Read the winning entry in 2023's Wonkathon, which asked contributors to answer this question: “How can we harness the power but mitigate the risks of artificial intelligence in our schools?”
Editor’s note: This essay is an entry in Fordham’s 2023 Wonkathon, which asked contributors to answer this question: “How can we harness the power but mitigate the risks of artificial intelligence in our schools?” Learn more.
Barrels of digital ink are being spilled fretting over the potential for artificial intelligence (AI) tools to disrupt learning in K–12 schools. Educators, policymakers, and parents are worried about the possibility that tools like ChatGPT will help students pass off computer-generated work as their own. While these are valid concerns, they miss the big picture: The greatest potential for AI to disrupt K–12 education lies not in how students learn, but in how educational institutions organize and operate.
AI tools will almost certainly have some positive impact on how students learn. General-use tools like ChatGPT and more K–12 specific aids like Khanmigo will change how students approach their schoolwork by helping them generate ideas and receive feedback faster than anything short of one-on-one support from a teacher or tutor. AI capabilities will enhance platforms connecting students, parents, and teachers. This is all exciting, and leaders should encourage thoughtful and safe use of AI tools in the same way they do for any other technology. But we ought to temper our expectations for these tools to drive transformative change in K–12 classrooms.
That’s because the new class of tools won’t fundamentally change how kids learn. It won’t erase the Matthew Effect—the concept that learning new information is easier when you have a higher level of background knowledge. As Robert Pondiscio noted last winter, “it takes knowledge to communicate knowledge—or even to have the discernment to judge whether an AI-generated piece of text makes sense or sufficiently responds to a prompt.” The real danger of AI in classrooms will be if it distracts from ongoing and increasingly successful efforts to provide students—particularly those in traditionally underserved communities—with access to solid, knowledge-building curricula.
This means leaders and policymakers should prioritize finding opportunities for AI-driven innovations to transform K–12 education at the organizational level. The challenges facing school systems in the wake of the Covid-19 pandemic are complex and interrelated: substantial learning loss, declining enrollment in public schools, a looming fiscal cliff, and staff shortages. AI can help fuel solutions to these challenges within and beyond the traditional public school system in three ways:
Realizing AI’s potential in all these areas will require vision, collaboration, and resilience from leaders and policymakers. Policies and incentives need to account for different ways of operating, even if it means accepting some near-term misses that occur when rolling out new AI-powered initiatives. Accountability and measurement systems need to consider student success that includes but goes beyond standard assessments of core subjects as more student learning shifts beyond the four walls of the schoolhouse. And stakeholders across the K–12 system need to approach AI not as a fad, but as a transformative technology that requires coherent and sustaining systems.
Perhaps most importantly, leaders and educators need to resist the temptation to become overly focused on—or even panicked about—how AI might change teaching and learning. The dawn of ubiquitous AI should serve as a reminder that children still need to develop a deep foundation of knowledge to use these tools well, and that the best use of AI in traditional schools is to free up the time of educators to do more work directly with students. Outside of schools, AI can help cultivate the “weirder” ecosystem of educational options needed for a system of education that empowers families to access the educational opportunities their children need to thrive. When used thoughtfully, AI tools have the potential to move us closer to an education system that provides a more diverse range of experiences to meet the unique needs of every student.
Editor’s note: This essay is an entry in Fordham’s 2023 Wonkathon, which asked contributors to answer this question: “How can we harness the power but mitigate the risks of artificial intelligence in our schools?” Learn more.
Having spent years in both university and school district administration roles, I’ve seen first hand how difficult it is to implement new, innovative ideas that have the potential to serve kids, teachers, and parents with fidelity. Artificial intelligence (AI) use in schools is going to be a complex effort that involves responsibility from various stakeholders, including vendors, school districts, teachers, students, and the broader educational community. It is going to require systems of accountability to be in place that are more agile, transparent, iterative, and responsive as the technology evolves and we learn more about its use cases, enjoy its benefits, and minimize its harms.
Here is my roadmap for how to ensure that, across a school system, anyone responsible for developing, supplying, procuring, implementing, and evaluating AI solutions is accountable for delivering what is best for students.
Start with the vendors
School systems have a lot of power through their procurement processes to ensure certain guardrails, parameters, and expectations are clearly articulated and adhered to by any vendors providing products or services. And while sometimes it may feel like procurement is where all exciting and innovative ideas go to die, a well developed RFP and procurement process that is anchored in a clear and coherent vision for what a district or school is looking for can be the first step in establishing strong expectations.
School systems should establish expectations from the outset by probing vendors to ensure their track record and core values are clear; there are clearly articulated and defined ethical considerations and data privacy policies in place; information about how AI algorithms work and what data is used to train, improve, and support the AI are transparent; and no ambiguity about how student data or information will be used. I’ve heard some call data “the new oil,” and districts have a legal and ethical obligation to protect students’ data and information.
Align use with developmental and pedagogical purposes
Beyond vendor selection, AI tools must continuously demonstrate and prove their efficacy and alignment to evidence-based, age-appropriate, and pedagogically relevant practices. The age of the students, their learning needs, and the specific educational context should all inform the choice and deployment of AI tools. Tools that are out of step with the age or learning objectives of the students will undermine the educational process and do not have a place in schools.
Some might argue that we’ve seen many tools, resources, materials, and technology used in classrooms that have proven little to no efficacy, alignment to evidence-based practice, or improved outcomes, so why are we over scrutinizing AI? The reality is, we should be scrutinizing and holding to account all of those other random things cluttering our schools and classrooms in the same way too.
Craft good policy
The days of no phones, no laptops, “no looking things up” are over. Policies that over-prescribe limits on the use of technology or uptake of new innovations are out of touch. Our school systems and classrooms should support, prepare, and incentivize educators and young people to tap into the technological resources they have at their disposal as they learn to critically think, read, understand and apply mathematics, and collaborate with their peers.
School districts must establish clear policies and guidelines that govern the use of AI within their systems to ensure classrooms are safe and productive learning environments. They must ensure AI policies and use are evaluated regularly and adjusted to reflect the evolving needs of students and advancements in technology. What we know today about AI will be drastically different five years from now, and policies should allow room for change, innovation, iteration, and learning while maintaining alignment to educational objectives and ethical standards.
Develop your people
Supporting teachers to understand and leverage the power of AI, integrate AI into their teaching methods, and adapt their pedagogy to accommodate these digital tools will be critical. AI can provide teachers with meaningful insights and identify areas where students need more support, which is desperately needed as we continue grappling with the effects of the pandemic on student learning.
But as AI tools advance to be able to surface needs and identify targeted supports, school leaders, districts, and policies should ensure that teachers’ professional judgment and triangulated data sources remain critical components in any intervention decisions. Humans must continue to make the final educational decisions that directly shape and impact students’ learning and continue to use professional judgment to ensure students are being served equitably and fairly.
AI tools must serve to augment, rather than replace, teachers, and should be used to enhance, not diminish, the role of teachers. Who doesn’t want to make the life of a teacher a little easier? Can we make teachers’ lives easier and, in the process, support them in personalizing learning to guide students towards mastery of grade-level content, critical thinking, and technological savvy? Sign me up!
Get to know and build skeptical trust in AI recommendations
Educators should also understand how AI systems make decisions and recommendations, particularly when it comes to recommendations about interventions or supports that target resources to specific groups of students. When users can’t grasp the basis of AI recommendations, the trust in these systems diminishes. Some call this the “black box,” where AI systems make decisions that seem arbitrary or inexplicable. AI tools should provide clear explanations of their reasoning, allowing users to have confidence in the recommendations they receive, and accountability policies should put in place checks and balances to ensure that no AI recommendations are implemented without scrutiny.
Involve students and parents
Parents and students should be at the table helping shape decisions for how AI is used at school and at home. They should be active participants in decision-making processes, informed about how AI is being used in the classroom, and encouraged to provide feedback and voice their concerns. Parents and students can become advocates for responsible AI use in schools and contribute to a safer and more effective learning environment if and when given the opportunity to do so. They can also play a critical role in ensuring accountability when things are not progressing positively or change is needed.
Pay attention to bias
At its core, accountability in AI use in schools must include a focus on values, ethics, and fairness. AI systems should be designed to be fair, unbiased, and free from discrimination. They should not perpetuate or exacerbate existing inequalities. Schools and districts must ensure AI tools are trained using representative, diverse, and vetted samples with safeguards against algorithmic bias, stereotyping, and the potential for discrimination in the use of AI. When an AI tool demonstrates bias, accountability policies should be in place to retrain and reprogram the tool to ensure it does not happen again.
Assess impact
No one fully understood how smartphones or social media would transform every aspect of our life in the span of fifteen years. AI is a dynamic field, and its impact on education is beyond what any of us could probably comprehend today. The only way we can keep up is by building strong guardrails and regularly assessing and evaluating the extent to which AI tools are enhancing educational outcomes. We must also constantly anticipate and respond to unintended consequences as they emerge. This should include information from academic assessments, surveys, and feedback from teachers and students. The data collected should be used to refine AI implementation strategies and inform policy decisions.
Let’s get this right.
Editor’s note: This essay is an entry in Fordham’s 2023 Wonkathon, which asked contributors to answer this question: “How can we harness the power but mitigate the risks of artificial intelligence in our schools?” Learn more.
How can we harness the power but mitigate the risks of AI in our schools? To paraphrase mega-wonk James Carville, “it’s about the data!”
Although AI has a sixty year history in education, today’s meaning of AI is heavily influenced by just one year of experiences with ChatGPT. This essay, however, is not about harnessing the power and mitigating the risks of ChatGPT or a concept of generative AI that is conceived narrowly in the image of ChatGPT. That would be shortsighted.
Instead, a wise policy should be grounded in sound, long-lasting foundations. The foundation of today’s AI is data. AI works by finding patterns in data, abstracting these to a model, and then applying the model to produce novel information that reproduces characteristics of the original data. Thus, the core problem in either harnessing the power or mitigating the risks of using today’s generative AI in education is the data.
(ChatGPT was not built using educational data nor information about educational processes. Although it can produce outputs similar to those produced in educational settings, it cannot predict the processes that enable humans to learn. This is because the bots that crawl the web to assemble the data for building today’s large language models don’t have access to learning process data.)
Sufficient data is also not available to today’s researchers, developers, or innovators who are experts in applying AI to education. Indeed, those who are best positioned to build education-specific “foundational models” complain of the difficulty of obtaining large, relevant data sets. Consequently, the path to powerful, low-risk AI in education must start with the problem of safely providing access to enough data to abstract out models specific to educational problems. It has to be data about the processes of teaching and learning.
When talking about big data in education, it is necessary to acknowledge the fallout of the well-publicized failure of InBloom, which was an attempt to centralize ed-tech data. The public will not give carte blanche for the centralization of private student data. But massive centralization is not necessary.
Instead, an AI model could be built using a privacy-protecting data sharing mechanism, such as a secure enclave. In an enclave, the data remains local. An API allows local servers to run a computation and return statistical, aggregate information about the data—importantly, not disclosing the data itself. Although untested, it would be possible for an AI-model-builder thereby to develop a model without ever having access to complete data, but instead only asking for certain statistical properties of local data to be reported back. The first component of my policy proposal is: (1) fund research into building AI models via privacy protecting mechanisms so that local communities could participate in building relevant models without disclosing all their data.
When AI starts by building extremely general models and then attempting to apply them to specific educational situations, risks abound. Thus, a second aspect of my proposal suggests that our efforts towards powerful, safe AI should begin with well-bounded problems. One that seems well suited to today’s AI is determining how to provide optimal supports for learners with disabilities to progress in mathematics problem solving. Although I believe parents are not willing to share their students’ data in general, I can imagine a collective of parents becoming highly motivated to share data if it might help their specific neurodiverse student thrive in mathematics. Further, only limited personal data might be needed to make progress on such a problem. Thus a second element of my proposal is (2) energize nonprofits that work with parents on specific issues to determine how to achieve buy-in to bounded, purpose-specific data sharing. This could involve a planning grant stage, which if successful, would result in money needed to establish a local privacy-protected method of sharing data.
The third part of my proposal would be to fund teams to build new education-specific models around targeted problems and privacy-protected data. There are already sufficient numbers of talented AI researchers who would step up to the challenge of crafting education-specific AI models given an important societal problem and access to enough relevant data. However, there would initially need to be support to (a) negotiate mutually agreeable limitations and benefits with parent and school groups who control the data and (b) determine ethical policies so that the benefits of the resulting AI models are available to those who participated in the research and so that they experience additional favorable returns should the AI models prove useful outside their originating collective. I believe existing federal grants (at multiple agencies) could support this. Thus, the third part would use existing grant funding mechanisms to build powerful and low-risk AI models via researcher-parent partnerships that focus on targeted uses of AI in education.
In summary, I argue against trying to domesticate the wild capabilities of today’s general-purpose generative AI tools. Instead, I argue for an alternative regime based around three commitments:
Harnessing powerful AI while mitigating risks would thus be achieved by privacy-protecting mechanisms for sharing data; building buy-in for limited data sharing; and forming collectives among scientists and members of the public who agree to mutually engage in solving a targeted educational problem.
The science of reading is thankfully supplanting dubious methods of teaching young children to read, as state policymakers across the country are increasingly expanding, refreshing, and/or adopting new early reading legislation to prioritize what works. This recent study, conducted by Michigan State University researchers John Westall and Amy Cummings, asks whether such legislation impacts short-term reading and math achievement, as well as economic and racial test-score gaps. They also investigate whether potential impacts depend (or not) on particular policies or groups of policies to move the needle.
They use three data sets, the first of which is a state-level roster of sixteen early literacy policies, such as requiring materials that reflect the science of reading, training teachers on them, and informing parents when their children are falling behind. They use SEDA data to measure academic outcomes across states, supplemented with national NAEP data to determine whether outcomes differ by low-stakes (NAEP) or high-stakes assessments (state tests). An event study design leverages, across states and over time, the differences in both the staggered adoption of early literacy policies and what they comprise. Analysts compare states’ average test scores several years before and after passage of any early literacy policy, and of the complete suite of sixteen policies, to a comparison group of states that do not (or did not yet) have these early literacy policies in place. They also examine third grade reading retention policies in particular since prior research has shown them to be particularly efficacious for younger students.
Before turning to the findings, a couple caveats merit mention. Since some states have amended their early reading policies piecemeal over time, Westall and Cummings chose the earliest adoption date and interpret their results as “intent-to-treat” effects (so states may or may not have particular policies in place), which is not as precise as it ideally would be. The categorization of states is also somewhat fuzzy. Treated states must have adopted at least one policy before the last year of test score data in the analysis (2018 for SEDA and 2019 for NAEP), so that leaves out at least five states that implemented policies in 2021, while including two in which it’s unclear whether they had policies at all (because the source document was missing evidence).
The key finding is that adopting any early literacy policy improves elementary students’ reading achievement on high-stakes assessments, but those effects fade out by middle school. Adopting a comprehensive suite of policies shows bigger and more sustained increases in reading scores (ranging from 0.03 to 0.1 standard deviations) that persist for several years, especially when third-grade retention requirements are included. The analysts find what they initially thought were spillover effects into math, but ultimately suggest that those gains were more attributable to Race to the Top interventions, which were occurring simultaneously in most states during the study period. There were no significant increases in low-stakes reading scores, except in states with comprehensive policies. They also find some suggestive evidence that early literacy policies reduce socioeconomic and racial gaps on high-stakes tests, but these effects are small.
On the one hand, it is encouraging to see any influence of state policymaking on student outcomes given the multitude of intervening variables from the statehouse to the classroom. The fact that the study picks up some positive (albeit fleeting) signals means that districts (and schools and teachers) are implementing the policies to some degree. On the other hand, we need more information relative to which methods of implementation are most effective to carry out particular policies—and lessons learned from states that have experienced favorable outcomes. These are the keys to long-term success and replication of early reading legislation.
SOURCE: John Westall and Amy Cummings, “The Effects of Early Literacy Policies on Student Achievement,” Education Policy Innovation Collaborative (June 2023).
Recent legislative efforts across the country have strengthened efforts to align reading instruction with the science of reading. These laws typically require teachers to use methods and materials aligned to the solid evidence base on how children best learn to read. To implement these laws, schools are required to hire teachers who are properly trained in those tried-and-true methods, and state licensure exam results can assist prospective employers in telling which candidates have the knowledge and skills needed to do the job right. But a new policy brief from the National Council on Teacher Quality (NCTQ) casts doubt on the evidence provided by many of those exams.
Author Hannah Putman takes a look at all twenty-five of the licensure tests used in various combinations across nearly all states to determine the readiness of teachers to teach young children to read. They are a mix of reading-specific tests and tests of general educational competence that can include math, science, and other subjects, as well. She compares each test with a detailed rubric that includes the five core components of the science of reading (phonemic awareness, phonics, fluency, vocabulary, and comprehension), plus an exhaustive list of related topic areas under each of those components, and ranks each test in one of four possible categories. (The brief offers a full list and further info on her methodology.)
An acceptable test assesses at least half of all topics identified within each of the five components, does not combine reading with other subjects (e.g., math or science), and includes few practices contrary to the research. The latter is an important point, as incorrect practice can negate correct practice even where the latter is strongly encouraged and supported.
A strong test—the highest ranking here—assesses an even higher percentage (75 or more) of the topics identified within each component, meets all of the other criteria for an acceptable test, and also addresses how to support struggling readers and English learners.
A weak test either addresses fewer than half of the topics in one or more components or combines reading with one or more other subjects or includes four or more practices contrary to the research (without clearly identifying that these are undesirable teaching practices).
An unacceptable test covers none of the topics in one or more components, or does not cover all five components adequately and also includes four or more contrary practices.
The bad news: While no test sunk to the depths of unacceptable, a whopping fifteen of the twenty-five qualified as weak. The most common negative factor was failure to adequately address all five components of the science of reading, and the three weakest tests addressed an average of fewer than 40 percent of topics in each component. The next-most common negative was combining reading with other subjects, such that strength in another subject could result in a passing score that masks a limited understanding of reading instruction specifically. On the upside, of the ten tests that met the criteria of acceptable, six exceeded the criteria and were rated as strong measures of aspiring teachers’ knowledge of reading instruction.
Unfortunately, weak tests are widely used across the country. Only twenty states use a strong or acceptable reading licensure test, the most common of those being the Foundations of Reading test, used by eleven states, including Fordham’s home state of Ohio. Twenty-eight states use weak tests, two use a combination of both acceptable and weak tests, and one state (Iowa) has no elementary teacher reading licensure test at all. Twelve states allow teaching candidates to choose among several different tests, although in most cases all options are weak, so the choice doesn’t matter much. Author Putman also cites the specific case of Oklahoma, which had an acceptable reading test in use, but in 2022 began allowing candidates to opt for a weaker test. (This was explained as an effort to combat a shortage of teachers in the Sooner State.)
Putman’s recommendations are straightforward. Testing companies should either shore up or discontinue their weak tests; she notes that both ETS and Pearson currently produce both low- and high-quality tests, so they should be readily able to refocus on the good stuff to the exclusion of the bad without much disruption to their business. State leaders should switch out their weak tests for stronger ones—no matter what political or administrative effort that requires—reduce the number of test options available for aspiring teachers, and close any testing loopholes that may exist. (For instance, some states do not require reading tests for special education or early childhood teachers.)
NCTQ is leading the charge to make sure that every avenue is covered to ensure that the vital components of the science of reading are taught, tested, and delivered to students, and also to ensure that all harmful teaching practices are eliminated in every corner of the education system, no matter how obscure.
SOURCE: Hannah Putman, “False Assurances: Many states’ licensure tests don’t signal whether elementary teachers understand reading instruction,” National Council on Teacher Quality (November 2023).
On this week’s Education Gadfly Show podcast, Tim Daly, the CEO of Ed Navigator, joins Mike to discuss the causes and harms of grade inflation—and how to fix it. Then, on the Research Minute, Amber reports on a new study that investigates whether exposure to certain peers and teachers explains achievement gains resulting from enrollment in early algebra.
Recommended content:
Feedback Welcome: Have ideas for improving our podcast? Send them to Daniel Buck at [email protected].
Cheers
Jeers
