I Got My Weekend Back

What teacher-facing AI saves, what it can quietly take away, and why not every recovered hour belongs back to the job.Photo by Pavel Bekker on UnsplashI. I Got My Weekend BackThere are many ways to determine whether a new technology works.We can assess how quickly a person completes a task,…

What teacher-facing AI saves, what it can quietly take away, and why not every recovered hour belongs back to the job.Photo by Pavel Bekker on UnsplashI. I Got My Weekend BackThere are many ways to determine whether a new technology works.We can assess how quickly a person completes a task, calculate how much more they produce, and compare the quality of their work before and after adopting it. In schools, we might examine changes in student achievement, teacher retention, or how much districts spend to achieve the same results.While researching how K–12 teachers use generative AI, I encountered another measure. One teacher told me that AI “gave me my weekend back” (Schoon, 2026, p. 77).She was talking about lesson planning. With generative AI, she could complete the initial part of the work earlier, leaving more time to review and improve what she had produced instead of carrying unfinished planning home. Her example was the clearest illustration of a broader pattern in my study.Several teachers described AI reducing the amount of work that spilled beyond the school day. They spoke of less take-home planning, lower stress, faster drafting, more manageable workloads, and, in some cases, reclaimed personal time (Schoon, 2026, pp. 81–83).“I got my weekend back” has stayed with me.Partly, I think, because it is such a modest claim.She did not say that generative AI had revolutionized her teaching. She did not claim it transformed student achievement, made her a better educator, or let her produce twice as much work. She got Saturday back. Maybe Sunday, too.For a technology surrounded by promises of transformation, that can sound almost disappointingly small. I am not sure it is.Teachers in my study generally did not describe generative AI as replacing the kind of thinking that made them teachers. One participant knew what she wanted to do instructionally but found converting those ideas into formal written plans burdensome. AI helped her put onto paper what was already in her head.Others described using it to produce first drafts, alter existing materials, design interventions, differentiate content, translate language, or begin communications they would later revise. The main reduction was often in the work required to move from an idea to an initial artifact. Notably, it was not in the judgment required to decide whether that artifact was any good (Schoon, 2026, pp. 77–78).That distinction became one of the central findings of my applied research project.GenAI could reduce effort, but the work rarely disappeared cleanly. Instead, it shifted. Teachers spent less effort confronting blank pages and sometimes more effort prompting, checking, revising, individualizing, monitoring, or deciding whether something produced by a machine actually belonged in front of their students.Some of the time saved became personal time. Some became better refinement. Some turned into new kinds of work that had not existed before (Schoon, 2026, pp. 77–87).So when a teacher told me that AI had given her weekend back, I did not hear a story about a professional learning to care less.I heard a story about a boundary being restored.That matters because conversations about AI and productivity often contain an unspoken assumption: when technology lets someone accomplish a task in less time, the value of that efficiency lies in what the person can do with the time left over.The saved hour becomes available capacity. Another lesson could be differentiated. Another assessment could be created. More feedback could be written. Tomorrow’s lesson could be improved.Maybe.But there is another possibility.Maybe the hour was never supposed to belong to the job.That was roughly where my thinking stood when I read a recent working paper by Alp Sungu, Benjamin Lira, and Angela Duckworth titled Generative AI Can Harm Teaching (Sungu et al., 2026).Their evidence made the question harder.And much more interesting.Because if AI can give teachers their weekends back while simultaneously making some students’ classroom experiences worse, then neither “AI saves teachers time” nor “AI harms teaching” tells us enough.We have to ask what, exactly, is being saved.II. A Warning Worth Taking SeriouslySungu, Lira, and Duckworth address a question for which we have had relatively little causal evidence.Teachers are already using these tools. Growing evidence suggests that generative AI can accelerate some forms of knowledge work. It is easy to demonstrate that a chatbot can produce a lesson plan, assessment, worksheet, rubric, parent communication, or differentiated text in seconds.What we have known much less about is whether putting those capabilities in teachers’ hands actually improves what students experience.Sungu and colleagues attempted to provide some evidence.Their randomized field experiment took place in schools in Turkey and analyzed students in grades 5 through 12 taught by 193 teachers across 14 schools. Teachers were assigned at the school-department level to either business-as-usual conditions or one of two groups that received access to a customized generative AI teaching-support tool. One treatment group also received weekly reminders intended to encourage continued use (Sungu et al., 2026).How teachers used the tool matters. Most interactions focused on preparing teaching materials rather than providing instructional support. Lecture preparation, homework and exam creation, and syllabus design accounted for much of the recorded use, and the median interaction consisted of only two teacher prompts.The authors interpret this pattern as suggesting that teachers often treated the system as a generator of instructional artifacts rather than as an iterative partner in pedagogical reasoning (Sungu et al., 2026).Then came the results.Giving teachers access to the AI tool reduced students’ intrinsic motivation by about 0.11 standard deviations. Students described the affected courses as less enjoyable, less interesting, and less important. Average academic performance did not change significantly across the full sample.However, among students taught by teachers in the study’s below-median baseline-performance group, AI access was associated with lower academic performance and lower student confidence. The decline in intrinsic motivation was also greater among students whose teachers reported heavier AI use before the intervention (Sungu et al., 2026).Those findings should make anyone advocating for teacher-facing AI uncomfortable.They make me uncomfortable.That is exactly why they are worth engaging rather than explaining away.I have spent considerable time studying whether generative AI can reduce some of the workload associated with teaching. It would be convenient if the evidence pointed only toward regained time, reduced stress, easier differentiation, and more manageable planning.But an intervention that improves adults’ working conditions while diminishing important aspects of students’ educational experiences would not be a successful educational intervention.Teacher well-being and student learning cannot be treated as competing scorecards from which we choose the result we prefer.At the same time, the most interesting part of the Sungu et al. paper, for me, is not the conclusion that AI can harm teaching.It is the question of how.The authors describe two broad possibilities. In the more optimistic account, AI enhances teacher capacity by taking on portions of planning and instructional-material production, allowing teachers to direct more attention toward differentiation, refinement, creative pedagogical problem-solving, and relationships with students.In the more pessimistic account, AI replaces some of the cognitive work involved in teaching itself. Teachers may spend less time considering the instructional decisions embedded within the materials they use. AI-generated resources may also lose some of the distinctive style associated with an individual teacher (Sungu et al., 2026).Precision matters here.The experiment established changes in student outcomes. It did not establish that reduced pedagogical reasoning or diminished teacher voice caused those changes. The authors present those possibilities as plausible mechanisms, and they acknowledge that different AI designs or workflows encouraging more active teacher involvement could produce different results (Sungu et al., 2026).Still, those hypotheses deserve to be taken seriously.Especially because they intersect so closely with what teachers told me.My participants often valued generative AI precisely because it could remove effort. Sungu and colleagues give us reason to wonder whether removing teacher effort can sometimes remove something students actually need.The temptation is to resolve that tension with a familiar slogan: AI should help teachers, not replace them.That is true.It is also too easy.The harder question is the one their study pushed me back toward:What kind of teacher effort are we trying to save?Some of the cognitive work involved in teaching is the profession’s very substance. Knowing the students. Anticipating misconceptions. Deciding why one example will work when another will not. Recognizing that an activity that looks perfectly reasonable on paper will fail with the twenty-six human beings sitting in this particular room.Adjusting an explanation because of the expression on one student’s face. Deciding that a technically correct answer is not the educationally appropriate one.If AI removes that effort, we may indeed have made teaching more efficient by making it worse.But other activities surrounding teaching consume enormous amounts of time without being synonymous with teaching itself.Before deciding that teachers should be wary of surrendering effort to machines, we also have to consider how much effort we already demand of them — and where that effort is currently coming from.Sometimes it is coming from Saturday.Sometimes Sunday.And that changes the question.III. The Classroom Did Not Meet AI Under Laboratory ConditionsBefore deciding what kinds of teacher effort should or should not be handed over to a machine, it helps to remember the conditions under which American teachers encountered generative AI.They did not meet it in a profession wondering what to do with its surplus capacity.The Learning Policy Institute’s 2026 national teacher-shortage scan estimated that approximately 387,843 teachers were working in assignments for which they were not fully certified, while another 37,569 teaching positions remained unfilled.Taken together, that amounts to at least 425,412 positions — about one in every eight teaching positions in the United States — that were either vacant or staffed by someone not fully certified for the assignment (Tan et al., 2026).That figure requires some care.It does not mean one in every eight classrooms is empty, nor does it suggest that nearly 400,000 people working in schools are incapable teachers. The measure combines actual vacancies with positions schools filled using teachers who had not yet met their state’s full certification requirements for those assignments.It is also a minimum estimate because vacancy and certification reporting varies across states. Nevertheless, the 2026 analysis marks the third consecutive annual increase in the combined shortage measure (Tan et al., 2026).Teachers already employed are not working abbreviated schedules.The 2025 State of the American Teacher Survey found that public school teachers reported working an average of 49 hours each week. That was an improvement from the 53 hours teachers reported in each of the previous two years, but it was still roughly ten hours more than their average contracted workweek (Steiner, Levine, Doan, & Woo, 2025).Other signs of improvement are worth acknowledging. The proportion of teachers who said they were likely to leave their current jobs by the end of the school year fell from 22 percent in 2024 to 16 percent in 2025 (Steiner, Levine, Doan, & Woo, 2025). The profession is not moving uniformly in one direction, and exaggerating a crisis adds little value when the numbers are already troubling enough.Working hours, however, tell only part of the story.A companion RAND analysis found that teachers had approximately 13 fewer hours of leisure time each week than comparable working adults. Forty-six percent of teachers said their jobs left them too tired for activities in their private lives, compared with 13 percent of similar working adults (Steiner, Woo, & Doan, 2025).Thirteen hours.That number changes how I hear the phrase “I got my weekend back.”Generative AI entered a profession in which many teachers were already borrowing time from somewhere else to make the job fit — an evening, an early morning, a planning period consumed by another responsibility, or a Sunday afternoon.This does not answer Sungu, Lira, and Duckworth’s concern.A teacher shortage cannot justify poorer instruction. Teacher exhaustion does not make cognitive offloading harmless. If AI causes teachers to surrender forms of professional judgment students need, the fact that teachers are overworked does not make that trade worthwhile.But workload changes the moral and organizational context of the trade.What do we mean when we say that AI “saves” a teacher an hour?If that hour previously belonged to the contractual workday, it may be reasonable to ask whether some of it could be redirected toward students. The teacher might spend more time providing feedback, working individually with a struggling student, improving tomorrow’s lesson, or collaborating with a colleague.But what if the hour AI eliminated was the hour between eight and nine on Sunday night?Was that newly available professional capacity?Or was it a debt the profession had been collecting from the teacher’s personal life?The distinction matters because much of the language surrounding workplace technology assumes that efficiency belongs to the organization. If a worker can accomplish in six hours what once took eight, the natural managerial question becomes what to do with the remaining two.That assumption is especially dangerous in teaching because the boundary around the workday has always been porous.There is almost always another worthwhile thing a teacher could do for students. One more lesson could be improved. One more family could be contacted. One more set of assignments could receive richer feedback. One more intervention could be individualized.There is no obvious point at which teaching announces itself finished.A technology capable of creating time can therefore do at least two very different things.It can create capacity for better teaching.Or it can return time that teaching never had a legitimate claim to in the first place.To understand which was happening, I had to examine more closely where teachers in my own study said the work actually went.IV. The Work Rarely DisappearedMy applied research project approached generative AI from a very different direction than Sungu and colleagues.I was not trying to determine whether AI caused changes in student achievement or motivation. Instead, I conducted a qualitative study with 13 educators from two K–12 schools in Northwest Indiana who were already using generative AI for substantive professional purposes.I wanted to understand a more fundamental question:What did using this technology actually do to their work?Did it reduce the effort required to teach? Did it simply shift that effort elsewhere? And when teachers saved time, did they actually experience less strain? (Schoon, 2026).The answer to the first question was fairly clear.GenAI often made things easier.Teachers described using it to escape the blank page. It could provide the beginnings of a lesson plan, draft a family communication, scaffold an existing text, generate an intervention resource, adapt materials for a particular learner, translate language, or provide a structure from which a teacher could begin working.The strongest pattern in my interviews was a reduction in what I eventually called front-end production effort: the labor required to move from an idea to something concrete enough to begin revising (Schoon, 2026, pp. 77–78).That is where a simple productivity story would end.The teacher used AI. The task took less time. Efficiency increased.But that was rarely the end of the story my participants told.Once the first draft existed, someone still had to decide whether it was right.Teachers checked factual accuracy. They changed wording. They adjusted reading levels. They removed material that did not fit their students. They reconsidered pacing. They tested lessons. They revised prompts when the first output was too generic. They checked whether materials aligned with their actual instructional goals rather than merely looking polished. They decided what student information could safely be entered into a system.Sometimes they decided that correcting what AI had produced would take longer than doing the task themselves.The work had not necessarily disappeared.It had changed addresses.Across the study, teachers frequently shifted from producing an instructional artifact to evaluating one. Instead of spending their effort asking, How do I make this? they increasingly asked, Is this accurate? Is it appropriate? Is it ethical? Does it fit these students? Would I actually teach this?For that reason, my findings did not support a simple conclusion that generative AI reduced teacher workload. More consistently, it redistributed effort, reducing some forms of production while preserving or increasing verification, customization, and professional judgment (Schoon, 2026, p. 81).That redistribution could still be enormously valuable.For many teachers, the exchange was worth it. The machine took over unpleasant or time-consuming starting tasks, allowing the teacher to spend more of the remaining time on work requiring greater expertise.In one case, a teacher generated initial lessons earlier and used the recovered time to rehearse, refine, and improve their delivery. Others said AI reduced stress, helped them keep up with planning, or prevented unfinished work from following them home (Schoon, 2026, pp. 83–84).That looks remarkably close to the optimistic possibility Sungu and colleagues describe: technology reduces lower-value production work and makes room for more meaningful pedagogical work.But I found another possibility.Sometimes efficiency produced more work.A generic first draft still required student-by-student customization. A supposedly fast AI interaction turned into an hour spent perfecting prompts. Saved drafting time became checking time.A teacher who could generate materials more quickly could also imagine a future in which faster production became the new baseline expectation: if AI makes differentiation easier, perhaps more differentiation will be expected; if a polished document takes twenty minutes instead of an hour, perhaps three polished documents begin to seem reasonable (Schoon, 2026, pp. 84–87).And some of the new labor had almost nothing to do with teachers generating materials at all.Student access to GenAI changed how teachers approached assessment and academic integrity. One participant no longer trusted homework completed outside class as reliable evidence of what students understood, so she increased the amount of direct formative assessment she conducted during class.Another described keeping student-monitoring software open constantly because students could upload screenshots of online assignments to AI systems and immediately retrieve answers. Teachers also described uncertainty about privacy, acceptable-use policies, approved platforms, and what responsible adoption should look like (Schoon, 2026, pp. 85–87).That creates an important complication in the idea of “AI-assisted teaching.” AI can change a teacher’s workload even when the teacher is not using it.When students can use GenAI, assessment practices may have to change. When administrators adopt new GenAI systems, teachers may have another platform to learn. Once AI-generated work becomes commonplace, expectations about speed and polish may change. Once schools recognize that routine materials can be produced almost instantly, what once counted as an impressive degree of individualization may quietly become the minimum.The appropriate unit of analysis, then, is not simply the prompt.It is the workflow around the prompt.My participants’ experiences suggested that saved time is fragile. It can survive long enough to become rest, family time, greater attention to students, or better instruction. But it can also be quickly reclaimed by checking, customization, monitoring, policy uncertainty, professional learning, or rising expectations.That leaves us with a more complicated way of evaluating teacher-facing AI.The question cannot merely be whether AI reduces effort.It has to be which effort it reduces, which effort replaces it, and what happens to the space left behind.And that is where I think the idea of “cognitive offloading” needs to become much more precise.V. Not All Cognitive Offloading Is EqualSungu and colleagues describe one risk of teacher-facing AI as cognitive offloading.If a teacher delegates the creation of a lesson or instructional resource to a generative system, the concern is not simply that the machine has typed the words. The teacher may also have avoided some of the reasoning that would otherwise occur while creating them: selecting examples, anticipating confusion, sequencing ideas, considering alternatives, or grappling with how best to explain something.That concern is persuasive.But I think “cognitive offloading” is too broad a category to tell us whether the offloading is good or bad.Teachers offload cognition constantly.A calculator relieves me of arithmetic so I can concentrate on something else. A gradebook remembers scores, so I don’t have to. A curriculum map prevents every teacher from independently reconstructing the order in which standards should be addressed.Spellcheck catches errors without requiring a writer to reread every word specifically for spelling. Search engines replaced much of the cognitive labor once required simply to locate information.We generally do not evaluate those tools by asking whether they reduced thought.We ask what kind of thought they reduced and what the person could do instead.Generative AI deserves the same distinction.Some cognitive effort is little more than friction surrounding the work.A teacher who knows exactly what she wants to communicate to families but spends forty minutes turning those ideas into a polished newsletter has unquestionably performed cognitive labor. The same is true of the teacher converting an existing passage into three reading levels, reformatting assessment questions, producing another set of examples, or staring at the first sentence of a lesson plan because a required template demands that an idea already clear in his head be translated into a particular written form.My participants repeatedly described GenAI as useful for this kind of front-end production. The benefit was strongest when teachers already possessed the professional knowledge needed to evaluate whatever appeared on the other side. AI could help them move from an idea to a workable artifact more quickly; the teacher still decided whether that artifact deserved to survive (Schoon, 2026).I see little pedagogical virtue in preserving friction for friction’s sake.There is no inherent educational value in a teacher spending forty-five minutes writing something that could responsibly be produced in fifteen.Difficulty is not evidence of quality. Time spent is not evidence of care.However, AI can participate in another category of cognition without completely owning the work.A teacher might ask a system for five possible ways to introduce a concept, common misconceptions students have about it, alternative examples, possible scaffolds, or questions that could reveal misunderstanding. Here, AI is not simply automating production. It is extending the teacher’s field of possibilities.But the usefulness of that assistance depends on what happens next.Someone still has to decide which suggestion fits. Someone has to notice that the “common misconception” is not actually common among these students, that the clever example assumes background knowledge they do not possess, or that the proposed scaffold solves a problem the class does not have.AI can widen the menu. It should not automatically order dinner.And then there is a third kind of cognitive effort, the kind I am much less willing to surrender.Knowing the students. Determining what matters. Recognizing when an answer that is technically correct is instructionally wrong. Understanding why yesterday’s discussion means today’s lesson needs to change. Deciding when to abandon the plan.Judging whether a text will challenge students productively or simply frustrate them. Recognizing that one student needs another explanation while another needs to struggle for another minute before receiving one.Understanding the ethical consequences of an instructional decision. These are not inefficiencies surrounding teaching. They are teaching.This is where Sungu and colleagues’ warning matters most. Their participants used the AI tool predominantly for instructional-material preparation, and interactions were generally brief, often involving limited back-and-forth. If the machine increasingly determines not merely how an instructional idea gets expressed but what the instructional idea should be, then efficiency can become substitution.But this distinction also prevents us from reaching the opposite, equally mistaken conclusion: that responsible AI use requires teachers to preserve all the labor they currently perform.It does not.The goal should not be to maximize the amount of cognitive effort teachers invest in their jobs.It should be to protect the cognitive effort that good teaching depends on. Some of the cognitive work of teaching is teaching. Some of it is simply what teachers have had to do before they could get around to teaching. We should be very careful not to confuse the two.VI. The Teacher Has to Know Enough to Know What Is WrongOne of the most troubling findings in the Sungu study also resonated most strongly with my own research.Across the full sample, giving teachers access to the AI tool did not significantly affect students’ academic performance. But that average concealed an important difference. Among students taught by teachers in the study’s below-median baseline-performance group, AI access was associated with academic performance approximately 0.13 standard deviations lower on standardized assessments. Those students also experienced a larger decline in confidence (Sungu et al., 2026).I want to be careful with the phrase “lower-performing teacher.”Sungu and colleagues did not directly observe teachers and rate their pedagogical expertise. Their baseline-performance measure divided teachers within departments according to their students’ average Term 1 scores.That is a meaningful indicator within their analysis, but it is not synonymous with a comprehensive measure of teacher skill. Student scores reflect more than the person standing at the front of the classroom.The result deserves attention without turning a statistical moderator into a label for the people involved.With that qualification, however, the pattern raises an important possibility.The people we might expect AI to help most may not automatically be the people best positioned to use it well.Sungu and colleagues propose what they call a skill-substitution mechanism. Teachers with less developed expertise may have greater difficulty critically evaluating AI-generated material, adapting it, rehearsing it, internalizing it, and turning it into something genuinely their own. In that case, the system does not fill a skill gap so much as conceal it beneath fluent output (Sungu et al., 2026).My study cannot test that mechanism. It was qualitative, small, and not designed to compare teachers by effectiveness.But my participants repeatedly described the inverse relationship.GenAI was easiest to use efficiently when teachers already knew enough to know what was wrong.Teachers with strong professional knowledge could quickly recognize when an AI-generated response was inaccurate, too simple, poorly worded, misaligned to a standard, or unrealistic for the time available.Their expertise allowed them to treat GenAI output as a starting point rather than a finished product. Because they could evaluate and revise the material efficiently, they were better positioned to benefit from the technology.The clearest reductions in effort in my study occurred when teachers already possessed enough professional knowledge to assess AI-generated content quickly and confidently (Schoon, 2026, p. 75).That suggests an uncomfortable formulation:Generative AI may be more of a competence amplifier than a competence equalizer.The same output can function very differently in different hands.For one teacher, it is clay.For another, it looks finished.That distinction has implications well beyond whether teachers know how to write effective prompts. Much AI professional development begins with tool capabilities: here is how to generate a lesson plan; here is how to differentiate a text; here is how to create a rubric.Those skills matter.But perhaps the more important AI skill is the ability to reject what AI gives you.That capacity is not fundamentally technological. It comes from content knowledge, pedagogical knowledge, knowledge of students, experience, ethical judgment, and the confidence to distrust fluent language.If safe and effective use of generative AI depends partly on expertise, then giving teachers access to increasingly powerful systems without strengthening the professional knowledge required to supervise them may be exactly backward.The teacher does not become less important as the AI becomes more capable.The teacher may become the quality-control system that capability depends on.VII. The Teacher’s FingerprintsSungu and colleagues offer another possible explanation for the decline in student motivation, and it is harder to measure.Instructional materials created with AI may simply feel less like they came from the teacher.The authors suggest that AI-generated materials can be technically adequate while bearing more of the model’s imprint than the teacher’s. If teachers use similar systems to generate similar instructional artifacts, lessons may begin to converge in language, structure, and style. Students who have spent months learning a teacher’s peculiarities may notice when those peculiarities disappear (Sungu et al., 2026).Again, this is a proposed mechanism, not something their experiment directly demonstrated.Students were not asked whether materials felt less authentic or whether they could detect AI-generated work. The study shows a decline in intrinsic motivation, with a steeper decline among students whose teachers reported heavier prior AI use. Loss of teacher voice is one plausible explanation.Still, I find the hypothesis difficult to dismiss.Classrooms accumulate fingerprints.A teacher has favorite examples. Repeated phrases. Running jokes. Stories students have heard too many times. Particular ways of explaining difficult ideas. Strange analogies that somehow work. References to something that happened in class three days ago. Knowledge of which students will laugh at something and which will roll their eyes.Over a school year, a class develops its own language.That language is inefficient.It is also intensely human.A standardized understanding of personalization tends to focus on differentiation: different reading levels, different scaffolds, different practice sets, different pathways through the same content. Generative AI is exceptionally good at producing this kind of variation quickly.But another form of personalization is much harder to mass-produce.Sometimes material is personalized because it bears evidence that a particular person made it for particular other people.The worksheet refers to the disastrous lab from Tuesday.The writing prompt jokes about something the class argued over last week.The example uses a student’s favorite basketball team because the teacher knows it will get his attention.The teacher intentionally uses an awkward sentence because she knows her students will catch the mistake and enjoy correcting her.None of these choices would necessarily raise the score on a rubric evaluating the material’s technical quality.They might nevertheless matter.One theme that emerged unexpectedly in my research concerned authenticity and professional legitimacy. Some teachers worried that AI-supported work might appear less authentic, less professional, or less deserving of recognition, even when they had substantially revised it and taken responsibility for what the system produced (Schoon, 2026).That concern initially led me toward questions of professional reputation: if a teacher uses AI, do colleagues or administrators assume that she has done less real work?Sungu and colleagues’ findings make me wonder whether authenticity might have another audience.Students.Some of what teachers add to instructional materials has educational value precisely because it is unnecessary for information delivery.The teacher’s fingerprints tell students this didn’t simply come from somewhere.Someone prepared it for us.That does not mean teachers should handcraft every worksheet from scratch. It certainly does not mean we should preserve Sunday-night lesson planning because students can somehow detect the love embedded in exhaustion.It means that when AI performs some of the production, the teacher still has to leave a recognizable imprint on the work.In my applied research project, one participant described this almost perfectly when discussing AI-generated material:“I will add me.”I originally understood that phrase largely as an expression of professional ownership.After reading Sungu and colleagues, I think it may also contain a pedagogical principle.The question is not whether AI touched the lesson.It is whether the teacher did.VIII. Who Owns the Saved Time?One part of Sungu and colleagues’ pessimistic account I find especially revealing.They raise the possibility that teachers may not use the time saved by AI to improve their professional performance. To illustrate the idea, they draw on an economic example involving taxi drivers who respond to unusually productive working conditions by reaching their earnings target sooner and stopping work rather than using the opportunity to earn as much as possible.In the educational version of the problem, AI might allow a teacher to create instructional materials faster without causing the teacher to reinvest the saved time in better teaching (Sungu et al., 2026).Economically, I understand the point.Educationally, I think it reveals an assumption worth interrogating.Suppose a teacher spends three hours every Sunday preparing for the coming week.She begins using generative AI responsibly. She does not hand over instructional decisions. She uses it to produce initial drafts, routine adaptations, parent communications, and other artifacts she then reviews. Eventually, she can accomplish the same work in ninety minutes.What should happen to the other ninety?One answer is obvious: she should reinvest it.She could improve Monday’s lesson.She could create additional differentiation.She could provide more detailed feedback.She could prepare an enrichment opportunity.She could contact families.She could study student data.Every one of those uses could benefit students.And this is what makes teaching unusually vulnerable to work intensification:There is always another morally defensible thing a teacher could do.The work has no natural ceiling.My applied research was grounded in Effort–Reward Imbalance, a framework developed by Johannes Siegrist that treats work not simply as a set of tasks but as a social exchange. Workers invest time, energy, skill, and commitment while expecting adequate returns in compensation, esteem, recognition, autonomy, security, and status. When high effort is not matched by adequate reward, sustained strain can result (Siegrist, 2016; Schoon, 2026).That framework mattered in my study because a faster task did not necessarily mean a better working condition.If a new expectation absorbed every efficiency, the teacher might produce more without experiencing any meaningful reduction in strain.Participants described exactly that possibility.One teacher worried that easier production could increase expectations for how much teachers ought to produce. Others described savings being reabsorbed through checking, customization, assessment redesign, student monitoring, policy interpretation, or learning new platforms.I described this as expectation creep: the possibility that time saved through GenAI could be absorbed by higher expectations for volume, speed, polish, or individualization until the workload benefit effectively disappeared (Schoon, 2026).That possibility changes the question I want to ask about teacher productivity.Not:Did teachers reinvest the time AI saved?But:Who owns the time AI saved?Some of it may absolutely belong back in the classroom.If AI eliminates thirty minutes of formatting during a teacher’s planning period and she spends that time talking with a student who needs help, that seems like an extraordinary use of technology. If AI gets the first draft of a lesson out of the way and the teacher uses the remaining time to rehearse, interrogate, improve, and personalize it, Sungu and colleagues give us reason to believe that this reinvestment may matter.But I cannot accept the premise that every saved minute automatically becomes an organizational asset.Especially when the minute came from outside the workday.A teacher who used to work from eight until nine on Sunday evening and now does not has not necessarily “failed to reinvest” a productivity gain.That judgment only makes sense if Sunday evening belonged to the school in the first place.This is why the workload statistics matter so much to the AI conversation.American teachers already work substantially beyond their contracted hours. They have less leisure time than comparable workers. Schools are already struggling to fill teaching positions. The problem we are trying to solve is not how to squeeze unused capacity out of an underworked profession.We should be careful about adopting technology on the promise that it will relieve teachers and then treating that relief as evidence that teachers can absorb more work.That would turn AI’s most attractive promise into another mechanism of intensification.There are, then, at least two ways teacher-facing AI can fail.In the first, AI removes the wrong work. It substitutes for professional reasoning, teacher voice, contextual knowledge, or the relational dimensions of teaching. Teachers save time, but students receive something worse.Sungu and colleagues give us good reason to worry about that failure.In the second, AI removes the right work — but nobody lets it stay removed. Every minute saved from routine drafting becomes another individualized product. Every faster communication becomes an expectation for more communication. Every streamlined plan justifies a more elaborate planning requirement.My participants give us reason to worry about that failure.We should refuse both.The goal is neither to protect every existing form of teacher labor nor to automate as much of teaching as technologically possible.It is to remove unnecessary labor, preserve necessary judgment, protect instructional quality, and allow at least some recovered capacity to remain recovered.Sometimes the best use of an hour saved by AI will be better teaching.And sometimes the best use will be absolutely nothing that can be entered on a teacher evaluation.That is not a flaw in the technology.It may be one of its most valuable outcomes.IX. What Responsible Teacher-Facing AI Should Actually OptimizeOn the practical implications, my position is closer to Sungu and colleagues’ than my disagreement over saved time might suggest.They do not conclude that generative AI has no place in education. Instead, they argue that schools should resist assuming that teacher-facing AI automatically improves instruction. They recommend professional development that helps teachers use AI as a complement to pedagogical judgment, stronger guardrails around its use in core instructional work, differentiated support for teachers who may be more vulnerable to substitution, and rigorous evaluation before broadly scaling systems (Sungu et al., 2026).I agree with all of that.Their study also suggests that implementation matters enormously. Teachers received a one-hour training session on the tool, and the authors acknowledge that alternative designs — particularly systems that require more active teacher engagement or provide stronger guardrails — could produce different results (Sungu et al., 2026).But I would add another set of questions before calling an implementation successful.Schools should ask not simply whether teachers can use the tool, but whether the tool is changing teacher work in the direction they actually want.Does it reduce net effort, or make one visible task faster while creating three invisible ones?Does it preserve teachers as the final evaluators of instructional quality?Does it make professional judgment easier to exercise, or easier to bypass?Do teachers have enough knowledge and support to recognize weak AI output?Are policies clear enough that teachers do not have to independently determine which uses are permitted?Does the system preserve professional ownership, or does AI-supported work become something teachers feel obliged to hide?Are teachers given time to learn and evaluate new systems, or is the cost of adoption quietly added to an already overloaded job?Most importantly:When the tool saves time, what happens next?My recommendations at the end of my applied research project argued that generative AI should be treated as a professional-practice issue, not simply a tool-use issue. The most productive uses in my study tended to preserve teachers’ responsibility for verification, customization, alignment with student needs, and final instructional decisions.I also recommended that school leaders explicitly examine whether saved time produces meaningful relief or is instead reabsorbed by additional expectations and new forms of labor (Schoon, 2026, p. 106).That distinction becomes even more important in light of the Sungu findings.A professional development session titled 50 Prompts That Will Save You Hours might teach exactly the wrong lesson.Speed is not the objective.More output is not the objective.Even “AI adoption” is not the objective.The objective is better professional capacity.Sometimes that means producing something faster. Sometimes it means making a better decision. Sometimes it means having enough time and mental bandwidth to notice the student who needs you. Sometimes it means retaining the energy required to make tomorrow’s lesson recognizably yours.And sometimes it means leaving work at work.Educational leaders therefore need to replace one of the most common questions surrounding generative AI.Instead of asking:How can we get teachers to use AI?ask:What kind of teacher work are we trying to create?Once that question comes first, the technology becomes subordinate to the answer.That is where it belongs.X. Back to the WeekendSungu, Lira, and Duckworth have produced evidence that advocates for generative AI in education should take seriously.Their experiment shows that teacher-facing AI can harm student outcomes under at least some conditions.It warns against one of the easiest mistakes we can make with generative AI: assuming that because a tool is designed to make instructional production faster or easier, the resulting instruction must therefore be better.My research led me to a different warning.Because faster does not necessarily mean easier, either.Teachers can save time on one task only to lose it to verification, customization, monitoring, new policies, new expectations, or new problems created by students’ own use of AI. Efficiency is real, but it is fragile. In the conclusion of my study, I argued that saved time can become rest, focus, or renewed attention to students — or be consumed almost immediately by new demands (Schoon, 2026, p. 112).Taken together, these findings leave me somewhere between enthusiasm and fear.Which is probably where we should be.If generative AI removes the part of teaching that requires knowing students, making judgments, noticing what does not fit, adapting to context, and bringing something recognizably human into the classroom, then we have saved the wrong work.But if it removes the blank pages, repetitive drafting, formatting, clerical friction, routine transformations, and unnecessary hours that have slowly colonized teachers’ evenings and weekends, we should not rush to fill the space back up.We should protect it.Sometimes responsible use of AI means a teacher spends the saved hour improving a lesson.Sometimes it means spending the hour with a student.Sometimes it means thinking more carefully about something that deserves more thought.And sometimes the teacher should go home.That is not abandoning the human work of teaching.It may be part of preserving the human being who has to do it.So I keep returning to the teacher in my study.She used generative AI.She still revised.She still refined.She still made the instructional decisions.And she got her weekend back.We should be ambitious about what artificial intelligence can do for education.But perhaps, in this one respect, we should also be satisfied with something that modest.ReferencesSchoon, R. R. (2026). “I will add me”: Generative AI, teacher work, and effort-reward imbalance in K–12 education [Applied research project capstone, Purdue Global].Siegrist, J. (2016). A theoretical model in the context of economic globalization. In J. Siegrist & M. Wahrendorf (Eds.), Work stress and health in a globalized economy: Aligning perspectives on health, safety and well-being (pp. 3–17). Springer.Steiner, E. D., Levine, P. R., Doan, S., & Woo, A. (2025). Teacher well-being, pay, and intentions to leave in 2025: Findings from the State of the American Teacher Survey. RAND Corporation. RR-A1108–16.Steiner, E. D., Woo, A., & Doan, S. (2025). To make teaching sustainable, help teachers balance work and personal demands: Findings from the 2025 State of the American Teacher Survey. RAND Corporation. RR-A1108–20.Sungu, A., Lira, B., & Duckworth, A. L. (2026). Generative AI can harm teaching. SSRN.Tan, T. S., Baez, G. A., & Kemper Patrick, S. (2026). State teacher shortages 2026 update: Teaching positions left vacant or filled by teachers without full certification. Learning Policy Institute.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!I Got My Weekend Back was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Generative AI Pub — Published August 27, 2026 at 14:11 UTC — Category: Image AI