As teachers embrace and incorporate AI tools, the technology poses a number of risks. But there is also hope that forward-thinking teachers may uncover replicable, scalable strategies for deploying it to make the job more manageable and free up time to build the relationships with students that matter so much for learning.
During my last year as a classroom teacher, I was tasked with grading 5,400 assignments. Here’s the math: 150 students x 4 quarters x 9 assignments (the number we were required to grade quarterly). Yes, some of those 5,400 were short quizzes or in-class presentations I could assess in real time. But because these were Honors and IB English language arts (ELA) classes, many other assignments were paragraphs or full-length essays. If I spent just ten minutes per essay per student, that added up to 25 hours of grading for a single assignment. By contrast, I had 7.5 hours of in-school prep time each week, and much of that time was absorbed by meetings, lesson planning, and family communication.
I loved teaching writing, but all those hours grading late at night and on the weekends snowballed into one of the largest forces that drove me out of teaching. “Set more boundaries!” I kept hearing. But nothing seemed skippable. Less planning time would’ve meant struggling through worse lessons, ignoring email would’ve meant missing updates from families and colleagues, and so on. I never figured out a way to streamline without losing something essential.
While my situation was perhaps extreme, it was hardly unique. Three-quarters of teachers who plan to leave identify “an unmanageable workload” as a key factor. Only 19 percent of teachers report working 40 or fewer hours per week during the school year (compared to 60 percent of otherwise similar working adults). And almost a third work 51 hours or more (compared to 9 percent of adult peers). Although it’s true that grading isn’t the only professional responsibility that teachers have outside regular hours, it’s high on the list for many.
With the advent of generative AI just months after I left teaching, the challenges of grading student writing ballooned, as students began submitting ChatGPT-authored work while schools scrambled to respond to the new technology. I was not sorry to have missed that shock to the system. But as I learned more, I became curious about how AI might ease teachers’ burden in other ways, especially by assisting with the time-consuming task of giving feedback on writing. Initially, I feared that AI would simply reward the most formulaic student writing or allow bias to pervade its judgments. Though I haven’t shed these concerns, I’m increasingly open to the possibility that, as long as there are humans who care about writing, we can teach AI to attend to the nuances of writing.
Earlier this year, I spoke with a high school ELA teacher who’s doing exactly that. She explained how she “put her brain” into her school’s AI program by feeding it assignment prompts, grading rubrics, past student work, and her own past feedback on student work. As it churned out scores and comments to students, the teacher gave the AI feedback to fine-tune it in accordance with her voice and instructional priorities. Even after months of this work, she still personally reviewed all student submissions and AI-generated feedback. And, she assured me, she was fully transparent about the process with her students.
This teacher isn’t alone. Though many educators remain wary of using AI for writing feedback, the use of platforms like MagicSchool AI for this purpose is becoming more widespread.
Yet this method of grading doesn’t seem portable across contexts. For starters, only an experienced teacher could provide such sophisticated bot-training. And even though this teacher thoroughly reviews all student submissions and AI feedback, I’m skeptical that the same would be true for many others. I imagine that it would feel more efficient to treat the AI much like a human teaching assistant, by reading the feedback closely during the first few attempts and thereafter trusting the assistant to stay on course, maybe checking a couple random samples periodically. And if a teacher isn’t reading closely, they’ll save more time, but they’ll also risk losing opportunities to learn the nuances of students’ strengths and needs.
Or, from an instructional perspective, is it enough for AI to synthesize student writing data and look for trends? Maybe it is. I honestly don’t know, and given the dearth of evidence to date, I don’t think anyone else does, either. There’s enough potential for it to be worth finding out, but with how many students, with which students, and for how long? It would be awfully nice to somehow track the use of AI-assisted writing feedback, and to formalize professional learning opportunities across schools to share both best practices and setbacks, so that teachers aren’t stuck reinventing the AI wheel via trial and error in isolated contexts.
There’s also the human element. To preserve trust, transparency with students about AI use is the bare minimum. Jessica Grose recently wrote for the New York Times about a group of students who felt “betrayed” by undisclosed teacher AI use. Teachers rightly don’t want students passing off AI writing as their own; it’s entirely reasonable for students to expect the same of teachers.
Even with complete transparency, though, I still wonder how students feel about being graded by AI. I’m doubtful that teens will buy that a teacher is really reading their work if some or all of the feedback is automated. And wouldn’t most of us, at any age, write differently for a robot audience than we would for a human audience? But maybe it’s not so terrible to have young people learning to write not just for fellow humans, but also for a robot audience. Dystopian though it may sound, that’s probably a workforce-ready skill now.
Most fundamentally, how might AI grading disrupt—or enhance—human connection? We know that student-teacher relationships are essential for teaching and learning. The act of providing feedback on student writing is inherently one that shows care: A teacher is taking the time to read, engage with, and respond to a student’s thoughts. Already something is lost when feedback is delivered through an online learning management system rather than by hand, just like a handwritten note still feels more personal than an email (source: I transitioned from accepting only hard-copy papers to requiring Canvas submissions from my undergrads circa 2014). By further depersonalizing the feedback process, relying on AI-generated feedback seems to sacrifice opportunities for authentic, one-on-one exchanges of ideas.
Yet, thinking back to that digital pile of 5,400 assignments, maybe AI grading can also support the most human elements of teaching. For example, it could give teachers more minutes in the day for relationship-building work, such as more carefully tailored lessons or family communications. Even if teachers don’t repurpose those minutes toward specific tasks, the shift to a lighter workload could help with relationships. As a teacher, I was slow in getting longer assignments turned around. If AI had helped me pick up the pace, then students might have been less annoyed with me, and they might have perceived more respect for their work. I certainly would have been less overwhelmed and better rested, which would have better enabled me to manage behavior and build rapport.
In a moment when many young people continue to feel lonely and disengaged from school, AI-assisted writing feedback has the potential to weaken the already-fragile threads holding together classroom communities and to create opportunity for strengthening those bonds. As with any instructional tool, an individual teacher’s skill in wielding it will factor into whether the benefits outweigh the risks. It also doesn’t have to be all or nothing: Teachers might rely on AI-assisted feedback for some writing tasks but not others, or they could provide both so that the student can compare the human and bot responses to their work. Personally, I remain wary of too much AI reliance—admittedly, a decade ago I was one of the last of my colleagues to switch from paper to digital submissions. But in my optimistic moments, I hope that forward-thinking teachers will uncover scalable strategies for deploying AI to make the crucial work of writing instruction more manageable and, above all, to build the relationships that matter so much for learning and for our humanity.
At this year’s SXSW EDU conference, the startup Apprentos stole the show, winning the “Launch Competition” with a slick pitch promising to “de-risk” apprenticeships for employers. In a celebratory Instagram reel and press release that followed, it looked like a breakthrough: a tech solution poised to hack, er, solve one of the thorniest problems in American education. The reality is more complicated. Apprentos—and startups like it—can help individual firms and students, no question. But compared with the scale of apprenticeship programs in countries like Germany or Switzerland, this is merely a drop in the bucket, a reminder that system-level change in U.S. apprenticeships is a generational project—one that startups alone are ill-equipped to drive.
Apprentos’s win is best understood in context. SXSW EDU’s signature event is hardly a new experiment. This was its fifteenth year spotlighting early-stage education ventures. Surveying past finalists offers a telling glimpse into the landscape. Some names are familiar (e.g., Amira Learning), others obscure even to seasoned observers (Tilli, anyone?), and a surprising number of organizations now appear to have vanished entirely, their broken URLs an eerie reminder of how quickly startups come and go. The competition’s framing is equally revealing. Prizes go to “startups on the bleeding edge of innovation,” not to initiatives that have proven durable in schools and classrooms. That’s not a critique—it’s a design choice—but the showcase only hinted at what it might take to shift our country away from a college-or-bust mentality that, while softening, remains deeply entrenched.
According to the U.S. Department of Labor, there are roughly 700,000 Americans in apprenticeships. It sounds impressive, but this represents less than half a percent of the U.S. workforce (the same figure is around 70 percent in Switzerland). By comparison, there are more than 19 million Americans in college. While there’s been no shortage of effort to close this gap—with recent ones led by Presidents Obama, Biden, and Trump—adoption remains stubbornly low. Despite high student satisfaction, solid wages, and strong employer and family support, the reality hasn’t caught up to the rhetoric.
That’s because the disconnect is largely structural and institutional. Much like efforts to persuade Americans to eat more caterpillars and crickets—the coverage of the next protein revolution is legion, yet startups have folded despite clear nutritional and environmental benefits—American families, schools, and employers run up against barriers that make apprenticeships hard to implement at scale. Vocational pathways remain widely viewed as second tier, early tracking is politically fraught, and U.S. employers worry that apprentices will walk out the door, along with firms’ investment in their human capital, at the end of their training.
Startups like Apprentos can help lower costs, reduce administrative friction, and make employer participation less risky. Those fixes are useful, but they address only the margins of a far bigger problem. Scaling apprenticeships requires system-level alignment: schools granting credit, postsecondary institutions recognizing credentials, employers investing in training, and a cultural shift that treats non-college pathways as first-class options. Even when students, parents, and employers are supportive in principle, rigid course credit requirements and packed schedules make it challenging for high schools to accommodate apprenticeships without displacing other essential coursework. Innovation matters, but without broader institutional buy-in, the most brilliant tools will remain novelties, just as the most compelling arguments for edible insects rarely translate into dinner plates.
Even if all the pieces align, the question of time, attention, and resources looms large. A recent NPR story documented something that should be obvious but is too often overlooked: Alongside employer investment, consensus among policymakers that public funding for apprenticeships—federal and state combined—is nowhere near the scale needed to make a dent. Targets like a million (or four million) apprenticeships sound bold, but the federal outlay—often measured in the low hundreds of millions—doesn’t match ambitions that would require billions. As with K–12 schooling, there are always more ideas than dollars, more policy goals than political bandwidth. So while it’s easy to say we should just “take a page from Switzerland,” that rhetoric ignores very real constraints and tradeoffs.
Switzerland’s apprenticeship system now generates considerable hype. Study tours abound, with visitors convinced they’ve seen the future, but its strength lies in inputs—deep cultural, societal, and labor market norms—that cannot easily be transplanted here. Poaching concerns are just one factor; the system’s durability rests on long-established expectations, close coordination among schools and firms, and a cultural acceptance of vocational pathways that the U.S. has yet to develop. The notion that we should (or could) copy it wholesale is wishful thinking.
Returning to the drumbeat encouraging Americans to eat insects, proponents often point to cultures where it’s routine, implying Americans just need to get over their mental block. But that block is a feature, not a bug (pun intended) of America’s values and habits. Norms are not arbitrary preferences; they are embedded systems of expectation and belief. What’s more, attempts to upend long-standing norms frequently yield unintended consequences.
Looking ahead, the lesson is not that apprenticeships aren’t worth pursuing, but that lasting change will be neither quick nor easy. Policymakers, employers, and educators must temper enthusiasm with patience, recognizing that cultural norms, institutional inertia, and finite resources limit what can realistically be achieved in the short term. A sustained, coordinated effort over decades may slowly broaden perceptions of vocational pathways, but it will require more than bold targets or splashy slogans. It will demand alignment across schools, employers, families, and communities, as well as the political will to invest for the long haul. The sizzle of innovative pitches is compelling, but the steak of meaningful, system-level transformation depends on time and steady effort—and calls for both ambition and humility.
For teachers, artificial intelligence can be a boon as well as a challenge.
K–12 administrators are also taking advantage of AI, and I’ve been curious about how it’s being applied to teacher evaluation. I just did a survey of my Marshall Memo subscribers around the world and got 1,280 responses. The sample is admittedly unscientific (readers who chose to respond in the middle of a busy week), but the results jibe with what I’ve been seeing in my school visits and in conversations with front-line educators. The headlines:
The survey also asked for written comments and suggestions, and Claude helped me analyze the 522 responses. Some insights:
As a former principal, I can certainly understand administrators using AI to get their evaluation paperwork done more quickly and efficiently, especially if the process is bureaucratic and burdensome. But there’s a real danger of skipping conversations after classroom visits (supervisors will “just send feedback,” worried one survey respondent), with teachers getting impersonal, robotic evaluations containing errors and misunderstandings. As AI is used more often in the years ahead, there’s a strong possibility that this downside will become more pervasive.
But what worries me most is AI transcriptions (audio or video) being converted directly into rubric ratings of teachers. I’ve heard about this in schools and seen it touted on commercial AI websites and believe it’s doubly flawed. Even when done by a competent human, giving rubric scores after a single classroom visit is not a good idea, especially after a short visit. And even the cleverest AI tool is incapable of capturing the complex dynamics of a classroom and producing accurate high-stakes evaluative ratings. Rubrics are best used at the end of the school year, summing up a teacher’s performance by drawing on information from multiple interactions and including an in-person comparison with the teacher’s self-assessment.
I’m also worried that using AI tools will reduce the number of teacher-supervisor conversations after classroom visits. Face-to-face debriefs, when handled well, are a chance to affirm and encourage effective teaching, give the teacher a chance to explain the broader context, engage them in non-defensive reflection, build relationships and trust, and improve classroom practices a little at a time—with significant cumulative impact.
There is some good news, however: Some enterprising educators have developed AI tools with great potential for improving teaching and learning. Three examples:
All three of these tools use AI for what it does best—quickly synthesizing and summarizing lots of information and saving educators time—while supporting the human work of focusing on what’s going well in classrooms and what can be improved. By streamlining the observation and evaluation process, they increase the likelihood that teachers and administrators will make time for face-to-face conversations (ideally in the teacher’s classroom when students aren’t there) and do the essential work of affirming and continuously improving instruction.
An important proviso: To implement the second and third tools, schools need to modify the traditional teacher evaluation process, specifically:
It’s also vital to guarantee that information within the AI platform is confidential.
Given these conditions (which a number of schools have embraced), well-crafted AI tools can rescue the teacher evaluation process from bureaucratic purgatory, build trust and collaboration, strengthen the quality of day-to-day instruction, and improve learning for all students.
Kim Marshall, formerly a Boston teacher and administrator, coaches principals, consults and speaks on school leadership and evaluation, and publishes the weekly Marshall Memo www.marshallmemo.com. He’s the author of Rethinking Teacher Supervision and Evaluation (Jossey-Bass, 3rd edition, 2024).
School leaders, teachers, and state policymakers all work to provide information to parents about the quality of education in their schools and about their own children’s academic performance. How do parents receive and—if needed—act on the information provided? A new working paper creates an experimental methodology to examine how motivated parents are by various types of academic performance information, including a unique effort to reveal how parents respond when test scores and report cards send conflicting signals.
Researchers Derek Rury from Oregon State University and Ariel Kalil from the University of Chicago used the online survey platform Prolific to recruit parents from across the United States. All individuals who signed the consent form and completed the survey were paid a small stipend, but initial questions were designed to weed out anyone who was not currently the parent or guardian of one or more K–12 students. The majority of parent survey completers (n=2,078) were White (68 percent), followed by 18 percent Black, 7 percent Hispanic, and 2 percent Asian. Nearly 70 percent were mothers. The average age of parents was 41, and the average age of their oldest child was 17. Just over half of parents had completed a college degree and almost 42 percent reported household incomes below $60,000 annually.
The full survey focused on a hypothetical fifth grader who was “enrolled in a school similar to the participant’s own” and who had a specific letter grade in class, score on a standardized test (from 1 to 99), and percentile rank (10, 30, 50, 70, and 90) compared to a nationwide sample of fifth graders. Each survey respondent received a randomized set of data for a child and was asked to offer advice on how much time and money the family of this child should invest in their child’s academic progress based on the academic data provided. The child’s name was also randomly altered to indicate a boy or a girl for different survey completers. The dataset ultimately included more than 23,300 individual investment decisions.
Finally, participants were asked a series of questions regarding their feelings toward grades and standardized tests, designed to reveal whether parents viewed tests as biased against certain groups, whether they felt like a bad parent if their child earned poor grades or test scores, whether they believed parental investments pay off more if the child is already a good student, and so on. Researchers also asked parents how much they valued both grades and tests when making such investment decisions for their own child.
On average, participants recommended that parents invest about $43 per week and a little over four hours a day on their child’s academic progress. But Rury and Kalil found that survey respondents recommended stronger compensatory strategies—more investment to support the child—when either test scores or class grades indicated poor student performance. On average, a one-unit decrease in grades (e.g., from B to C) increased recommended investment by 0.14 standard deviations; a one-unit decrease in test score percentile category increased investment by 0.12 standard deviations. They also found clear evidence that parents placed more weight on grades than test scores when making investment recommendations. That is, the signal to invest was stronger for low grades than for low test scores.
And when the signals are mixed? In scenarios where grades were high but test scores were low, respondents did not recommend investing time or money. In scenarios where grades were low but test scores were high, they did recommend investing, consistent with the non-mixed signals responses. In short: The positive signal of high grades appeared to override the danger signal of low test scores. These findings cut across income, parental-education, and race/ethnicity categories, although Hispanic parents exhibit particularly pronounced grade-weighting. The findings also echo those established by Learning Heroes, whose survey and focus group studies have been sounding similar alarms for years.
The mechanisms at work in Rury and Kalil’s results aren’t immediately apparent, with the researchers’ best guess centered on interaction between the common belief (40 percent of respondents) that standardized tests are biased against “certain groups,” the fact that more respondents reported feeling distressed over their own child’s poor grades than over their poor test results, the high number (almost 30 percent) of respondents who said test scores were a reflection of family income, and several other widely-held beliefs captured in the data. They recommend that future research dives deeper into the specifics of these anti-test biases.
In conclusion, Rury and Kalil sound the alarm that their findings, coupled with rampant grade inflation, indicate parents are not receiving accurate information about how their children are really doing in school. From the parent perspective, this could lead to underinvestment in a child’s academic needs. And from the educator sector perspective, students will get less help at home unless grades and test scores start to match up—and tell the truth. From the policy perspective, it all sounds like a recipe for continued academic malaise.
SOURCE: Derek Rury and Ariel Kalil, “Interpreting Performance: Evidence on Signal Weighting in Human Capital Investment,” University of Chicago, Becker Friedman Institute for Economics (February 2026).
Bibb Hubbard, founder and CEO of Learning Heroes, joins the Education Gadfly Show to discuss grade inflation and the disconnect between what parents think report cards are telling them and how students are actually performing. As families rely on grades and teacher feedback to understand student progress, can stronger school-family communication help give parents a more honest picture of how their children are doing?
Then on the Research Minute, Amber Northern looks at new research on Chicago’s merit-based free community college program and what it suggests about college choice, degree attainment, and whether free tuition at two-year colleges steers students away from four-year success.
Recommended content:
Feedback Welcome: Have ideas for improving our show? We would love to hear them. Send them to [email protected]
