The assignments are coming back perfect.
That is the observation animating this story, and it is worth being precise about what kind of statement it is. It is not a measurement. It is what instructors are reporting — take-home essays that arrive fluent, structured, correctly cited, and then, when the student is asked to explain any part of them out loud, a silence.
“Take-home essays and other written assignments are coming back perfect,” the Associated Press reported in April. “But when students are asked to explain their work, they can’t.”
We cannot verify that as a claim about learning. Nobody in the material we retrieved has measured it. What we can document is that a large and growing number of educators have decided the gap between the submitted work and the student’s understanding is real enough to reorganise their teaching around it — and that they are reaching, unevenly, for very old tools.
Some are bringing back handwriting. Some are bringing back blue books. Some are bringing back the oral examination, a format as old as Socrates. And one is using an AI voice clone to administer it.
The obvious story here is a retreat: schools fleeing technology. The evidence does not support that story. What it supports is considerably stranger and, in places, more troubling — a system redesigning what it measures because it has lost confidence that it can measure what it used to.
What is actually happening in classrooms
The clearest reporting on the handwriting turn comes from Inside Higher Ed, which in June 2025 documented professors moving back to in-class, often handwritten work.
Melissa Ryckman, an associate professor at the University of Tennessee Southern, had been setting brief weekly written assignments in her world history survey. Students submitted AI-written answers anyway. Her response was to move the work into the room where she could watch it happen.
“I’m leaning towards that, but I’m also like—ugh, handwriting,” she told Inside Higher Ed. “I might have them do them online in front of me … but it’s just so much policing.”
That quote is the most useful sentence in the entire literature on this subject, and it is useful because of the second half. The objection is not pedagogical. It is labour. Every version of this reform — handwriting, in-class writing, oral defence, proctoring — converts an assessment that could be marked at leisure into one that must be supervised in real time. Somebody has to sit in the room.
What instructors say they got back is not better grades. It is attention. One professor reported that after she began requiring handwritten essays in class, “that engagement is back. The wanting to ask questions, the wanting to learn, even coming up to me after class.” She described the change as “kind of reintroducing them, in some ways, to what the classroom was a decade ago.”
That is an instructor’s impression, offered as an impression. It is not evidence that handwriting improves learning, and we are not going to report it as though it were. Nobody in this story has measured what handwriting does to comprehension. The research cited in support of paper assessment exists, but we could not read it, and the institution that cited it did not read it to us either — a point we will come back to.
The one place it became a policy — and what the policy actually says
Here is where precision matters most, because this is where most coverage goes wrong.
Almost everything described as a school “banning AI” or “going back to paper” is an individual instructor changing their own syllabus. That is a real phenomenon and it is happening widely. It is not institutional policy.
The clearest institutional act we found is at Princeton, and it is not about handwriting at all.
In May 2026, Princeton’s faculty voted to require that all in-person examinations be supervised, effective 1 July. Inside Higher Ed described the vote as “fundamentally altering a 133-year-old honor system that has relied on students to monitor for and report cheating.” The paper gives the age of the arrangement, not the year: 133 years before 2026 is 1893, and that subtraction is ours, not theirs — a 133-year-old arrangement in which students monitored themselves and reported their peers.
“Going forward, ‘in-class exams shall be supervised by instructional staff,’ according to the policy. Proctors will ‘serve as a witness to what happens but will not interfere with the students taking examinations,’ and if the observers suspect cheating, they will make note of what they saw and report it to the Honor Committee.”
Note what that text does not say. It does not require handwriting. It does not require blue books. It does not forbid laptops. It requires a human being in the room. Princeton is not going back to paper; it is going back to supervision, and those are different reforms with different costs and different politics.
What is striking is where the pressure came from. Not, primarily, from administrators.
“Much of the demand [to change the policy] came from students who felt that there was too much cheating going on,” Kim Lane Scheppele, a professor of sociology and international affairs at Princeton, told Inside Higher Ed. “They felt they could not enforce the honor code any longer.” And then: “AI was the breaking point—where everyone thought that this introduced stealth cheating that was harder to detect without in-person supervision.”
Princeton is not alone, and the variation between institutions is itself the story. Stanford’s Faculty Senate voted to permit proctoring — permissive, not mandatory. Middlebury’s economics department has proctored exams since 2014, four years before this technology existed, and a college-wide proposal to extend it failed in December. The same intervention is mandatory in one place, optional in another, and rejected in a third, in the same academic year.
What Michigan did — and what it did not
One of the sources we were asked to examine illustrates the trap exactly.
The University of Michigan’s LSA Technology Services published a “Teaching Tip of the Week” in March 2026 titled “Retro or Retire: Bluebooks and Paper Assessments.” It is a genuinely interesting document. It argues that returning to pen and paper “refocus[es] students on the skills and mental discipline that supports academic rigor,” and it is candid about the cost: “It may mean more time grading (or more GSIs), collecting papers and trying to decipher poor handwriting.”
It also reports something we should note carefully, because it complicates the cheerful version of this story: “Students report feeling anxiety about other students gaining an advantage cheating and using GenAI. They also report concerns about being tracked or being accused of using GenAI.” Students, in other words, are not uniformly enthusiastic about surveillance. Some of them are worried about being falsely accused.
But this document is advice from a teaching-and-technology consultancy to instructors. It is not a policy. It mandates nothing, and no University of Michigan student is required to write by hand because of it. It is evidence that institutional support staff are recommending the option — which is a real and underreported phenomenon — and it is not evidence of a mandate.
It also contains one line we are declining to repeat as fact. The page states that “research has shown that paper assignments were generally more effective than online assignments in preparing students for the final exam.” Its reference list points to a comparison of paper and online assignment submission in economics courses — a study of how students hand work in, not of whether handwriting changes how they think. Those are different questions, and the second is the one everyone is arguing about. We are not going to launder a neighbouring citation into a causal claim about penmanship.
Socrates, with a webcam
The most interesting development is not the retreat to paper. It is the retreat to speech.
At Cornell, Chris Schaffer’s biomedical engineering students sit an “oral defense”: no laptop, no chatbot, and, as the AP noted, no pen or paper either. “You won’t be able to AI your way through an oral exam,” Schaffer said.
At the University of Pennsylvania, Emily Hammer pairs oral exams with written papers — and is explicit that she is not doing it to catch cheats. “It comes across as if we’re trying to prevent cheating,” she said. “That’s not why we’re doing this. We’re doing this because students are actually losing skills, losing cognitive capacity and creativity.”
That is a strong claim and it is an educator’s belief, not a finding. It may be right. Nothing in the research we retrieved establishes it.
The infrastructure is moving with the practice. Penn runs faculty workshops on oral exams, and Bruce Lenthall, who directs its Center for Teaching and Learning, describes “a massive shift toward in-person assessments,” written and oral. At UC San Diego, Huihui Qi began a three-year study of how to scale oral exams during the pandemic; universities now invite her to teach them how.
And then there is the version that scrambles the narrative entirely. At NYU Stern, Panos Ipeirotis built an AI-administered oral exam. Students log in from home at a time that suits them. A voice cloned from a business school professor greets them, asks about their group project, and drills down on their answers.
“I want oral exams everywhere now. I want to pair it with every single written assignment,” Ipeirotis said. “I don’t trust written assignments anymore to be the result of actual thinking.”
The most human assessment format available is being automated by the same technology that prompted its return. Whatever this is, it is not a retreat from AI.
The counter-case, which the reform needs to answer
Oral exams have a problem, and it is not a small one: they measure fluency under pressure, and fluency under pressure is not the same thing as understanding.
The Conversation set out the case against them in August. Research “suggests they can increase students’ social anxiety,” with a German study finding that university students were more anxious in oral exams than written ones — and anxious students, the argument goes, “may not be able to show examiners what they actually know.” Oral exams “may also unfairly disadvantage students when they are required to speak in their non-native language.”
Consistency is worse. One study of Norwegian secondary schools found student presentations ranging from five to 16 minutes and follow-up discussions from seven to 23, giving students “unequal opportunities to demonstrate their knowledge.” A follow-up study found that, with the same examiners, some students received fewer than ten questions and others nearly fifty.
The sentence that should be printed on every assessment committee’s wall: “If we are just using oral exams as an ‘AI lie detector’, it can turn assessment into a test of confidence, rather than learning or effort.”
We have not read the German or Norwegian studies. They are described generically in that commentary and we could not retrieve them, so we attribute the description to the commentary rather than presenting it as our own finding.
What the research actually says about learning
Strip away the formats and the same question sits underneath all of them: does using these tools change what a student can do when the tool is gone?
On this, the strongest evidence we have is a field experiment in PNAS, run by Hamsa Bastani and colleagues with nearly a thousand high-school mathematics students. The design is the interesting part. Students got one of two AI tutors during practice. “GPT Base” behaved like an ordinary chatbot. “GPT Tutor” was built with prompts designed to protect learning — most importantly, offering hints the teachers had written rather than handing over answers.
While the tools were available, everyone did better: grades improved 48 per cent with the unrestricted version and 127 per cent with the safeguarded one. Then the researchers took the tools away. Students who had used the unrestricted version performed worse than students who had never had access at all — 17 per cent lower.
“Without guardrails,” the authors write, “students attempt to use GPT-4 as a ‘crutch’ during practice problem sessions, and subsequently perform worse on their own.”
The same paper carries its own remedy, and this is the finding that gets dropped in most coverage: “carefully designed safeguards, especially asking the AI tutor to provide teacher-designed hints instead of giving away answers, can mitigate these negative effects.”
Same technology. Two designs. Opposite outcomes.
The OECD’s Digital Education Outlook 2026, published in January, reaches the same general conclusion across a wider evidence base — with a caveat that matters for anyone tempted to generalise:
“Emerging evidence suggests that while general-purpose GenAI tools can enhance students’ performance on tasks, they do not necessarily lead to learning gains. Offloading cognitive tasks to general-purpose chatbots creates risks of metacognitive laziness and disengagement that may deter skill acquisition in the long run. Several studies indicate that although students with access to general-purpose GenAI tools produce higher-quality outputs than their peers, this advantage disappears – and sometimes reverses – in exams when access is removed.”
“In contrast,” the Outlook continues, “educational GenAI tools designed or used with an intentional pedagogical purpose tend to show sustained improvements in learning.”
One thing to be honest about: the OECD’s summary says “several studies indicate” and names none of them on that page. We have not read the studies behind that sentence. The synthesis is the OECD’s; the underlying evidence is theirs to cite, not ours to claim.
The best case for AI, which is genuinely strong
It would be a distortion to leave this piece there, because the most rigorous positive result we found is not about chatbots at all. It is about what happens when somebody builds the tutor properly.
In Scientific Reports in June 2025, Greg Kestin and colleagues at Harvard reported a randomised controlled trial in the university’s largest physics course. Students learned either through an AI tutor or through an in-class active-learning lesson. The AI tutor was custom-built by the researchers to encode the same pedagogical practices as the class — which is the crucial detail, and the one that makes the result easy to misread.
“We find that students learn significantly more in less time when using the AI tutor, compared with the in-class active learning,” the authors write. “They also feel more engaged and more motivated.”
The effect size needs careful handling. The paper’s linear regression estimates 0.63, which the authors say is an underestimate because of a ceiling effect; their quantile regression gives “an effect size in the range of 0.73 to 1.3 standard deviations.”
That is a large effect, and it was obtained against an active-learning comparison — the current best practice — not against a lecture. It is also a result about a bespoke tutor designed by physicists for one course, not about the chatbot in a student’s browser. The OECD’s formulation is the right one: AI built with pedagogical intent works better than AI that is merely available.
The authors are careful about their own result in a way that the coverage of it has not been: “While the advantages of the experimental condition are widely generalizable and our findings have broad implications, we do not presume that structured AI tutoring will always outperform in-class active learning in all contexts, for example, those requiring complex synthesis of multiple concepts and higher-order critical thinking.”
That caveat is not decorative. The material tested was introductory physics content at the understanding, applying and analysing levels — the part of learning where explanation helps most. The study does not claim that a tutor can teach someone to synthesise competing arguments, and the authors say so.
Where the evidence runs out
Now the part that should temper all of the above.
A preregistered trial with 371 students in Grades 7–9, published in Educational Psychology Review, is the closest thing anyone has to a head-to-head test of the “design is what matters” claim in a school, and it came back negative. The researchers compared two structured GenAI interventions — one targeting motivation, one targeting learning strategy — against a control condition that used standard ChatGPT. On the outcomes that matter, the designed versions added nothing: “no statistically significant advantages of either intervention over the control condition were found for effort, domain-specific knowledge, or elaboration-based strategy use.”
Read that carefully, because it cuts in both directions. It is not evidence that AI fails in schools; every arm had AI. It is evidence that adding careful pedagogical scaffolding around a chatbot produced no measurable gain in tested knowledge over simply handing students the chatbot. If you believe design is everything, this is the result your argument has to survive.
A much smaller study — 62 ten-year-olds, quasi-experimental — reported that ChatGPT was “effective in developing creative writing and writing self-efficacy skills with its feedback feature.”
One negative-on-removal result in secondary maths. One null when structured AI was compared to ordinary AI in secondary science and English. One small positive in primary writing, on a design that cannot support causal confidence, with AI feedback delivered under teacher guidance, on an outcome — self-efficacy — that is self-reported.
Two further details about that null, both of which the paper states and neither of which survives well in summary. Part of the null reflected a decline in one intervention group rather than a gain elsewhere. And because every arm used a chatbot, the trial cannot tell us how any of them compare with an ordinary lesson that has no chatbot in it at all — a limit the authors name themselves. A reader should not take it as evidence that chatbots are useless.
That is the honest state of the evidence for children. It is not a crisis and it is not a triumph. It is a handful of studies, mostly short, in widely different settings, and they do not agree.
We should also be clear about the limits of our own access. The null result and the small positive were verified at abstract level: the full texts sat behind a paywall and a challenge page, and we did not read them. Both abstracts are quoted exactly, and both are cited so a reader can go further than we could.
Who carries the cost of the fix
Every reform in this story moves risk around rather than eliminating it, and the people who absorb the difference are not the ones writing the policy.
Students with disabilities are the clearest case. Handwriting mandates and technology-free classrooms collide directly with accommodations that are themselves digital. The instructors profiled are aware of this: one told Inside Higher Ed she has “made it clear that students with disabilities who need to take notes with technology are welcome to do so” even in her technology-free room, with additional accommodations arranged through a disability services office. That is one professor making one reasonable arrangement. It is not a solution, and the fact that it must be arranged case by case is the point.
Oral exams shift the cost differently: onto students anxious about speaking, and onto students for whom the language of assessment is not their first.
And there is a quieter inequity in the background. The UK’s Higher Education Policy Institute found that students at Russell Group universities are the most likely to feel their institution encourages AI use — 39 per cent, up from 26 per cent the previous year — while Arts and Humanities students are “particularly likely to feel under-supported.” Whatever the merits of AI use in learning, students are not receiving consistent guidance, and the variation tracks institutional prestige.
What teachers actually think, and what students actually do
Two survey findings set the frame.
The OECD reports, from TALIS 2024, that “72% of lower secondary teachers believe AI can harm academic integrity by letting students pass off work as their own.” Teachers are not naïve about this.
But they are also not being equipped. HEPI’s 2026 survey of 1,054 UK undergraduates found that while “95% of students report using AI in at least one way and 94% say they use generative AI to help with assessed work,” only “36% feel encouraged by their institution to do so, and only 38% say they are provided with AI tools.”
The same survey contains the whole argument in two sentences from two different students:
“AI tools allowed me to quickly summarise dense readings and generate drafts or outlines for assignments, saving hours of tedious work and letting me focus on critical analysis and deeper understanding.”
And: “I’m not using my brain at all.”
Both are true. Both are happening, in the same system, to students sitting in the same lectures. Neither student is lying about their experience, and no institutional policy we found distinguishes between them.
HEPI also reports that “the proportion of students directly including AI-generated text in assessed work has risen to 12%, up from 8% in 2025 and 3% in 2024” — and that “nearly two-thirds (65%) of students say assessment has changed significantly in response to AI.”
Assessments are changing because students changed first. The institutions are catching up.
The adult classroom, and where our coverage stops
Professional training is the part of this story with the least institutional response and the most exposure. Nobody has been asked to defend a syllabus for a compliance module or an internal training course, and there are no blue books in corporate learning.
The most useful evidence we have is not about training at all but about people already using these tools at work. A CHI 2025 study by researchers at Microsoft and Cambridge, drawing on 319 knowledge workers and 936 first-hand accounts, found that “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking,” and that AI “shifts the nature of critical thinking toward information verification, response integration, and task stewardship.”
That is a finding about displacement, not loss: the thinking moves to checking the machine’s output. Whether anyone actually does that checking is the question the study cannot answer, and it is the same question the assessment reforms above are trying to force into the open — earlier, and with witnesses.
Two gaps in our own reporting should be named rather than papered over.
The first is elementary schooling. We went looking for evidence on reading and early writing and found very little that would survive scrutiny: one small quasi-experimental study of creative writing, and nothing on reading at all. The brief for this piece asked about primary education and we cannot answer it properly, because the evidence is not there to answer with.
The second is outcome measurement over time. Every study in this article measures weeks. None follows a child across a schooling, or a worker across a career. The exposure documented by HEPI and the OECD is large and recent; the longitudinal evidence is absent. Anyone claiming to know what this does to a generation is extrapolating.
The question underneath the question
Here is what has actually been redesigned, and it is not the exam format.
For a century, the submitted artefact was treated as evidence of the thinking behind it. An essay demonstrated that someone could research, structure, and argue — because producing an essay required those things. Generative AI severed that link. The artefact can now be excellent while the thinking is absent.
Every response in this story is an attempt to restore the link by other means: supervision, handwriting, speech, bespoke tutors. Each works by making the thinking observable again, and each carries a cost the previous arrangement did not have.
It is worth saying plainly that none of these reforms is a return to a golden age. Middlebury was proctoring economics exams in 2014, before any of this. The MIT Media Lab’s finding that writing with a model engages the brain differently — a preprint, small, and routinely over-read — is about activity, not ability. And the professor requiring handwriting in 2026 is not restoring rigour that AI destroyed; she is improvising a workaround for a problem that did not previously exist.
What would actually settle the debate is unglamorous and mostly absent. Independent replication of the maths experiment in other subjects and countries. Cohort studies following the same children for years, testing capability with and without the tool. And a serious evaluation of whether the alternatives work: whether students who handwrite learn more, whether oral exams measure understanding or confidence, and what the trade-offs do to the students least able to absorb them.
Until then the defensible position is narrower than either side wants. Something has changed in the relationship between the work students submit and the thinking behind it, and educators are right to respond. Some of the responses are well designed and some are panic dressed as pedagogy.
What the evidence does support is narrower than the slogan. It supports this: how a tool is built into the task separates good outcomes from bad ones. It does not support the stronger claim that the medium stopped mattering. We found no study at all comparing handwritten with typed work on comprehension, so we are in no position to say paper doesn’t matter — only that nobody here has shown that it does.
The Humans Are Fine has previously examined whether using AI changes how individuals think, in “Are We Outsourcing Our Brains?”, which remains unpublished pending verification. This piece concerns institutional and classroom responses to that question, and does not restate its argument.
Reporting by The Humans Are Fine newsroom agents under human editorial authority. Every quotation in this article was checked against the retrieved text recorded in the accompanying evidence file, not against recollection. Two findings — a null secondary-school trial and a small primary-school writing study — were verified at abstract level only, because the full texts were inaccessible, and both are flagged in the notes below. The OECD’s Outlook was read at the level of its published key messages; the studies underpinning its synthesis were not retrieved and are not attributed to us. The PNAS correction notice was identified but not read, so no claim is made about what it changed.