When Chatbots Finish the Homework, Real Work Gets Scarce

College Board surveys found that 84% of U.S. high school students used generative AI for schoolwork by May 2025. The share had been 79% in January.

Take-home essays, problem sets, and reading summaries now finish in a chat window. The grade still pretends the student did the labor, so the only schoolwork that still proves a mind is the kind a chatbot cannot submit.

Elizabethton High School Took the Case Files

In 2017, sociology teacher Alex Campbell at Elizabethton High School in Tennessee handed his class a problem police had not closed in more than three decades. He told them it would be the hardest course they had ever taken. They went after the Redhead Murders, a string of unsolved roadside killings of young women across the South from the late 1970s through the early 1990s.

Students tracked down an FBI profiler, interviewed a victim’s brother, built timelines from old case files, and flagged people who had not been counted as victims. They pulled in a true-crime podcaster already working the same ground, wrote their own profile of the killer, and landed on a theory the Tennessee Bureau of Investigation later supported. The work became the 2024 podcast Murder 101. A later class fed a three-episode Prime Video docuseries of the same name, which launched worldwide on July 13, 2026, and cracked the service’s top 3 series rankings that month.

Campbell said the project worked because it was hard, not because it was easy.

I make it obvious to them up front that this is going to be very hard, but they’re going to be invested, and so that makes it easy to do hard work. If you ask people about the most meaningful experiences of their life, they’re almost always hard.

Alex Campbell, sociology teacher, Elizabethton High School

Kids can spot a fake in about two seconds, he said. If the answer is in the back of the book, they know the problem is fake. A cold case with open files and living families is not that.

84 Percent of High Schoolers Now Outsource Schoolwork

Campbell’s class is a rare room. Most American high school work still looks like the packet a chatbot can complete before lunch.

THE FOUR YEARS THAT BROKE THE ASSIGNMENT

1. 2017: Campbell assigns the Redhead Murders and tells students the class will be brutal.
2. 2022: ChatGPT arrives; take-home writing and summaries become instant.
3. 2024: The Murder 101 podcast lands, and Gallup’s Voices of Gen Z study finds engagement slipping.
4. January to May 2025: College Board’s student surveys show GenAI use for schoolwork rising from 79% to 84%.
5. December 2025: RAND’s American Youth Panel finds homework AI use still climbing, along with fear that it dulls thinking.
6. July 13, 2026: Murder 101 premieres on Prime Video in more than 240 countries and territories.

College Board Research, drawing on surveys of students, parents, and educators from June 2024 through June 2025, found that 84% of high school students used generative AI for schoolwork by May 2025, a 5 percentage-point jump in four months. Half used it to brainstorm, edit or revise essays, and/or hunt sources. ChatGPT was the default: 69% said they used it for assignments and homework that May.

RAND asked a broader enrolled sample, from middle school through college, a blunter homework question. The share who 62 percent used AI for homework help in December 2025, up from 48 percent in May. ChatGPT use in that panel rose from 45 percent to 53 percent over the same span. Students most often wanted better explanations of assignments (38 percent), brainstorming (35 percent), facts (33 percent), and drafting or revising (33 percent).

Chris Hulleman, a University of Virginia researcher who founded the school’s Motivate Lab and has studied student motivation for more than 20 years, said the social contract changed. Students now ask why the task exists. “And if it’s boring or seems stupid to them, then yeah: Why wouldn’t I use this free thing that’s at my fingertips?”

Students Fear Getting Dumber and Still Hit Send

The same teenagers who lean on the chatbot also say it might be making them worse at thinking. That split is the tell. They are not confused about the cost. They are choosing speed anyway.

WHAT STUDENTS AND TEACHERS SAY THE TOOL COSTS

• Student fear: Two-thirds of high schoolers in College Board’s October 2024 BigFuture survey agreed that using AI too much will make them dependent and less intelligent, even as 85% said students can benefit from it.
• Thinking harm: By December 2025, 67 percent of RAND’s enrolled youth agreed that more AI for schoolwork harms critical thinking, up from 54 percent in February 2025.
• Teacher alarm: More than 8 in 10 AP teachers in the College Board brief agreed AI makes students overly dependent, less likely to build critical thinking, less able to express original ideas, and less likely to engage deeply with material.
• Grade signal: In an ACT Research brief dated January 2026, 71% of 4,903 surveyed students at least slightly agreed that high school grades no longer accurately reflect performance because students can use AI to cheat.

Among RAND respondents who skipped AI for homework, 78 percent said it harms thinking. Among those who used it, 60 percent said the same thing and used it anyway. College Board’s student sample was just as split on the trade: 50% neither agreed nor disagreed that benefits beat risks, with the rest divided down the middle.

Darren Lam, then 18 and a recent graduate of Staten Island Technical High School in New York, described classmates who refuse the tools on principle and others who let the model do school. The second group, he said, is “just not even really at school anymore, basically the AI is the student.” Even lighter users, he said, hand the bot the ugly assignments or buy time they would not have if they did the work straight.

Colorado math teacher Dylan Kane, who writes the newsletter Five Twelve Thirteen, compared the shift to indoor plumbing. Pipes freed people from hauling water and also erased the daily walking that used to come with the chore. Chatbots, he argued, are automating cognitive effort faster than schools have found a replacement workout. “People are getting dumber,” he said, “and this is a cross-generational phenomenon.”

OpenAI chief executive Sam Altman told a Stanford CS153 audience he had expected schools to rebuild around the tools. “I thought the education system would redesign itself in one year. I struggle to point to any significant systemic change in three and a half years.” He warned of atrophy in critical thinking if teaching and testing stay as they were before chatbots.

The Grade Game Makes Honesty Feel Naive

Ask students why they go to school, said Eric Anderman, an Ohio State University educational psychologist who studies motivation and cheating, and the answers split. Some say they are there to learn. Many say they are there to get grades. When teachers talk tests and letters more than mastery, cheating rises. When the message shifts to learning, it falls.

Anderman tells future teachers that “‘A’ stands for ‘artificial.’ We made it up.” He wants less talk about the test. The letter still sorts college files, scholarships, and parental peace, so the market for a clean draft has not moved.

Ben Riley, founder of Deans for Impact and the think tank Cognitive Resonance, put the peer math in one line: students “don’t want to be suckers.” If classmates boost grades with a model, doing the reading the long way feels like a donation.

WHY THE SHORTCUT STILL WINS THE HOUR

• The sort: ACT’s January 2026 brief found 81% at least slightly agreed it is easy to use AI to cheat on high school assignments, and 59% agreed users get better grades than students who do not.
• The arms race: School filters block ChatGPT on Wi-Fi; phones and hotspots walk around the block. Rewrite tools sand the voice so detectors argue with themselves.
• The time cut: The assignment is still due at 11:59 p.m. The bot returns a draft in seconds, so the honest night of work is the expensive choice.
• The prisoner’s problem: Rebecca Winthrop, a Brookings Institution senior fellow and co-author of the 2025 book The Disengaged Teen, said students want to learn and dislike hacks, then fear being the only one who refuses.

A USC Center for Generative AI and Society report led by Stephen J. Aguilar, an associate professor at the USC Rossier School of Education, found the same default in college. In a national survey of about 1,000 U.S. college students, most reached for what the researchers call executive help, a fast answer with little effort. Far fewer used instrumental help, the kind that clarifies a concept so the student can do the next problem. Professors who steered them toward the second path saw more of it. Left alone, students took the short road even when they learned less.

College Board’s December 2025 follow-up, built on more than 10,000 student, parent, and teacher responses, found the heaviest classroom users were already the students with the most academic room: seniors and students earning mostly As and Bs, who were nearly twice as likely as ninth graders to use GenAI for brainstorming, essay revision, research, and explanations. Use also rose with parental education. Two-thirds of students whose parents hold graduate degrees reported frequent use. Frequent use was 65% among Black and Hispanic students, 58% among Asian students, and 55% among White students.

What a Real Problem Does to a Classroom

The Gallup and Walton Family Foundation Voices of Gen Z survey, fielded April 26 to May 9, 2024 among 4,157 people ages 12 to 27, including 2,317 enrolled in K-12, already showed a school day that felt thin before the 2025 usage spike. Only 49% said their schoolwork challenged them in a good way. Only 46% said they got to do what they do best. Just 17% strongly agreed they had learned something interesting at school in the last seven days. Gallup’s write-up found that fewer than two in 10 strongly agree that classwork feels important, interesting, challenging, or lined up with their talents.

The 25% of students with the highest engagement scores were more than twice as likely as the lowest 25% to be thriving (76% versus 32%) and more than four times as likely to strongly agree they have a great future (61% versus 15%). When students named what lit them up, 60% credited a teacher who made the material interesting, 46% wanted hands-on work, and 35% wanted a tie to the real world. Winthrop’s line for the pre-ChatGPT culture was “a very transactional educational process.” The bot did not invent that process. It finished the transaction without the student.

Campbell’s answer is a ladder, not a speech. First a conversation with a victim’s brother. Then police and the files. Then timelines, profiles, and maps. “Before you know it, they’re building timelines,” he said. “They’re doing profiles, they’re doing geographic studies, and then they slowly begin to step up and realize they’re learning the skills, they’re meeting the people.” Inviting teenagers to work a problem no expert has closed, he said, tells them they deserve meaningful work with real consequences.

That idea is old. In the mid-1980s, educator Seymour Papert, co-founder of the MIT Media Lab and a designer of the Logo computer language, popularized “hard fun” after a student called a tough coding task fun. Papert wrote that the kid called it fun because it was hard rather than in spite of being hard. Kane wants schools to treat thinking the way gyms treat walking once elevators made stairs optional: honest about the lost labor, then deliberate about the workout. For some students that workout is still reading, writing, and math. For others he would grow chess, robotics, woodworking, debate, and drama, and try to get students good at them, not merely expose them.

Other campuses have run versions of that bet for years: students designing houses with engineering firms, building a hospital app, helping Kansas City rethink airport services, earning FAA drone certification to modernize farming on family land in New Mexico. Ulcca Joshi Hansen, an education researcher and author of the 2021 book The Future of Smart, points to project-based campuses such as Iowa BIG, One Stone in Idaho, Rock Tree Sky in California, and the Big Picture Learning network, which wraps high school around internships with local businesses and nonprofits. Youth programs like Boys and Girls Clubs and summer camps, she said, already invite kids into hard, real activity around things they care about.

Paper Tests Still Decide What Counts as Knowing

Hansen’s obstacle is not a missing lesson plan. It is the scoreboard. The accountability system, she said, still tells schools they do not know math if they can build something; they only know it if they can show it on a piece of paper. ESSA, the federal K-12 law, already allows portfolios, projects, and extended-performance tasks in state systems, and it created a narrow Innovative Assessment Demonstration Authority. Few states have made that waiver work. Sit-down tests remain the thing a superintendent can defend.

Hulleman has watched that efficiency logic for two decades. If higher scores on standardized tests are the primary output, he said, “you’re not going to have a very interesting system. And that’s sort of what we repeatedly do.” After COVID, Campbell heard students treat campus as optional. They could watch YouTube, use Khan Academy, and turn work in through Canvas. His reply is that school has to offer challenging, doable work they cannot get anywhere else. Most buildings still offer the packet.

HOW HIGH SCHOOLS POLICE THE CHATBOT

Separately, 55% of principals said their high schools had not blocked GenAI tools on the school network, and 45% reported some access limits. Nearly 6 in 10 parents (57%) said it is better for students to use the tools for schoolwork than to skip them. More than 85% of administrators called learning to use AI valuable. The adults want literacy. The product they still collect at 11:59 p.m. is a file a model can write.

Educator Priten Soundar-Shah, author of Ethical Ed Tech: How Educators Can Lead on AI and Digital Safety in K-12, flipped the labor. Students read, watch lectures, or even talk to chatbots role-playing historical figures on their own time. Class is for discussion, debate, and writing under his eye. “I don’t do any instructional time during class,” he said. Live presentations are the test. “There’s no point in offloading your conversations to an AI chatbot if you’re going to have to show up to class and still demonstrate that knowledge tomorrow.” He has also moved toward mastery grading, with retakes until the student actually has the skill.

His Paris test for students who say translation earbuds kill French class is simple. Who would you rather trust in a pinch: a fluent speaker, or someone waiting for the buds to catch up? Students get it. They are more conflicted than the adult panic implies. They still sit in a system that pays out for the file, not the fluency.

Who Gets the Work That Still Requires a Mind

If the only K-12 tasks that still demand a human are public, supervised, or tied to a real consequence, those tasks become a tracking system. Campbell’s students got case files, families, and a streaming series. Big Picture students get internships. A teenager in a test-prep factory gets a packet and a chatbot, then a letter grade that ACT’s own respondents already distrust.

Winthrop said the job market no longer forgives coasting on basic literacy and numeracy. “You’ve got to have a lot of higher-order thinking skills,” she said. “The bar has shifted, and it’s much higher.” The students who still practice those skills in front of a teacher will hold a credential the transcript can no longer guarantee. The students who do not will have a GPA and a chat history.

WHERE THE EVIDENCE SPLITS ON AI IN CLASS

• Default use dulls the work: Aguilar’s USC team found students left alone pick executive help. Kane, Altman, and RAND’s 67 percent are describing that path, the essay mill sitting on top of a packet curriculum.
• Guided use can be a tutor: The same USC work found professors who asked for instrumental help got more of it. El Salvador President Nayib Bukele, describing a World Bank summary of PISA for Schools, said 1,198 students in 171 public schools in a broader reform that includes AI tutors posted reading, math, and science results comparable to averages in Germany and Sweden. He said the schools were not a random national sample, that AI was not proven as the sole cause, and that teacher attendance in those schools rose from 76% to 97%.
• Scale is the unsolved part: Bukele said more than 1,000 schools were already in the program and that every public school in the country is supposed to be inside 18 months, with national PISA in 2029 as the later test. U.S. high schools, by College Board’s count, have not made a comparable redesign. About 13% encourage GenAI in all classes; about 40% try to ban it.

Hansen has watched a decade of invitations to let go of the old machine. Most campuses, she said, keep tweaking it and bludgeoning kids to do more of it. That has not worked. Campbell’s students, meanwhile, walked a red carpet at the historic Bonnie Kate Theater in downtown Elizabethton when Murder 101 premiered, then watched their own class on Prime Video. The killings are still unsolved. The files are still open. That is the point of the assignment.