The Homework Problem Is Now an AI Problem

The question 0:01

If a student's homework scores rise and their exam scores fall at the same time, what exactly has the homework been measuring?

This is WAC Signal, from the World Assessment Council. Evidence and ideas for the changing world of learning.

Episode five. The homework problem is now an AI problem.

The signal 0:32

RAND's American Youth Panel found that the share of US students, from middle school through college, using AI for homework help rose from forty-eight percent in May 2025 to sixty-two percent in December 2025, driven by middle and high schoolers.

Younger students see what is happening around them. In Pew Research Center's survey of US teens aged thirteen to seventeen, fielded in September and October 2025, fifty-four percent had used chatbots for schoolwork help, and one in ten did all or most of their schoolwork with a chatbot's help. Fifty-nine percent said students at their school use chatbots to cheat at least somewhat often.

And students are uneasy about it. In RAND's December 2025 survey, sixty-seven percent agreed that using AI for schoolwork will harm critical thinking, up from fifty-four percent in February. Even among students who used AI for homework, sixty percent agreed.

So the use is rising while the people using it worry about what it does. The question is whether the worry is justified, and what that means for the work we set outside the classroom.

The evidence 1:53

The most eye-opening study of 2026 comes from secondary schools in central China. Economists David Strömberg, Victor Lei and Yanhui Wu followed twenty-six thousand, eight hundred and eleven students in grades seven to twelve for thirty months, linking monthly closed-book exams, high-stakes entrance exams and homework records across nine subjects.

Because students began using AI at different times, the researchers could compare each student before and after first use with classmates who had not yet started. Within six months of adopting AI, homework scores rose eighteen percent, and homework time fell thirty percent. But monthly exam scores fell twenty percent.

On the high-stakes entrance exams, the penalty was eighteen to twenty-four percent, and it took about two years to show in full. The losses were concentrated among roughly eighty percent of users whose very short homework times and very high homework scores suggested they were outsourcing the work. Students who kept spending about as long on homework as non-users saw only small losses.

There was a clear dose response. Students reporting up to one hour of AI use a week lost about five percent. Those reporting five hours or more lost about thirty percent.

A second large study points the same way, from a different setting. Researchers from the University of California, Irvine, and McGraw Hill analyzed a ten-year panel of three point two million interactions on the ALEKS math platform. They compared word problems that can be pasted into a chatbot with graphing problems that cannot easily be outsourced.

After ChatGPT's release, time spent on the copy-and-paste problems fell by twenty-six point nine percent among college students and thirty-one point three percent among high schoolers, while time on graphing problems held steady. On randomly assigned proctored retention questions, the odds of a correct answer to word problems fell twenty-five percent relative to graphing problems, even as unproctored performance rose.

The randomized evidence agrees. In a field experiment with nearly one thousand high school students in Turkey, Hamsa Bastani and colleagues found that a ChatGPT-like interface raised practice scores by forty-eight percent. But when access was removed, those students did seventeen percent worse on an exam than peers who never had it. As the authors put it, quote. Without guardrails, students attempt to use GPT-4 as a crutch during practice problem sessions, and subsequently perform worse on their own.

And the students did not notice. In that trial, those in the unrestricted group did not believe they had learned less.

The tension 5:17

The Chinese panel is not a randomized trial. Students chose when to start using AI, so some of the gap could reflect who adopts and when. It is a working paper, and the ALEKS study is a preprint that does not observe AI use directly.

There is also a debate about how the losses can be avoided. The authors' own estimate of the learning penalty fell from around twenty-five percent to around sixteen percent over the period they studied, and it is not clear whether better models or more experienced users explain the decline. Critics note that the group who used AI well was tiny. By the lead author's account, only twenty AI-adopting students ever spent seventy-five minutes on homework in a month.

And AI is not bad for learning as such. In a proctored experiment with two hundred and eleven Middlebury College undergraduates, reported in an IZA working paper, access to off-the-shelf AI while studying raised unaided test scores by zero point two seven standard deviations, and the gains held a week later. The difference lay in use. Students who used AI to explain concepts kept their gains. The short-run gains of students who had AI write for them vanished once it was removed.

Then there is detection, the response many institutions reached for first. In a blind test at the University of Reading, answers written entirely by AI were submitted into real psychology modules. Ninety-four percent went undetected, and they scored on average about half a grade boundary higher than real students.

Software detectors are no better, and they are less fair. Stanford researchers ran seven widely used detectors on ninety-one essays written by non-native English speakers. The average false-positive rate was sixty-one point two percent. When Vanderbilt University switched off Turnitin's AI detector in August 2023, it noted that a one percent false-positive rate, applied to the seventy-five thousand papers it had submitted the year before, could have meant around seven hundred and fifty student papers wrongly flagged.

The implication 7:52

The core problem is a measurement trap. A homework score, an essay drafted with help, or performance on practice problems captures what a student and a machine can do together. It says little about what the student can now do alone. A study that reports only the first number will look like a success.

That means rising grades on unsupervised work should be treated as a possible warning sign, not proof of learning. And because the penalty in the Chinese panel took up to two years to appear in full, an institution that waits for exam results to tell it something is wrong will find out late.

It also means take-home work can no longer carry the weight it used to as evidence of what a student knows. That is not the same as saying homework has no value. The evidence suggests homework still works when students keep doing the thinking.

The WAC view 8:57

In our AI in Learning Report 2026, we argue that the biggest risk of generative AI in education is not the cheating institutions catch. It is the learning that quietly does not happen when students hand the hard part of a task to a machine, often while their grades go up.

So we would separate two jobs that homework has been doing. Practice should help students learn, and it can include AI, ideally tools that withhold answers and keep students working. Assurance, the evidence of what a student can do, should come from secure, unassisted tasks, such as supervised exams or interactive orals.

That is close to what one regulator is now describing. Australia's higher education regulator, TEQSA, set out three pathways in 2025: program-wide reform with several secure assessment points across a degree, at least one secure task such as an interactive oral in every unit, or a mix of the two.

And we would stop using AI detector scores as evidence of misconduct. The error rates are high, and they fall hardest on second-language students. The money and effort are better spent on a mix of supervised assessment and open, AI-inclusive tasks.

Finally, teach students how to use AI for learning. More than eighty percent of US students told RAND that their teachers had not explicitly taught them how to use AI for schoolwork. The students who kept doing the work themselves, and used AI to explain rather than to produce, are the ones who fared best.

What to watch 10:56

One. Whether the large 2026 studies, from China and from the ALEKS platform, hold up in peer review, and whether independent researchers replicate them elsewhere.

Two. Whether study modes and answer-withholding tutors from the major AI companies are tested in independent randomized trials. In our search, we found none for a mainstream study mode as of October 2026.

Three. How quickly universities and school systems move from detection to redesigned assessment, and whether they fund the staff time and invigilation that secure tasks require.

The sources for every figure in this episode are listed on the episode page, along with the full AI in Learning Report 2026, at worldassessmentcouncil.org.

WAC Signal is produced by the World Assessment Council, with AI-assisted research and AI voices, and every claim is checked against its source before publication.

This has been WAC Signal. Evidence and ideas for the changing world of learning.