 293 standardized is also the degree to which consumers can access it from non-professionals in 294 an organization, or as a commodity.

296
Within professional organizations (firms, schools, hospitals), we are seeing a move away from 297 tailored, unique solutions for each client or patient towards the standardization of service. 298 Increasingly, doctors are using checklists, lawyers rely on precedents, and consultants work with 299 methodologies. More recently, there has been a shift to systematization, the use of technology to 300 automate and sometimes transform the way that professional work is done -from workflow systems 301 through to AI-based problem-solving. More fundamentally, once professional knowledge and 302 expertise is systematized, it will then be made available online, often as a chargeable service… when 303 professional work is broken down into component parts, many of the tasks involved turn out to be 304 routine and process-based. (Susskind & Susskind, 2016a) 305 306 In the case of education, a detailed standardized curricula is a prerequisite for 307 intelligent tutoring systems to actually challenge the professional role of teachers rather than 308 simply be instructional tools used for limited purposes. So far most intelligent tutoring 309 systems have been developed in STEM fields which have more standardized curricula than 310 the social sciences and humanities. There have also been more tools developed for K-12 311 education than for more complex higher education. While there are pressures towards 312 curricular standardization in higher education, such as from learning outcomes assessment 313 and discipline-specific accreditation, the process is more advanced at the primary and 314 secondary level. 315 316 National curricula for K-12 education are common in many countries, albeit not in the 317 United States where K-12 education is under state and local control. Over the last two 318 decades, however, there have also been moves in the U.S. towards national educational 319 standards starting with the Bush-era No Child Left Behind policies which financially 320 punished schools found "underperforming" on standardized testing. Since the racial and 321 class background of students is the principal driver of their aggregate achievement, this 322 policy was widely perceived as a way to defund already struggling public schools and 323 encourage private education. Teachers and schools complained that No Child Left Behind 324 encouraged "teaching to the test" and cutting programs like art and music to spend more 325 time on exam preparation. 326 327 During the Obama era, the fight over K-12 standardization shifted to the Common Core 328 State Standards Initiative. The proposed Common Core was a comprehensive elaboration 329 of competencies in each field that should be achieved at each grade level. Originally 330 proposed by conservatives, and embraced as a bipartisan initiative, many states signed on. 331 As the Tea Party mobilized against all Obama policies, the Common Core was demonized 332 as a federal takeover of local education, and eventually shelved. China, by contrast, with the world's largest education system, adopted national 335 curricular standards for K-12 education in 2003, and in 2018 announced updated standards 336 for high schools. The Chinese enthusiasm for national curricular standards and the use of 337 standardized testing as a meritocratic gatekeeper to class mobility is often attributed to the 338 reliance on national civil service exams since the 6 th century Sui dynasty. Today Chinese 339 education is singularly focused on the university entrance exam or GaoKao, implemented 340 in 1952. While 90% of Chinese high school students scored well enough on the 2020 GaoKao 341 to attend some form of higher education, scores strictly determine the prestige of the 342 university that a student can attend. China's long history with curricular standardization 343 and testing-based meritocracy was a strong influence on education in Vietnam, South Korea, 344 Taiwan, and Japan, all of whom are in the top ten in international comparisons of learning 345 (Jones and Whiting, 2020 Countries like China that have adopted national curricula probably will have a head 348 start in the widespread adoption of interactive instructional tools compared to countries that 349 allow localities to determine curricula. Investments in an intelligent tutoring system, and 350 programming the curricular goals to be achieved, will be a much more attractive investment 351 with these larger markets. On the other hand, the local autonomy and diversity of curricula 352 in places like the United States could also encourage more curricular experimentation, and 353 the creation of intelligent tutoring systems for basic math and reading comprehension could 354 be widely applicable without uniform curricula. We will have to see which kinds of systems 355 implement more widespread use of intelligent tutoring. The key to rationalizing a work process is measuring the time and cost of the labor 361 inputs, and the quantity and quality of the outputs, to determine the most efficient methods 362 for maximizing production at an adequate level of quality. As noted above, curricular 363 standardization is more advanced in K-12 instruction than in higher education, and K-12 364 teachers have less autonomy over what they teach than higher education instructors. But 365 the pressures in higher education to measure and redesign teaching have been increasing 366 for decades, from accreditors, Board of Trustees, politicians, parents, and students. 367 Universities are obliged to demonstrate that they are measuring both general educational 368 skills, literacy, and numeracy, as well as the specific skills and knowledge promised within 369 college majors. They are also being asked to demonstrate that they use these ongoing 370 assessments to map and redesign their curriculum to improve learning outcomes. 371 372 Using learning assessments to guide curricular re-design is one aspect of the 373 rationalization of teaching. Or as Ovetz (Ovetz, 2015) puts it "faculty autonomy over course 374 design, content, delivery, and student assessment have been challenged, and even 375 displaced, by the efforts to replace content-based assessment of learning, represented by the 376 grade and degree, with competency-based standards, rubrics, departmental and student 377 learning objectives, badges, micro-credentials, pathways, and certifications." Learning 378 assessment pushes education from relying on faculty grading and the granting of degrees 379 to external assessment and to the "the differentiation of instructional duties that were once 380 typically performed by a single faculty member into distinct activities performed by various 381 professionals, such as course design, curriculum development, delivery of instruction, and 382 assessment of student learning" (Gehrke & Kezar, 2015). Intelligent tutoring systems can 383 then be based on these validated curricular maps and pedagogies. Just as learning outcomes assessment can be used to rationalize and redesign curricula, 388 a focus on ensuring student competencies encourages the redesign of secondary education 389 and university degrees. Competency-based education has competed with the credit-hour-390 based degree for a long time, with the degree always being more prestigious. Now the tide 391 seems to be shifting back towards competency, as self-guided curricula and assessment 392 become more sophisticated, and the inflexibility and expense of degrees becomes less 393 attractive. At the secondary level, most US states have adopted policies facilitating the 394 completion of high school degrees through testing, such as by reducing or waiving 395 requirements for class time (Brodersen et al., 2017;NASSP, 2021). For instance, the state of 396 Ohio allows students to earn high school credit by demonstrating competency in a subject 397 area rather than requiring a specific number of hours of classroom instruction (Deye, 2018). 398

399
The unbundling of education from a comprehensive transformative experience capped 400 by a degree into a set of specific skills that can be assessed directly completes the process of 401 curricular standardization and learning outcomes assessment. While understaffed 402 secondary education may welcome competency-based pathways for the students who 403 prefer them, the switch from time-based certification to competency-based assessment 404 threatens the business model of higher education. For instance, more than 2 million US high 405 school students take Advanced Placement tests each year. Advanced placement allows 8% 406 of college freshmen to place out of an average of 10 credits, and up to a year's worth, of 407 introductory courses, shortening time-to-degree and costing their schools tens of thousands 408 of dollars per enrollment (Evans, 2018). Extending competency testing to the rest of the 409 curriculum could cut the time and cost of a college degree in half (Craig & Williams, 2015). 410 411 Brief curricula designed to confer badges for specific competencies can also adapt more 412 agilely to labor market expectations. Instead of hiring a graduate with a computer science 413 degree, employers might require competency badges for ten key skills, each requiring an 414 intensive month of instruction. A competency-based model would also lend itself to life-415 long learning, drawing adults back for short, just-in-time training in skills that are 416 immediately applicable to their work rather than to a two-year Master's degree. 417 Badgification and competency-based education in turn simplifies the programming of 418 intelligent tutoring systems to impart these streamlined and standardized skill sets. Online education is also contributing to the emergence of intelligent tutoring by 423 encouraging curricular standardization and reducing instructor autonomy. Distance 424 learning actually began in the 19 th century with experiments in education through 425 correspondence, followed by experiments in recording lectures for broadcast on the radio, 426 television, or on film strips (Mirrlees & Alvi, 2014). By 1958 there were 31 educational 427 television stations and 150 university closed-circuit television experiments in the United 428 States. While administrators and reformers promoted distance learning as a way to educate 429 more students and reduce costs, faculty saw televised education as "the threat of 430 technological unemployment, the degradation of the teacher's status and role, and the 431 dehumanizing of the teacher-pupil relationship" (Zorbaugh, 1958). Critics like David Noble 432 charged that new instructional technologies "like the automation of other industries, rob 433 faculty of their knowledge and skills, their control over their working lives, the product of 434 their labor, and ultimately, the means of their livelihood" (Noble, 1998 Since online education dispenses with the expensive infrastructure of schools and 437 universities it is already significantly cheaper than the traditional model. Online services 438 increase price competition and shift the balance of power from the provider of services to 439 the owners of the curricular product and its consumers. Online education is a potentially 440 global marketplace, far more price-competitive than public schools or local universities. 441 Following this logic, there has been widespread attention to the potential of online 442 education and massive open online courses (MOOCs) as a solution to the rapidly inflating 443 cost of education. Even before Covid a third of American students in higher education were 444 enrolled in distance learning, with about 15% exclusively taking online courses (Lederman,445 2018) that are disproportionately taught by contingent faculty. 446 447 While many American universities have established their own online offerings, 448 MOOCs grew quickly by offering universities partnerships with the technology and 449 marketing expertise of companies like Coursera, edX, and Udacity. The carefully designed 450 MOOC is the joint property of the university and external firm, with the teacher as a 451 contracted content provider. By 2019 five companies accounted for 90% of the MOOC 452 market (Shah, 2019). While a student in rural Oklahoma or Bangladesh might have 453 previously taken a small continuing education course at a local community college, they 454 now had the option of receiving a more prestigious MOOC certification from Harvard or 455 MIT alongside several thousand other students. Businesses are increasingly open to 456 accepting MOOC certification as a credential for hiring and promotion. For instance, the 457 Information Technology Certificate Program offered by Google through Coursera involves 458 courses in system administration, operating systems, and network security, resulting in a 459 Google badge at a fraction of the cost of a comparable community college degree. While 460 MOOCs have not drawn many students away from traditional curricula or degrees, during 461 the Covid crisis of 2020-2021 participation in distance learning was closer to 100% and online 462 education is poised to grow quickly. Perversely K-12 teachers have been simultaneously professionalized and 467 proletarianized. The educational requirements for American K-12 teachers have increased 468 over the last century, and most now have a college degree. But their salaries have remained 469 10% to 25% lower than comparably educated occupations for decades (Strauss, 2016). In 470 2019 dollars the average annual salary for American K-12 teachers reached a high of $63,000 471 per year in 1990 and has stagnated since then (NCES, 2021). In response to declining wages, 472 the number of American students training to be K-12 teachers has also been declining for 473 decades, and those who remain in the field have turned to unprecedented labor militancy 474 (Camera, 2019). With schools restricted from paying higher wages by austerity budgets, or 475 from hiring more teachers by the growing teacher shortage, public education in the United 476 States is primed to embrace intelligent tutoring systems and labor-extending software as 477 proposed by McKinsey. While this essay has mostly focused on the trends in the United 478 States, the shortage of teachers in the developing world is generally even more acute. Once 479 the cost of automated teaching tools declines it may be even more attractive in poor 480 countries without enough teachers (Edwards & Cheok, 2018).

482
As with the proletarianization of K-12 teachers, the decline in professional status and 483 autonomy of university faculty is making it easier to adopt intelligent tutoring systems in 484 higher education. In the United States in 1969 about four out of five (78%) higher education 485 faculty were tenured. By 2018 that had fallen to only one out of four (27%) (Flaherty, 2018). 486 Contingent faculty may have multi-year contracts or may be hired course-by-course, and 487 include post-doctoral fellows, clinical faculty, and visiting professors. What they have in 488 common is lower salaries and considerably less political influence within universities 489 compared to tenured faculty. While higher education administrators and tenured faculty 490 generally agree that hiring tenure-track faculty should be prioritized, financial pressures on 491 universities make adjunctification hard to avoid. For-profit educational institutions in the 492 U.S. rely even more heavily on adjunct faculty, with 90% of instructors being contingent 493 (Proper, 2017). Even adjuncts on full-time, multi-year contracts earn less than comparable 494 tenure-track positions. Reliance on adjuncts is also a boon to administrative flexibility since 495 many adjuncts are hired weeks before courses begin, and their classes can be canceled if 496 they don't fill (Kezar et al., 2019). 497 498 Adjuncts represent the unbundling of teaching from research, governance, and service 499 expectations. While tenure-track faculty condemn the exploitation of adjuncts in 500 aggregate, they also benefit from it since they have been able to shift the teaching of large 501 introductory courses to adjuncts so they can spend more time on research or on teaching 502 smaller classes for advanced students. Since adjuncts do not participate in governance and 503 are doing more of the teaching, they are less able to resist attempts to measure and 504 standardize curricula. The next stage of unbundling is when for-profit schools, like 505 University of Phoenix, hire experts to design curricula and hire adjuncts to teach them. 506 Finally, the university itself can be unbundled into stand-alone online programs, 507 microdegrees, and badges (McCowan, 2017). The proletarianization of teaching and the subordination of teachers to planned 512 curricula and learning assessment do not in themselves threaten the occupation, only its 513 professional autonomy, prestige, and compensation. For capitalist rationalization to fully 514 commodify the occupation it needs to turn the core skills of the profession into software. 515 The technologies to automate these core skills have been in development for decades. 516 One of the first "intelligent tutoring systems" was called SCHOLAR, introduced in 518 1970 to teach South American geography. Tools for teaching STEM fields in general, and 519 computer science in particular, have been over-represented, both because of the expertise of 520 the developers, and the more objective, rule-based nature of STEM disciplines 521 (Mousavinasab et al., 2021). After fifty years of development, these interactive learning tools 522 now incorporate Bayesian logic, data mining, machine learning, and natural language 523 processing. A 2019 meta-analysis of 19 studies of intelligent tutoring systems in K-12 524 education found that students who used them had higher test scores than students in 525 teacher-led classroom instruction, with learning comparable to one-on-one instruction (Xu 526 et al., 2019). These results were the same as those of two previous meta-analyses of more 527 than 50 intelligent tutoring systems in K-12 and higher education (Kulik & Fletcher, 2017;528 Ma et al., 2014): "These results held across grade levels (elementary through higher 529 education), content domains, and study quality (e.g., randomized controlled trials and 530 quasi-experiments)" (Ma et al., 2014). A core teaching skill is the ability to talk to students about the material and to give them 536 context-appropriate advice about what and how to study. Until now intelligent tutoring 537 systems are incapable of understanding and generating conversational speech, and 538 instructional chatbots have been limited interfaces to "frequently asked questions." For 539 instance, universities have experimented with interactive chatbots to guide students 540 through routine inquiries about admissions or course registration (Engati Team, 2021). 541 Admithub and Pounce are chatbots that schools offer to admitted students to steer them 542 towards putting down a deposit. Beacon, a "digital friend" developed by Staffordshire 543 University, recommends readings and makes connections with tutors (Newton, 2021). 544 545 Now intelligent tutoring systems are beginning to use natural language programming 546 (NLP) for open-ended conversational interactions with students (Pérez et al., 2020). NLP 547 models have achieved startling breakthroughs, as with the GPT-3 system introduced by 548 OpenAI in 2020 (Heaven, 2020), ensuring more incorporation into tutoring systems. A team 549 at the University of Bath found a significant increase in learning after incorporating the 550 ability to parse natural language queries and provide hints and links into their intelligent 551 tutoring system Korbit. Their system personalizes the answers and hints based both on 552 previous student queries and their ongoing performance in the course (Kochmar et al., 553 2021). A GPT-3 system from Open AI can summarize books of any length (Wu et al., 2021), 554 and Google, Facebook, and Microsoft have developed their own document summarizing 555 tools (Wiggers, 2021). A GPT-3-based system called Learn From Anyone allows a student to 556 assign any well-known public figure as "teacher" -such as Aristotle for philosophy or 557 Einstein for physics -and the system responds to students' queries in that person's style 558 (Gandhi, 2020). Multiple-choice exams scored by computers have been in use for fifty years. The much 564 more difficult job is the assessment of text, which researchers have been attempting since 565 the 1960s (Page, 1966). Auto-grading of tests and papers is advancing rapidly, and can now 566 not only gauge spelling and grammar, but also the coherence of an argument, its relevance 567 to the prompt, and the complexity of words and syntax. Software can then report on the 568 specific strengths and weaknesses of a student's writing. Essays with unusual features can 569 be flagged for human review (Hussein et al., 2019). 570 571 These systems can be gamed by students who understand what is being scored, by 572 including a lot of big words for instance, and the systems can't yet judge the accuracy of 573 factual claims. But it is almost as much work to write convincing gibberish as it is to write 574 actual prose. As a scientist at the Educational Testing Service, Nitin Madnani, noted "If 575 someone is smart enough to pay attention to all the things that an automated system pays 576 attention to, and to incorporate them in their writing, that's no 577 longer gaming, that's good writing" (Smith, 2018). The main form of cheating in writing assignments however is plagiarism, and 580 automated plagiarism detection tools like TurnItIn have made catching work copied from 581 the Internet painless. Students can also use these new tools to improve their own writing 582 and check for accidental plagiarism. The tool Grammarly for instance gives students 583 suggestions to improve word choice and tone, clarity, formality, and fluency, as well as 584 flagging potential plagiarism. In other words, these tools already provide much of the 585 feedback on writing that a teacher would. The next step is to have an AI help write your 586 paper. The Rytr AI Writing Tool for instance can produce thousands of words of passable 587 prose in 40 different styles given just a few prompts, and more "intelligent authoring" tools 588 are coming to market (Dale & Viethen, 2021). Teachers' largely intuitive sense of students' capacities and struggles is now being 594 complemented by data analytics. Big data and machine learning provide rapid, quantitative 595 assessment, using multiple factors to predict who will be in academic difficulty or require 596 additional attention. Many universities now have data analytics platforms that use dozens 597 of student characteristics to predict whether prospective students will enroll, how successful 598 they will be, what courses they should take, and when they need an intervention to remain 599 enrolled. Online learning management systems (LMSs) like Blackboard provide a moment-600 to-moment picture of student engagement, study habits, and performance that, combined 601 with hundreds of other predictive facts, give instructors and advisors a clear idea of who 602 needs help and with what. For instance, Solutionpath's Student Retention, Engagement, 603 Attainment and Monitoring (StREAM) platform gives teachers and advisors a real-time 604 dashboard of students' 'engagement score' that combines class attendance and the 605 timeliness and quality of assignments and tests with LMS records of interactions with course 606 materials, peers, and the library. 607 608 The next step in applying these Big Data approaches to education are personalized 609 learning recommendations, with a "learning experience platform" (LXP) offering up texts, 610 videos, podcasts, exercises, or course recommendations like Netflix or Youtube videos 611 (Williamson et al., 2020). Intelligent tutoring system informed by this kind of real-time Big 612 Data will be far better at tracking and motivating student learning than any one instructor 613 could be (Kaklij et al., 2019). What kinds of instructional content or exercises should be 614 suggested for an 18-year-old Latina would-be computer scientist who did well in high 615 school but is juggling a second job, versus the 30-year-old Vietnamese adult male with 616 dyslexia who commutes to take night classes to be a nurse? When and for which students 617 does a bad grade on a paper raise a red flag? Are learning problems and successes tied to 618 specific courses, instructors, or pedagogies rather than student characteristics? To the extent 619 that teachers attempt to weigh these kinds of factors, they can be prone to biases that Big 620 Data generally should not be. Social-emotional skills have been considered the most difficult to automate, and these 626 skills are also central to the teacher's role. Ironically, the legendary ELIZA software 627 developed by Joseph Weizenbaum at MIT in the 1960s to simulate Rogerian therapy 628 demonstrated that very simple code, repeating the patient's statements as questions, could 629 elicit trust and deep emotional sharing. Today, rapid progress in natural language 630 processing has been accompanied by improvement in recognizing human emotions from 631 verbal and nonverbal cues. Chatbots that can detect depression and other mental health 632 issues are widely available (Ahmed et al., 2021;Jovanović et al., 2021), and there is 633 accumulating evidence that interacting with mental health chatbots can improve symptoms 634 of depression and anxiety (Abd-Alrazaq et al., 2020). 635 636 Education researchers are now incorporating emotion recognition into intelligent 637 tutoring systems to gauge student engagement, frustration, and mental health 638 (Khadimallah et al., 2020;Newton, 2021). Chinese educators are experimenting with real-639 time monitoring of students' faces in the classroom to generate an engagement dashboard 640 for the teacher (Waltz, 2020). Researchers are developing similar tools to gauge student 641 engagement in the online classroom, at least when students have their cameras on (Sharma 642 et al., 2019). Incorporating emotion recognition into intelligent tutoring systems will allow 643 the software to learn what the student finds engaging or boring, and adjust the content and 644 pace of instruction accordingly. Futurists have a poor track record in predicting how new technologies will impact 649 employment and occupations, and predictions of rapid automation have often been wrong. 650 Powerful occupations can resist the implementation of technologies that threaten their 651 autonomy, or selectively adopt technologies to extend their productivity and authority. 652 Subordinating professions to organizational rationalization is a political project which co-653 evolves with their social and occupational power, the systematization of their skills, the 654 subdivision and outsourcing of their tasks, and the capacity of automation technologies. 655 This essay has situated the development of intelligent tutoring systems within the 656 standardization of curricula, the growing use of learning outcomes assessment and online 657 learning, and the declining autonomy of the teaching profession. These social dynamics will 658 shape how quickly intelligent tutoring systems are created and deployed. 659 660 For more than a century education at all levels has gradually attempted to measure and 661 redesign teaching to maximize productivity and quality. As skills and tasks were re-662 assigned to specialists, they were also subjected to the monitoring and control of school 663 administrators. Standardized curricula and external testing challenged curricular autonomy 664 and grading. In higher education the most privileged professionals, tenured faculty, are 665 being supplanted by contingent instructors, further enabling the standardization and 666 commodification of instruction. The distillation of instruction into knowledge products, 667 such as texts and online courses, complemented efforts to measure learning outcomes and 668 standardize curricula. The principal obstacle to these efforts, the basis of professionals' 669 occupational autonomy, was the irreducibility of their core bundle of skills. For teachers, 670 those skills are the ability to assess students' learning and emotional state and to structure 671 curricula and communicate with students in ways that illuminate and advance their 672 learning. The decades-long gestation of intelligent tutoring systems is now poised to whittle 673 away at these core skills with natural language processing, predictive analytics, adaptive 674 responses, and affective intelligence.

676
The expanded use of intelligent tutoring systems has the potential to address the 677 shortage of teachers in many countries, reducing the cost of education, personalizing 678 learning, and enabling access to life-long learning. Many of the problems with education 679 will remain, however. The algorithmic data used to inform predictive analytics and learning 680 systems will inherit biases from historical educational data in the same way criminal justice 681 or natural language algorithms inherit biases from their training data. There will be better 682 and worse intelligent tutoring systems, and presumably the better ones will be more 683 accessible to the affluent. The content and pedagogical goals of intelligent tutoring systems 684 will remain as political as curricula have always been. Will religious ideas about evolution 685 be included? How will education about sexuality, imperialism, and racism be framed? Will 686 the curricula for the children of the affluent stress creativity, leadership, and civic character 687 development, while systems for the working class focus on marketable skills? Centralizing 688 curricula and pedagogy into intelligent tutoring systems will reduce both good and bad 689 forms of diversity, and give corporations and the state even more influence than they had 690 over millions of human instructors, making these systems a new terrain of political struggle. 691 As with the debate over automation of other fields, it is time to move past debating whether 692 there should be intelligent tutoring systems to how they should be used. 693 Dodd, C.