Teaching to the Tech

Holding the purveyors of AI EdTech to high standards will spur innovation in the present and set today’s students up for success as tomorrow’s citizens, workers, thinkers, and entrepreneurs.

Executive Summary

Nearly 50 million American children are enrolled in K-12 public schools. Few areas of federal policy will have more impact on the strength and flourishing of our nation than federal education policy. Yet over the past generation, we have largely allowed for-profit tech companies, not our elected representatives, to set education policy, with screens becoming the conduit for a great deal of instruction and assessment. The results have been grim. Reading and math scores have declined to the lowest levels in decades, while student mental health declines and soft skills atrophy.

Despite this, even as many schools are finally grappling with the fallout from recent uses of EdTech, we are at grave risk of repeating these mistakes on a larger scale with AI in the classroom. American policymakers must not repeat our past mistakes.

As the White House moves to “advance artificial intelligence education for America’s youth,” it risks confusing the legitimate educational goal of fostering AI literacy with the more questionable aspiration to improve educational outcomes across the board by embedding AI in the classroom. There is little evidence that AI can advance such educational outcomes. Available evidence suggests the opposite: that student reliance on AI is accelerating the decline of basic cognitive and social skills. Current AI tools and systems are prone to expose children to harmful and age-inappropriate content, and even to engage in criminal activity, such as the production of deepfake pornography.

It is critical that our schools invest in AI education. But teaching students about AI does not mean using AI to teach in every case. Using AI in the classroom effectively will require careful forethought, creative pedagogy, and conscientious design. Our dynamic technology industry is capable of designing and deploying age-appropriate AI tools that can supplement the hard work of teachers and students, but it will not be incentivized to do so if outsize profits can be reaped through the sale of products that maximize engagement at the cost of learning outcomes.

The federal government is a major source of funding for public education. It must use its purchasing power to demand the best returns for taxpayer money and the best outcomes for America’s children. The administration should establish a federal AI EdTech certification program, requiring any companies wishing to offer their AI products and services to school districts using federal funds to meet robust guidelines of safety, transparency, and pedagogy. Congress should codify these requirements, enshrining them into law. This paper outlines twelve principles that should guide the administration and Congress as they develop these standards.

Far from stifling innovation or threatening America’s global leadership, holding the purveyors of AI EdTech to high standards will spur innovation in the present and set today’s students up for success as tomorrow’s citizens, workers, thinkers, and entrepreneurs.

I. Introduction

Few tasks are more essential to the survival and flourishing of the American republic than the education of our children. If our citizenry is not steadily replenished with rising cohorts of thoughtful, literate young adults capable of responding to the challenges of the day, engaging in public debate, and making their voices heard not only through the ballot box but the countless channels of civil society, we cannot sustain a free and democratic republic. Likewise, without new generations of graduates who are versed both in key technical skills and the humane arts of reading, writing, and speaking well, of hashing out disagreements and building consensus, and above all equipped with the competence and confidence to learn and adapt on the job, we cannot sustain a thriving and growing economy. And if our leadership class is not steadily replenished with entrepreneurs, statesmen, and thinkers who have cultivated self-discipline, long-term thinking, and creative problem-solving, we cannot and will not sustain America’s national leadership in industry, security, and innovation.

Sadly, these are a task at which America has been struggling and in large part failing. Seeking to reconcile the demands of universal education with needs of a complex and differentiated economy and the challenges of persistent racial and class inequalities, one presidential administration after another has trumpeted their determination to “leave no child behind,” but without delivering convincing results. Today, educational outcomes are on a discernible decline by almost any measure, and perhaps most troublingly, the achievement gap is growing: already-struggling students are being left further and further behind.

Within this context, it is no surprise that the White House should seek to reverse this decline and reassert America’s global leadership by technologically supercharging our educational system through the introduction of artificial intelligence (AI). In an April executive order, “Advancing Artificial Intelligence Education for America’s Youth,” the administration declared, “To ensure the United States remains a global leader in this technological revolution, we must provide our Nation’s youth with opportunities to cultivate the skills and understanding necessary to use and create the next generation of AI technology.”¹

Indeed, although the stated goal of the EO is to equip students and teachers with “AI literacy,” which suggests that AI is meant to be a new subject alongside chemistry and literature, other actions indicate that the administration envisions AI as a tool of education across all subject areas. Several avant-garde school districts have already undertaken such comprehensive “AI-ification” of their curricula, and in a recent Department of Education grantmaking priority published in response to the executive order, the agency declared its intention “to improve education outcomes through Artificial Intelligence.” In the Supplemental Priority and Definitions, the Department elaborates:

Beyond serving as a subject of study, AI also offers powerful opportunities to enhance teaching and learning. When used effectively, AI tools have the potential to support personalized instruction, increase classroom engagement, and improve student outcomes. Educators are beginning to use AI-powered platforms to analyze student progress, identify learning gaps, and tailor support to individual needs. This work is only the beginning.²

This goal is much broader than the laudable aspiration to equip students and teachers for “AI literacy” in preparation for an increasingly AI-driven world. Unfortunately, this is exactly the kind of mission creep that we have seen over the past generation with computer and digital technology generally. In the 1980s and 1990s, schools rightly responded to the growing digitization of the economy and society by introducing computer labs as a standard feature of every school, and building space into the curriculum to instruct students in the use of these powerful new devices. Students learned to create art with Microsoft Paint and InDesign, how to write and edit in Word, how to organize data in Excel, how to prepare presentations in PowerPoint, and much more. Given how many of these tasks and skills could be readily applied to academic work in existing subjects, there was soon an irresistible temptation to expand computers from another topic of education, to a tool of education across the board: increasingly books would be replaced with e-books, handwritten essays and tests with typed submissions, step-by-step equation solving with automatically calculated spreadsheet cells.

The computer’s transition from topic to tool took place not in response to research on how to improve learning outcomes, but often in the face of it. For instance, studies have consistently shown that screen-based reading and note-taking results in lower comprehension than the use of books and notepads. Schools simply followed the path of least resistance—and the money. Swarms of “EdTech” salesmen plied their products to administrators, and EdTech companies often dumped their products on school districts as charitable donations, ostensibly as a show of their dedication to America’s youth, but more often as a way to find stable markets for otherwise unsuccessful products, such as the iPad and Chromebook.

This bonanza has been fueled in part by educators’ extremely broad construals of (or outright noncompliance with) relevant child and student privacy protection statutes. As Meg Leta Jones has documented,

The Children’s Online Privacy Protection Act (COPPA) requires companies to obtain parental consent before collecting data from children under 13, and the Family Educational Rights and Privacy Act (FERPA) protects student records from commercial exploitation. Yet through administrative excess and enforcement abdication, federal agencies have systematically dismantled these legislative protections, transforming schools from educational institutions serving families into data collection intermediaries doing the bidding of Big Tech.³

FTC Guidance allowing schools to consent on behalf of parents to data collection, together with schools’ decision to treat EdTech companies as “school officials” authorized to access student records under FERPA, have encouraged companies to use all this new classroom time-on-device to extract large troves of student behavioral data—a gold mine in the currency of the digital economy.⁴

This evolution from “Tech Ed” to “EdTech” may have paid rich dividends for Google and Zoom’s investors, but increasingly at the cost of America’s children, which is to say at the cost of America’s future. We are in grave danger of repeating the same mistake today with “AI Ed,” only at much greater speed. As its first practical action in response to the executive order, the White House Office of Science and Technology Policy announced “A Pledge to America’s Youth: Investing in AI Education,” an attempt to establish public-private partnerships “to provide resources that foster early interest in AI technology, promote AI literacy and proficiency, and enable comprehensive AI training for educators.”⁵. The signers of the pledge (139 listed at time of writing) include a who’s who of tech companies, from Amazon, Anthropic, and Apple to Zoom. They even represent tech VCs, like Y Combinator, and lobbyists, like NetChoice and TechNet, not to mention such esteemed educators of America’s youth as Meta and Roblox. All these companies and organizations are lining up at this publicly funded trade show for the opportunity to share their wares with school districts. With parents desperate for their children to succeed in the new AI economy, teachers’ unions looking for ways to make their members’ jobs easier, and administrators looking to make budgets go further, these companies will have little difficulty finding willing buyers for their products.

But will the results redound to the benefit of American students and American society, producing a cohort of thoughtful citizens, skilled and competent workers, and confident leaders in the rising generation? That remains to be seen, but experience suggests it is unlikely to happen without sustained effort, careful research, thoughtful product design, and judicious planning. We will not succeed in equipping our students to “thrive in an increasingly digital society”⁶ without taking the time to first learn the lessons of the recent past.

This paper considers the past, present, and future of educational technology in order to map out a blueprint for how this administration can effectively harness AI to, as the EO states, “foster[…] a culture of innovation and critical thinking that will solidify our Nation’s leadership in the AI-driven future.”⁷ It first reviews the lessons learned from the often careless deployment of EdTech in American classrooms over the past 20 years, both in terms of learning outcomes and in terms of collateral harms to mental health and child development. Next, it considers the current trajectories of AI-assisted learning and where it is in danger of repeating these past mistakes. Finally, it concludes by presenting a vision for the goals of education in support of American dynamism, and how AI tools and systems, rightly governed and deployed, can play a critical role in equipping the next generation and fostering American dynamism.

II. Learning from the Past: EdTech and our Educational Crisis

Although new technologies have often found their way into our classrooms, a key inflection point appeared in 2010. The iPhone had been out for three years, but only around 2010 were kids and teens beginning to adopt smart phones—with their growing universe of third-party apps, especially social media.⁸ Pretty soon, students were bringing their smartphones to school, encouraged in part by anxious parents eager to keep tabs on their kids at all times. While such devices can hardly be considered “EdTech,” they do represent a profound intrusion of digital technology into classrooms that until then had largely been analog. Students had always passed notes in class, but now they could text, comment on Instagram photos, or play Candy Crush Saga. Most schools caved to these new norms, allowing distracting screens to disrupt the academic setting and undermine classroom discipline.

Around the same time, the iPad was released, which was marketed specifically for classroom use. Apple, then a media darling, persuaded numerous school districts to offer grant programs for iPad purchases, offering major discounts to encourage early adoption. Despite the lack of evidence that these devices improved educational outcomes, schools rushed to jump on the bandwagon.⁹ By 2013, Apple was doing $1 billion in EdTech sales per quarter, with 4.5 million iPads in use in US schools.¹⁰

Other companies soon rushed to join the market, using paid teacher-influencers to prototype products and pitch them to fellow educators.¹¹ Google mimicked Apple’s tactics to encourage Chromebook use in public schools starting in 2012, and by 2015, the Chromebook had surpassed the iPad.¹² By the beginning of 2020, there were 40 million Chromebooks in educational use worldwide¹³ and schools were adopting “1:1” device policies under the assumption that every student must have their own laptop or tablet, or risk being left behind in a digital economy. COVID accelerated this trend, driving a 74% annual spike in Google’s Chromebook sales.¹⁴

At the same time as this vast expenditure was transferring wealth from taxpayers to tech companies, American students saw unprecedented declines in academic outcomes. In 2022, National Assessment of Educational Progress (NAEP) reports of reading and math scores showed substantial declines, reaching the lowest levels in decades.¹⁵ This trend was mirrored across many other OECD countries.¹⁶At the time, this was generally chalked up to the effects of COVID (which is to say, to entirely substituting screen-based learning for human learning) and it was hoped that students would “bounce back” from this traumatic experience. Instead, the 2024 numbers shocked many educators by showing no improvement in math scores and a continued decline in reading scores,¹⁷ with a historic 33% of eighth graders reading at a “below basic” level. Perhaps most troubling was the fact that behind this overall decline was a substantial divergence, as students in the 25^th percentile saw their outcomes plunge while higher-performing students were more modestly affected.¹⁸

In other words, even as educational technology was sold to schools, parents, and the American public as a way of “closing the digital divide,” by which struggling students in poorer households were being left behind, the opposite was in fact happening: the saturation of the classroom with screens seemed to be driving a growing divergence in performance. Indeed, data indicate that higher-income students in intact households were likely to spend less time on screens at home and at school (with elite private schools often trumpeting their screen-free policies), while screentime became the default for students from lower-income families both at home and in the classroom.¹⁹

Of course, the mere fact that test scores have declined as EdTech has expanded does not demonstrate causation. However, numerous studies have found a compelling link. As early as 2015, a comprehensive OECD review concluded, “Students who use computers very frequently at school do a lot worse in most learning outcomes…And perhaps the most disappointing finding of the report is that technology is of little help in bridging the skills divide between advantaged and disadvantaged students.”²⁰ A study of the 2018 NAEP data showed “a highly negative correlation between devices and student learning. For instance, students who use a computer for English language-arts work for less than 30 minutes a day score 23 points higher on a reading exam than students who use a computer for English language-arts work for four or more hours a day.”²¹ And a massive review of 126 studies of EdTech by the global research center J-PAL concluded, “Initiatives that expand access to computers…do not improve K-12 grades and test scores.”²² (It did, however, also note that “Educational software designed to help students develop particular skills at their own rate of progress have shown enormous promise in improving learning outcomes, particularly in math,” an observation relevant to our consideration of AI educational tools below.)

Why might this be the case? Particularly in the domain of reading, it has long been clear that both comprehension and retention tend to be significantly higher with print than with screens. Nicholas Carr argued this case forcefully in his 2010 book The Shallows, and numerous studies since have borne out this observation.²³ Similar trends have been observed for the benefits of paper over screens for writing²⁴ and doing math problems.²⁵ The causes of this superiority of physical over digital learning media may relate in part to our need for what philosopher Matthew Crawford calls “embodied cognition,”²⁶ but more simply, screen-users are more prone to multi-tasking. As Jared Cooney Horvath observes, “multitasking is among the worst things that students can do for learning and memory. When we multitask, we go slower, our accuracy drops, and our learning decreases significantly.” ²⁷ Since teens are habituated to spend more than 80% of their time on screens consuming entertainment or scrolling social media, Victoria Rideout and Michael B. Robb, The Common Sense Census: Media Use by Tweens and Teens, 2019 Common Sense Media, 2019. their brains default to distraction when using a digital device to do schoolwork—lasting on average less than six minutes before being distracted by some other activity on the device.²⁸

Little surprise then that 56% of K-12 educators reported that off-task behavior on school-issued devices was a major source of distraction during learning time.²⁹

Meanwhile, as ordinarily deployed, especially in 1:1 device programs, screens tend to isolate students from one another and the teacher, physically and attentionally, diminishing motivation to learn and causing social skills to atrophy. Students learn much more effectively when they learn together. Collaborative problem-solving engages all of our faculties, not only improving learning outcomes,³⁰ but also helping students develop the “soft skills” of persuasion, negotiation, division of labor, and conflict management that are necessary to thrive in the workforce and other adult contexts.

The social disruption caused by digital technology, as Jean Twenge, Jonathan Haidt, and others have documented, has helped drive a crisis of youth mental health since 2010. While classroom technology may not be the primary vector for these harms, schools’ highly permissive approach to smartphones in classrooms have been inseparable from a larger transformation in the relationship of education to digital technology over the past generation. Moreover, many studies suggest that excessive screen time as such is a key driver of many harms.³¹ With students already spending unhealthy amounts of screen time outside of class, the use of screens for learning has only intensified these harms, especially since many school-issued devices lack basic controls or restrictions to ensure that students can use them only for school-related tasks. Given the data-mining incentives of digital technology, one is tempted to conclude that this is a feature, not a bug, as a 2022 report showed that 96% of “educational” apps shared children’s browsing data with third parties.³²

Most troublingly, school-issued devices have become a primary vector for youth exposure to pornography, a widening crisis that has been receiving increased attention. The average age of first exposure to pornography is around 12 years old, with 30% of all young people reporting that they had watched pornography at school (often on a school-issued device).³³ Given the immense time and effort that many parents expend on ensuring that their children are protected from pornography at home, it is an appalling injustice for their tax dollars to be spent on enabling Google to provide their children easy access to pornography in the classroom.³⁴ And as harmful and addictive as hardcore pornography is to developing adolescent minds, even worse are the numerous stories of children being connected to sexual predators through school-issued technology.³⁵

In light of these myriad collateral harms of screens in schools, together with declining educational outcomes, it is little surprise that dozens of states have begun to aggressively roll back the tech policies of the 2010s, with 18 states now banning cell phones bell-to-bell and many districts beginning to reconsider their other classroom device policies.³⁶ It is strange and unfortunate, then, that in the midst of this long-overdue nationwide reckoning over the damage done by the last generation of digital technology in the classroom, many schools are now rushing pell-mell to fill this void with the newest generation of EdTech: generative AI.

III. Navigating the Present: New Challenges of AI in Education

While it is not inevitable that the deployment of AI in K-12 education will produce similar harms, if allowed to follow the path of least resistance and maximum profit for EdTech companies, it is likely to.

From the standpoint of educational outcomes, many of the downsides of existing EdTech are likely to apply to AI as well. While we may imagine in the not-too-distant future the possibility of physical AI teaching bots, for the time being, “using AI” almost inevitably means “sitting in front of a screen.” Given the harmful results of excessive screen time in the classroom to date, it should be incumbent on those touting AI’s educational benefits to demonstrate how we will avoid intensifying these harms if we begin to substitute even more screen time for face-time with human teachers. Such screen time is poor preparation for an adult workplace in which students will have to be prepared to listen attentively to and persuasively engage with customers, supervisors, and co-workers. Moreover, if AI educational applications are simply loaded onto general-purpose Chromebooks or tablets with easy access to other online distractions, even the most promising AI tools are unlikely to achieve the hoped-for learning outcomes.

The collateral harms of screens in schools that have provoked a parental backlash in recent years also seem likely to be mirrored with AI technology. Just as social media use has driven many teens into spirals of isolation, depression, self-harm, and even suicide, troubling stories are rapidly proliferating of similar, but even more intense, effects from young people engaging with human-like AI chatbots.³⁷

These bots, by hijacking our natural desire for companionship, connection, and affirmation (a desire especially strong in the adolescent years), and offering an endlessly sympathetic and sycophantic listening ear, are causing many young people to detach from reality and develop dangerous psychological dependence. One of the most common uses of AI technology, moreover, has been pornographic, with many chatbots offering sexually explicit roleplay, and numerous widely-available apps promising to create deepfake pornography, often based on pictures of classmates. According to the New York Times, schools across the country are facing an “epidemic of deepfake nudes,”³⁸. an epidemic that the recent passage of the TAKE IT DOWN Act will probably only modestly curtail.³⁹

Advocates of AI-based education may well reply that their tools are not designed for such deplorable purposes. However, without more robust safeguards, they are liable to be put to such uses and worse. Earlier this year, for instance, Forbes documented how KnowUnity’s SchoolGPT, readily provided step-by-step instructions for synthesizing fentanyl, while a competitor marketed by CourseHero offered advice for date rape drugs and suicide.⁴⁰ Moreover, even if educational AI apps are more constrained, educators will have to take extreme care lest they normalize for students dangerous habits with AI that they go on to apply outside of the classroom. For instance, students who spend time in class using AI to generate and manipulate images may be led to experiment outside of class with altering (and nudifying) pictures of their classmates. Students who develop habits of psychological dependence on AI tutors in the classroom may be tempted to log on to companion apps like Character.AI or Replika when they get home. Educators may reply that this is a problem for parents, not for them; however, if it is actually AI literacy and AI competency that we are educating for, this must include taking great care to instruct students in the perils of AI tools and habituating them away from unhealthy uses. Indeed, in the recent highly-publicized OpenAI teen suicide lawsuit, it was reported that the victim began his ultimately tragic interactions with ChatGPT by asking questions about his schoolwork.⁴¹

Not only is educational AI at risk of repeating and compounding many of the mistakes of the existing suite of EdTech, but it also poses certain unique challenges that educators are only beginning to grapple with.

Most notably, AI seems perfect for cheating. Professors, teachers, and administrators at nearly every level of education have reported a massive breakdown in academic integrity over the past three years driven by generative AI. Even students are joining in the lament; as one high-school student wrote recently in The Atlantic, “AI has softened the consequences of procrastination and led many students to avoid doing any work at all…. The dominant worldview seems to be: Why worry about actually learning anything when you can get an A for outsourcing your thinking to a machine?”⁴² Almost any kind of traditional take-home assignment, it seems, can be completed in part or in whole by AI, with the plagiarism almost undetectable. For many of today’s students, the problem seems not to be so much a lack of AI literacy, but rather a lack of any other kind of literacy, facilitated by their mastery of ChatGPT.

While OpenAI has responded to the problem by releasing a new “Study Mode,” ostensibly to encourage students to work through writing and problem-solving step-by-step, early tests suggest that it is still little more than a plagiarism machine.⁴³ Research into the academic impacts of generative AI is still limited, but early studies seem to confirm what we could readily have guessed all along: that use of generative AI, while increasing output in the short run, does so at the cost of memory, understanding, problem-solving, and critical thinking.⁴⁴ We know that learning tends to be correlated with mental effort; thus, exciting new technologies that promise to reduce mental effort, as popular as they may be with students and even with teachers, will tend to undermine learning.

The widespread adoption of AI in education may have other more diffuse but equally significant impacts. As anyone who has ever been a teacher knows, education depends on authority, unfashionable though that word may be. By replacing the authority of a human teacher with the authority of a seemingly-infallible (although frequently hallucinating) AI tutor, we risk reducing our teachers to little more than classroom babysitters, and poorly-respected ones at that.⁴⁵ Needless to say, not only will this have negative academic effects (humility is a prerequisite to sustained learning), but it will also encourage moral malformation and prepare our students very poorly for navigating authority relationships in the adult workforce.

IV. Twelve Principles for a Brighter Future

None of the above is meant to deny that AI tools and systems can help us to reform and reimagine the K-12 classroom in ways that could provide immense advantages for the next generation. For example, AI phonics and language tutors can provide personalized, infinitely patient feedback and conversational practice for every student in a classroom, allowing students to progress at their own pace and freeing teachers to focus their attention where it is most needed. Similarly, a well-designed Socratic writing tutor could provide an immense force multiplier for teachers, since writing feedback and instruction is extremely time-consuming. Imagine a bot which, instead of simply fixing spelling and punctuation errors, told the student it had found 25 mistakes in their submission and prodded the student to identify and correct each one. It could then move on to highlight poor word choice and syntax and prompt the student to think up better ways of expressing their ideas. Finally, it could streamline some of the needlessly tedious aspects of writing, such as formatting citations. AI tools could similarly play a beneficial role in STEM education, if students, working collaboratively in teams, used generative AI to mock up designs to solve engineering problems and then produced the parts on 3-D printers before assembling structures or machines. Teachers could use AI to improve lesson plans and personalize assignments, and administrators could use AI to improve data management and tracking.

But before we get carried away with optimism about the power of technology to transform education, it is worth remembering that not all human problems are readily susceptible to technological solutions. The basic challenge of getting an often stubborn, distractible child to buckle down and learn—indeed, to learn to love to learn—has been with us for millennia, and it is not obvious that we have yet progressed much further than Plato. Technology may assist us in the task, as it always has (from the abacus and clay tablet to the calculator and laptop) but we are foolish if we look to it for transformative results.

Moreover, while innovation can often be entrusted to market forces, the realm of education is one doomed to market failure. Why? Because while it is the task of markets to cater to revealed customer preferences, it is the task of educators to resist and fundamentally modify those preferences. The successful educator does not find out what six-year-olds or sixteen-year-olds already like and simply give it to them; he discerns what it is they ought to like and finds ways to mold their desires accordingly, teaching them the value of deferred gratification and enabling them to develop skills rather than following the path of least resistance. Mass-marketed educational products will tend to respond to market demand by maximizing engagement through customization and gamification. These may generate immense profits, but they will not create successful learners. Moreover, we have long recognized that, without clear standards and liability regimes in place, the market will tend to prioritize profit over safety, a tradeoff that we cannot accept in the case of our children.

Thus, recognizing that education is a public good that cannot be effectively supplied by market forces, we have rightly entrusted it in large part to states, communities, and charitable institutions. It is a fundamental task of government to discern, in conversation and collaboration with parents and communities, what measures and tools are required to effectively educate the next generation within an ever-changing social and economic context, and to set standards, fund programs, provide training, and measure outcomes accordingly.

The following twelve principles should guide how the federal government steers AI educational policy to avoid repeating the mistakes of the past generation of EdTech and to set our children up for success in an increasingly AI-driven future. These principles fall under four headings: Goals, Governance, Pedagogy, and Design.

Goals

Clarity about objectives: No quest can succeed if it does not know where it is headed. To date, the goals of the Trump administration’s guidance on educating for an AI future lack clarity. To begin with, the two distinct objectives—educating for AI literacy and using AI to improve educational outcomes across the board—should be clearly distinguished from one another rather than treated as interchangeable. “AI literacy” may be an important skill for today’s students to develop, but ordinary literacy (in its broadest sense of reading, writing, and critical thinking) and numeracy will remain the bedrock of a successful education. Even in a mature AI-driven economy, policymakers should anticipate that the skills most in demand will tend to be distinctively human and relational skills, rather than mere technical computing competence. This should certainly inform our pedagogy today.
Benchmarks for success: Once these objectives are defined, the administration should determine what would constitute evidence of success, as well as what collateral harms are acceptable as the price of this success. To state the obvious, if graduation rates increase but only because students are using AI to complete their assignments, this will not constitute success. But what if students demonstrate increased technical prowess in the use of AI systems and improvement in certain STEM fields, but at the cost of declining literacy and reasoning skills, how shall we assess this result? If our school districts are to invest billions of dollars into AI education, we had better establish in advance clear benchmarks for assessing whether this has been a good investment of taxpayer funds. The NAEP’s “Nation’s Report Card” remains the gold standard for such assessments, and while the Department of Education should ensure that its “Technology & Engineering Literacy” assessment is regularly updated to include AI literacy, this category should remain one among many benchmarks of educational success.

Governance

Privacy protections: Any use of AI systems in schools should be accompanied by robust privacy protections to ensure that no student data is tracked or retained beyond what is necessary for parents, teachers, and administrators to effectively supervise learning. Schools should ensure that any products rigorously comply with the full legislative intent of COPPA and FERPA, and that wherever possible, these systems operate locally rather than in the cloud. They should beware of companies purveying products at prices that are too good to be true, so that they can collect student data: if you’re not paying for the product, you are the product.
Parental involvement: It is essential that parents be thoroughly informed about the purpose and functions of any AI tools or systems being adopted in their schools, and have the opportunity to give meaningful input before major pedagogical changes are made. Privacy protocols and product safeguards—or the lack thereof—should be disclosed to parents, and parents should, within reason, have the ability to opt their children out of using tools or participating in forms of instruction that they believe will be harmful to their child’s development.
Teacher authority: In order to preserve teacher authority, it is critical that any adoption of AI tools or systems in the classroom be subject to the discretion of teachers, rather than imposed from above. Will AI tools function as teaching aids, as technology often has (from the whiteboard to the laser pointer), or as the real delivery vehicle for instruction? Much hinges on how we answer this question, for at the heart of education is imitation: students, whether they realize it or not, tend to become like their teachers. If we are to educate for a human future, we must ensure that at the head of every class is a human teacher, not a human babysitter for a digital instructor.

Pedagogy

Prioritization of foundational instruction: Inasmuch as our goal is improving AI literacy, we must remember that learning about AI is not the same as learning with AI. While there are practical skills for using AI effectively that require hands-on practice, that must be accompanied by understanding of how large language models work, how they are trained, and what sorts of problems they excel at. The advent of AI calls not for a doubling down on mere technical training, but the revival of both the humanities (equipping students to engage with questions in philosophy of language, philosophy of mind, and ethics), and technology studies (so they understand how AI systems work). Effective AI pedagogy will begin with such foundational instruction, and introduce the use of age-appropriate AI tools gradually, with most hands-on use of AI reserved for upper grades during class periods and contexts specifically dedicated to such instruction (as in the “computer labs” that played such a key role in 1990s and 2000s tech education).⁴⁶
Collaborative learning: We have seen that many of the harms of screen-based learning come from its tendency to isolate students from one another, causing fundamental social skills to atrophy, and rewiring children’s brains for non-human engagement. If we are not to repeat this mistake with AI, we should seek to ensure that most student use of AI takes place in collaborative group settings, in which students are learning to brainstorm, plan projects, and execute them in conversation with one another, with the bot functioning at most as simply one more conversation partner or assistant.
Limitation of screen time: Although for some applications, 1:1 use of AI in which students are sequestered from one another behind individual screens may be necessary (e.g., personalized language learning mentioned above) such use should be carefully limited. AI educational tools are being promoted to schools on the basis that they will greatly improve learning efficiency, cutting down on the class time that is often wasted by teachers struggling to convey the same concept to twenty different learners with different needs and styles. If this is so, then AI-based learning ought to significantly reduce passive instructional time, freeing up large portions of the school day for collaboration, conversation, and teacher-led Socratic discussion. Schools should reinvest these efficiency gains in human development, rather than falling prey to the temptations to use screens as babysitters.

Design

Task-oriented: Given the disastrous track record of past EdTech in encouraging students to “multi-task” or simply click away to more entertaining apps on their devices, it is critical that AI educational software be designed to be rigorously task-oriented. Such programs should be designed along the lines of standardized testing software, such that the user cannot click away to any other application and must follow defined learning pathways until tasks are completed or concepts mastered.
Non-humanoid: The dangers of generative AI’s human-like user interface are already being well-documented, with children proving especially prone to emotional manipulation and psychological dependence of companionable, ingratiating chatbots. Although such design features help to make AI tools more engaging—understandably desirable in a classroom setting—these perils are grave enough that any child-facing AI applications should be designed to minimize human-like conversational feature and deliberately accentuate a more robotic interface, so that students are not tempted to blur the lines between reality and artificial personhood.
Prompting rather than prompted: To date, much of the use of AI tools in education has encouraged cheating and undermined learning because of bots designed to obsequiously respond to user prompts. Any good teacher, however, knows that in a well-functioning classroom, it is the teacher posing most of the questions, and the students learning to generate answers. Accordingly, effective educational AI tools should instead be engineered to prompt users rather than respond to prompts. There is a world of difference between an AI writing tool that produces a thesis statement on demand from a student and one that requires the student to generate a thesis statement and continually prompts him to refine it.
Aligned to child safety: AI tools created for use in the classroom should also be designed with rigorous guardrails that cannot be easily jailbroken and ensure that students can only use them for queries related to the subject under study. They must not respond to inappropriate queries, provide students with access to age-inappropriate information, or engage in extraneous relationship-building conversations with students. Of course, what is “age-appropriate” for a sixteen-year-old is quite different from what is appropriate for an eight-year-old, so products must be designed with calibration to different age levels and aligned with the developmental needs of their users.

To design and deploy AI educational tools according to these principles will require significant forethought, research, carefully-conceived pilot programs and rigorous assessments. But to unleash these tools of unparalleled power on 50 million American schoolchildren without such forethought and standards, restrained only by the sales resistance of local school boards to EdTech salesmen, is an unacceptable risk that will cost us far more over the coming years.

V. Recommendations

Accordingly, the White House’s Artificial Intelligence Education Task Force, while marking an important start, must have its remit broadened and deepened if the above concerns and priorities are to be addressed. In particular, the timetable announced in the initial April 2025 executive order, which requires that AI educational resources “are ready for use in K-12 instruction” by the end of 2025 simply does not allow time for appropriate definition of benchmarks and quality controls. The “Principles for Responsible Use” for AI outlined in the Department of Education’s “Dear Colleague Letter” point in the right direction, but should be built out extensively and given teeth.

The White House should issue a supplementary executive order highlighting the above twelve principles, extending the timetables for the Artificial Education Task Force, and convening a Working Group on Technology in the Classroom to assist it. This Working Group should bring together researchers, frontline educators, parents’ advocates, technologists, and child development experts to build out these principles into robust guidelines that can serve as the basis for a federal EdTech certification program. The Artificial Education Task Force should then formalize these guidelines as a basis upon which to evaluate any public-private partnerships proposed under the “Pledge to America’s Youth” program. Any companies subsequently wishing to benefit from federal grants or contracts to offer their AI products and services to school districts should have to demonstrate how their products or programs meet these safety, transparency, and pedagogy standards.

This supplementary executive order should also direct the Department of Education, the NSF, and the NIH to allocate funding for rigorous longitudinal research into the effects of various AI tools and systems on learning and child development, so that such research can be used to guide the continual refining of these certification standards. The White House should also announce an AI Education Innovation Challenge that will award large grant prizes to entrepreneurs that successfully demonstrate breakthroughs in AI educational tools that can conform to the principles articulated above.

Congress should seek to codify these important measures in legislation that would protect America’s children and ensure that AI EdTech does not follow in the same pernicious grooves as much of recent EdTech has done. This legislation should clarify the original intent of COPPA and FERPA to ensure that data collection in schools is limited to educational rather than commercial purposes, is stored locally wherever possible, and is subject to robust disclosure and parental consent requirements. This legislation should also establish clear product liability standards for any AI EdTech, whistleblower protections for teachers, and processes for parents to opt out of forms of AI-based instruction. This legislation should also assign the Working Group on Technology in the Classroom, or an analogous body, to a permanent role within the Department of Commerce or the Department of Education, where it will advise upon and codify safety, transparency, and pedagogy standards and require companies receiving federal grants for educational software to undergo regular audits to ensure compliance with such standards.

Objectors will complain that such policies will slow innovation and reduce American competitiveness, but this is false. In reality, however, innovation is spurred by the need to overcome obstacles, challenges, or constraints. Often, this has meant responding to demanding design standards set by the federal government, which have pushed companies to heights of achievement they would otherwise have deemed impossible. By taking a leading role in defining the benchmarks which AI systems and applications must meet in order to justify deploying them on America’s youth aided by taxpayer funds, the federal government can spur innovation and ensure that our AI technology continues to lead the world. After all, many of the problems that must be solved by AI labs to create effective classroom-ready technology are the same problems that must be solved for even larger-scale and higher-stakes AI deployment (such as in national security applications): how to align systems with real user needs and interests, how to secure them against data leaks and jailbreaking, and how to subordinate them to effective local institutional governance.

Such an approach to AI in K-12 education will not threaten America’s global leadership or weaken our position vis-à-vis China, which is itself carefully designing its AI educational tools to ensure that they aid rather than undermine learning.⁴⁷ If we take prudent steps now to ensure we guard, nurture, and authentically educate the next generation in this technological age, we will secure our nation’s future. If we do not, we will set the next generation up for illiteracy, weakness and stagnation.

Brad Littlejohn

Brad Littlejohn is Director of Programs and Education at American Compass. He also serves as a Board Member and Coalitions Advisor for the Digital Childhood Institute. You can follow him on X at @WBLittlejohn.

@WBLittlejohn