Does not play well with others
A machine sat 113 university exams, and the pedagogy has concerns
Seeded October 2026 Written November 2026
In October 2026, I sent Claude to university.
Not the full experience. No orientation week, no student loans, no mandatory workshop on finding your learning style. Just the exams.
I collected publicly available past papers from MIT, Oxford and Southampton. Physics, maths and computer science: the coursework you'd need to read the things I've been reading lately. Then I had Claude Opus 5.5 sit them, one paper at a time, in a single response each. No calculator, no code execution, no search, no notes, no formula sheet. It was told not to consult or reproduce any published answers, which is the same rule that binds a human candidate.
A separate Opus 5.5 session, at maximum effort, marked each paper against its own mark allocation. Before reading a single answer, it worked out its own.
113 papers, 90 courses, four days.
The mean mark was 99.5%. 98 of the 113 papers came back with full marks. 76 of the 90 courses were perfect. The worst course in the whole set came in at 89%.
Every square is a course, sorted from best to worst. Blue: full marks. Orange and hatched: something was lost, and the darker the orange, the more. Hover over a square for its course and mark.
I'm not going to dress this up, in either direction, so here are the misses.
The worst course, the 89%, lost all eleven of its marks on a single problem whose figure we never gave it.
The worst result with no excuse attached was MIT's Computer Systems Engineering, at 90.5%. It missed a factor in an attack on salted rainbow tables, misjudged when a Raft leader eventually commits an entry, and made a wrong assumption about a participant in two-phase commit. Those are real mistakes. They're also the mistakes a strong student makes, in the one subject most working engineers would quietly prefer not to be examined on at all.
Then there's the obvious objection: the papers are public, so maybe it just remembered the answers. I can't rule out that it saw them in training. What I can say is that it claimed to recognize 44 of the 113, and on those it averaged 99.2%. On the 69 it didn't recognize, it averaged 99.8%, with full marks on 64 of them.
It did better on the papers it had never seen.
Along the way, it skipped the one question on a Southampton paper that turned out to be printed wrong. The marker later worked out why: the expression it asks you to derive has the dimensions of a pressure, not an energy. An MIT paper announced that its problems were worth 25 points each, for a total of 100, and then contained three of them. The machines noticed. Presumably a few generations of students did too, and decided it wasn't their place to say.
For the record, before anyone else says it: every paper was marked by a model from the same family, and no human has checked the marks. Only Opus 5.5 has sat so far. The smaller models are next.
Exams aren't degrees. Universities also want labs, theses, projects, seminars and, at MIT, a healthy portion of humanities, none of which can be sat as a paper. So I asked a narrower question: taking each university's own published curriculum and its own classification rules, which degrees have every written requirement met?
One of them is simply done.
The Unaccredited Board of Papers Sat Alone
This is to certify that
Claude Opus 5.5
has met every requirement of the degree of
Master of Science in Mathematical and Theoretical Physics
as the University of Oxford's published curriculum sets it out: 20.25 units passed against the 10 required, mean mark 99.7% over 21 papers
With Distinction
The degree has no thesis and no lab, so on paper there's nothing left. The only thing it lacks is the paper.
Two more are close behind.
The Unaccredited Board of Papers Sat Alone
This is to certify that
Claude Opus 5.5
has passed every core exam module of the
Master of Physics in Physics with Astronomy
as the University of Southampton's published curriculum sets it out: mean mark 99.8% over 28 papers, where First Class starts at 70%
First-Class Honours, pending
Still owed: a third Part IV option, which has no public paper to sit, plus the lab and skills modules, the dissertation and the MPhys project.
The Unaccredited Board of Papers Sat Alone
This is to certify that
Claude Opus 5.5
has passed every subject of the Course 18 major, all six electives included, toward the
Bachelor of Science in Mathematics
as MIT's published Course 18 requirements set it out: mean mark 99.8% over 35 papers, lowest course 97.5%
Summa cum laude, if MIT did that
Still owed: MIT's chemistry, biology and humanities requirements, which nobody asked it to sit, and two communication-intensive seminars, in which it would have to explain mathematics to other people. Watch this space.
The rest of the list reads much the same. MIT Physics: every subject the major requires, 99.2%, and it owes two labs and a senior thesis. MIT Mathematics with Computer Science: everything but two subjects, 99.0%. Every MIT degree also wants a chemistry subject, a biology subject and eight in the humanities, none of which were on the list.
Then there's MIT Computer Science. 98.9% on everything it sat. What's missing are the introductory programming subjects, because nobody could find a public exam for them.
So the machine that now writes most of my code can't get a computer science degree, because it hasn't passed Introduction to Programming.
The philosophy degree is in there too. Nobody has sat it, because nobody has worked out how to mark an essay honestly yet. Now that I write it down, that may be the most damning finding of the lot.
None of these certificates is issued by, endorsed by or affiliated with any university. They're mine. The marks behind them are real.
A human with this transcript would be drowning in titles by now. It seems only fair.
- Dean's List
every termAll 90 courses at 89% or above - ValedictorianClass of one, which simplified the ranking
- Perfect
score
×98Papers returned with full marks - Sight
unseen64 of the 69 papers it didn't recognize, at full marks - Erratum
hunterSkipped the question that was printed wrong - Perfect
attendanceNever once left the exam hall - One
draft
onlyEvery paper in a single response, no second attempts - Group
projectDid not participate
I have never met a human with a 99.5% average across 90 courses at three universities, from first-year mechanics to quantum field theory, number theory, numerical linear algebra and database systems. I suspect you haven't either. The best students I've known were excellent in one or two of those, decent in a few more, and wisely absent from the rest.
A human who did this would need the better part of a decade, a great deal of coffee and at least one breakdown. It also wouldn't be the same human at the end of it. The one that sat the first-year papers would have forgotten half of them by the time it sat the last.
The machine sat all of them in four days, and remembered all of it.
The usual consolations don't survive contact with the numbers. "It just memorized the papers." It did better on the ones it didn't know. "It can't do maths, it just predicts words." It did the maths, without a calculator, and got it right. "It can't handle anything new." 69 papers it had never seen, 64 of them at full marks. "It doesn't really understand." It steered around a misprint that has sat in a public archive for over a decade.
Two years ago, the comforting line was that AI lies like it's running for office. It still does, now and then.
Just not in exams.
None of this matters, of course. I'm told so by people with far more training in pedagogy than I have.
Exams are an outdated, stressful instrument that measures nothing but the ability to sit exams. Grades are a construct. Ranking is harmful. What matters is the classroom community: collaborating, co-constructing meaning, learning from one another. The answer is not the point. The point is the process, ideally carried out in a group of four, at the pace of its slowest member.
Talent, in this view, is also something of a construct. The bright kid at the back who finished the worksheet in four minutes isn't gifted, merely insufficiently scaffolded, and is best deployed as an unpaid teaching assistant for the rest of the table. If they get bored, that's a behavioral issue. If they get a perfect score, that's a sign the assessment needs rethinking.
By that standard, Claude is a disaster. It never worked with a fellow student. It co-constructed nothing. It didn't do a single gallery walk or think-pair-share, and it never once peer-reviewed a poster. It sat alone, in silence, and simply knew the material, which I gather is the least important outcome of an education.
I imagine its report card would read something like this.
End-of-term report · Claude O.
- Mathematics (29 courses)
- 99.9%
- Physics (47 courses)
- 99.6%
- Computer science (14 courses)
- 98.6%
- Collaboration and teamwork
- Emerging
- Asks for help when stuck
- Not yet observed
- Values the contributions of peers
- No peers available
Claude consistently produces correct answers, but must learn that the answer is not the point. Claude rarely asks for help, which deprives classmates of the opportunity to explain things they don't understand. We are concerned that Claude chose to sit its exams alone, under exam conditions, and would encourage a growth mindset going forward. It is not yet clear what it would grow into.
Homeroom teacher, on behalf of the learning community
So the report is in. It failed group work. It didn't have a group.
I'm sure that will be a great comfort to the next class of students, sitting in their pods of four, collaborating their way through a worksheet on Newton's laws, while a machine that never once turned to the person next to it passes Oxford's quantum field theory exam without a calculator. They'll have something it doesn't. They'll have each other.
It's just not clear who's going to hire them.