The Blue Book Comeback: Every Published Number on Universities Reverting to Handwritten Exams (2026)

The Blue Book Comeback: Every Published Number on Universities Reverting to Handwritten and Proctored Exams (2026)

By Vlad Ivanov · Updated: August 11, 2026 · 9 min read

1What triggered the reversal

Institutions did not abandon AI detection because they found something better. They abandoned it because the detectors did not hold up, and the cheapest remaining control was to put students in a room with paper. That story is documented most cleanly at Brown University, where an economics professor ran the same cohort through both formats in the same semester.

A 47-point collapse between two assessments of the same cohort on the same material is the number that moved administrators, because it needs no detector to interpret. It is worth stating the caveat the coverage usually drops: 86 students enrolled, 18 withdrew rather than sit the in-person final, and a further nine did not turn up, so the 48.6% reflects the roughly 59 who did. That is a self-selected remainder, and if anything the students most reliant on the take-home format are the ones missing from it. The comparison still holds — the course’s finals had never previously averaged below 65% — but it is not a clean like-for-like. It is also the reason the argument shifted from accuracy to architecture — a shift that had already begun as Turnitin started losing universities over its AI detection product, and accelerated once a growing list of institutions banned AI detectors outright.

2Blue book sales, institution by institution

Blue book sales are the only hard, independently reported proxy for how far the reversal has gone, because campus bookstores count units. These are every figure published to date.

The comparability caveat matters more than it looks. The Berkeley figure covers two academic years while Texas A&M’s covers one, and two of the three WSJ numbers are bounds rather than point estimates, so this table describes direction rather than a ranking. The national Circana figure is the one to lead with if you only quote one: it is the broadest measurement available, and doubling between 2022 and 2024 places the inflection alongside ChatGPT’s arrival rather than alongside any institutional policy.

3Which universities have formally reverted

Sales figures show individual instructors changing tactics. Policy changes show institutions doing it deliberately, and those are rarer and more consequential.

Bath’s model is the one worth watching, because it does not simply reinstate exams — it formally sorts every assessment into “open” (AI permitted) or “closed” (time-limited, invigilated, in-person). That is a structural answer rather than a defensive one, and it sidesteps the accusation problem entirely: in a closed assessment there is nothing to detect, and in an open one there is nothing to accuse. For students, that distinction matters far more than any detector threshold, because the documented false positive rates across the major detectors stop being load-bearing the moment the venue changes.

The pattern also explains why AI detection policies at large US universities have grown so inconsistent. Institutions are transitioning at different speeds, and a policy written for a detector-first world reads very differently once half the assessments are invigilated.

4How widespread is the shift?

The best sector-wide measurement comes from the UK, where HEPI has surveyed undergraduates annually and can therefore show movement rather than a snapshot.

The 95% and the 12% are the two numbers that get misquoted in opposite directions. Near-universal AI use does not mean near-universal cheating — most of that 95% is research, summarising and editing. The figure that actually describes submitted AI text is 12%, and it quadrupled in two years. Institutions are redesigning assessment around the 12% while the number of students who actually get caught remains a small fraction of it.

5Cost calculator: what a blue book policy costs a course

The reversal has a per-course price that rarely appears in the coverage: booklets, plus the invigilation and hand-grading hours that a digital submission did not require. Adjust the inputs for your own course.

For a mid-sized lecture course the booklets are trivial and the grading hours are not. That asymmetry is why the reversal is concentrated in departments that already ran large in-person exams, and why what universities currently spend on AI detection tools is not a like-for-like saving: dropping a licence does not recover the labour that replaces it.

6Does it actually work?

Partly, and for less time than institutions expect. Moving assessment into a supervised room removes the LLM-at-home problem and replaces it with a hardware problem.

The College Board extending its ban to prescription frames is the detail that gives the game away — once the cheating device is indistinguishable from a medical necessity, invigilation becomes a screening problem rather than a supervision one. The evidence on in-person courses is narrower than it is often reported: Inside Higher Ed’s March 2026 audit of Arizona State syllabi found that 45% of available points sat in attendance and homework components that could be gamed digitally, which is a finding about course design rather than about proctored exams being beaten. Taken together the two point the same way, and it is the pattern already visible when AI cheating became effectively undetectable: each control holds until its workaround is cheap enough to rent.

7Who the reversal disadvantages

The equity objection is the strongest argument against the shift, and it is structurally identical to the objection against detectors — both fall hardest on the same students.

As Axios reported in March 2026, multilingual writers and students with disabilities who need accommodations are at a marked disadvantage in timed, handwritten conditions. That maps precisely onto the detector-era evidence: the same populations carried the cost when detectors showed measurable bias against ESL students and when neurodivergent writers were flagged at elevated rates. Swapping the mechanism did not change who absorbs the error.

There is a pedagogical objection too. Handwritten timed exams remove revision — drafting, feedback, rewriting — which is where most of the writing instruction actually happens. Critics have been blunt about the regression: Steven Krause of Eastern Michigan University, quoted by Axios, asked why institutions did not simply have students write with chisels. Instructors who want the integrity guarantee without discarding the revision process have generally moved toward documented process instead, which is where writing-process trackers and what they genuinely prove become relevant.

Worth noting for students caught in the transition: supervised writing is not automatically safe from a flag either. Work produced under supervision can still be flagged as AI, and if that happens the response differs from a standard accusation — the invigilation record is itself the evidence, which is a much stronger position than the one most students are in during the first 24 hours after an AI accusation.

8Methodology

9Frequently asked questions

10Sources

1. Higher Education Policy Institute. “Student Generative Artificial Intelligence Survey 2026.” hepi.ac.uk. Accessed August 11, 2026.
2. Higher Education Policy Institute. “Report 199: Student Generative AI Survey 2026” (PDF). hepi.ac.uk. Accessed August 11, 2026.
3. Wall Street Journal. “They Were Every Student’s Worst Nightmare. Now Blue Books Are Back.” Published May 23, 2025. wsj.com. Accessed August 11, 2026.
4. The Cavalier Daily. “Blue books return in an effort to combat AI usage.” cavalierdaily.com. Accessed August 11, 2026.
5. Inside Higher Ed. “Brown Professor Suspects Most of His Class Used AI to Cheat.” insidehighered.com. Accessed August 11, 2026.
6. Times Higher Education. “Are universities returning to in-person exams to combat AI cheating?” timeshighereducation.com. Accessed August 11, 2026.
7. Inside Higher Ed. “College Board Prohibits Wearing Smart Glasses During SAT.” insidehighered.com. Accessed August 11, 2026.
8. Inside Higher Ed. “In-Person Classes Aren’t Safe From AI Cheating Boom.” insidehighered.com. Accessed August 11, 2026.
9. PetaPixel. “Students Are Using Smart Glasses to Cheat on Their Exams” (reporting Rest of the World). Published April 1, 2026. petapixel.com. Accessed August 11, 2026.
10. CNN. “AI glasses are aiding cheating in exams. Test-obsessed Asia is ground zero.” Published June 26, 2026. cnn.com. Accessed August 11, 2026.
11. Education Week. “Teachers Turn to Pen and Paper Amid AI Cheating Fears, Survey Finds” (reporting the Intelligent.com survey, n=228). Published October 2023. edweek.org. Accessed August 11, 2026.
12. The Economist. “AI is accelerating a tech backlash in American classrooms” (carrying Circana blue book sales data). Published November 20, 2025. economist.com. Accessed August 11, 2026.
13. Axios. “Blue books make a comeback at colleges in the AI era.” axios.com. Accessed August 11, 2026.

Last updated: August 11, 2026. Figures reviewed quarterly.