The Friday Harvest, Issue 003: Both Your Instruments Went Dark

Dean Le Blanc
College faculty member. Still grades every week.

THE ONE THING

Both of the instruments you used to know what a student did went dark in the same week.

On Monday, MIT’s Ad Hoc Committee report advised against relying on AI detectors, and noted that MIT’s own Committee on Discipline does not consider detector output alone sufficient to bring a case. That closed the first door, and most of us had already quietly stopped trusting it anyway.

On Thursday, OpenAI Economic Research published a randomized trial run with Bocconi University. More than 1,000 first year undergraduates, four groups, one real assignment. Students with ChatGPT access scored almost a full point higher on a five point rubric. Students trained in causal reasoning produced measurably more original and more distinct ideas, and scored no higher at all.

That is the second door. The grade.

A rubric is a claim about what matters. It held for a long time, because producing a clear, well structured answer used to require the understanding underneath it. This week a vendor ran the controlled experiment on its own product and published the result: the instrument sees the polish and misses the thinking.

Detection told you what a student did not do. The grade told you what they understood. Neither one is reporting reliably now, and only one of them was ever the point.

ON YOUR DESK

August 27. OpenAI Economic Research and Bocconi University published a randomized experiment with more than 1,000 first year undergraduates. Four arms: ChatGPT access, causal reasoning training, both, or neither. ChatGPT access raised the rubric score by almost a full point. The causal reasoning group produced a wider range of more distinct ideas and did not score higher.

The gain in student thinking was real, it was measured, and the grade could not see it.

Source: OpenAI

August 25. MIT released the final report of its Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training. It advises against relying on AI detectors, on the reasoning that detection invites an arms race with humanizers, and it asks every subject to post a clear generative AI policy in the syllabus and on the course site.

The enforcement option is formally closed at one of the institutions everyone else cites.

Source: MIT

August 26. OpenAI expanded ChatGPT for Teachers to 55 more school systems across 20 states, reaching over 100,000 additional educators and staff. It also announced a 16 state National Data Privacy Agreement through the Student Data Privacy Consortium framework.

Read the second half of THE TOOL before you go looking for it.

Source: OpenAI

August 27. Google’s NotebookLM added books as expert sources, so a student can now interrogate a book inside the notebook rather than read it alongside one.

Your reading list is now an input to something.

Source: The Verge

THE TOOL

ChatGPT for Teachers. If you teach in higher education, you cannot have it.

This is the announcement that will land in your inbox forwarded by somebody who did not read it closely, so here is the honest version.

On August 26 OpenAI expanded ChatGPT for Teachers to 55 additional school systems across 20 states, bringing it to more than 100,000 additional educators and staff. The program now covers more than 100 K to 12 organizations across 30 states and over 300,000 educators and staff. It ships with a 16 state National Data Privacy Agreement, administrative controls, and data that is not used to train models by default.

It is a serious product and the privacy work underneath it is the most substantive thing any vendor has done for schools this year.

Can you get it? No, and the page says so plainly. In OpenAI’s own words, ChatGPT for Teachers remains a free tool for verified U.S. K to 12 educators through June 2028. Not higher education. There is no waiting list for you, no institutional path, and no version of this where your provost negotiates you into it.

That matters more than it sounds, because it is the second time this month a headline that read like faculty news turned out to be K to 12 news. If you teach dual enrolment or early college, some of your students sit inside districts that now have it, and their teachers have a managed workspace you cannot see into.

What is actually usable from it. One number, from OpenAI’s own privacy preserving analysis of educator usage between January 1 and July 16 of this year. Educators sent more than 1.9 million messages about time saving tasks. The largest single category was report cards and progress reports at 900,000 messages. Lesson planning came second at 800,000.

Sit with the order of those two for a second.

The biggest thing teachers reach for AI to do is not build the lesson. It is write about the student afterward. That is the same job that just lost its instrument.

WORTH KNOWING

The two studies disagree about everything except the finding. Igor Chirikov’s Artificial Intelligence and Grade Inflation, UC Berkeley CSHE Working Paper 26-3, analyzed more than 500,000 grades at one large research university from 2018 to 2025. In courses with more AI exposed tasks, the share of A grades rose by 13 percentage points, about 30 percent relative to the 2022 baseline, and the increases were larger where homework carried more weight. That is observational, at scale, from a critic. The Bocconi trial is randomized, small, and run by the vendor. One is looking at half a million real grades and one is looking at four controlled arms, and they land in the same place.

The rubric in the Bocconi study was not a bad rubric. It measured whether the recommendation addressed two standard marketing goals. That is a defensible, teachable, gradeable criterion and most of us have written its equivalent. The failure was not in its quality. It was in its silence: it had no row for an idea nobody else had, so a group of students who demonstrably produced more of them registered as identical to a group that did not.

Where the primary source and the coverage diverge, use the primary source. The trade press has been writing the Berkeley figure as “about 30 percent more A grades,” which is the relative change. The paper’s own number is 13 percentage points. Both are true and they are not the same claim, and a faculty audience will check.

FROM THE KERNEL

Monday: department AI policy, and why one guideline naming three kinds of AI does more work than a ban.

Tuesday: the faculty teaching session on the course you already built.

Wednesday: MIT says stop using AI detectors, and your syllabus policy now names your own AI use too.

Thursday: the student session on the AI policy on page four of a syllabus nobody read.

Friday: the rubric rewards polish, and the one row to add before your next stack lands.

ONE THING TO TRY MONDAY

Twenty minutes, one assignment, and you can do it before your first class.

Pull up the rubric for the next major assignment you will actually grade this term. Not a model rubric, not the one in the shared drive. The one you are going to use.

Read it row by row and ask each row a single question: could a well organized, entirely conventional answer earn full marks here? For most rows the answer is yes. That is not a failure. Organization and coverage are real criteria and they should stay exactly where they are.

Now count the rows where the answer is no. Rows that require something only this particular student could have supplied. Rows that ask why this approach and not the obvious one. Rows that reward an idea the reader has not already seen twenty times in the same stack.

If that count is zero, you have found it, and you are in the large majority.

Write one row. Keep it short, keep it observable from the outside, and anchor it to something the student has to produce themselves. The alternative they rejected and why. The assumption their recommendation depends on. The condition under which their own answer would be wrong. That last one is the causal reasoning move the Bocconi training taught, and it is precisely the move their rubric had no place to record.

Then tell your students the row exists before they start. A criterion nobody knows about is a trap. A criterion announced in week one is a curriculum.

That is one row, on one assignment, and it will be better than anything a tool would have written, because you are the one who knows what a conventional answer looks like in your discipline.

Then count the rest of the term. The other assignments in this course, the other sections you teach, the rubrics you inherited from whoever taught it last, and the program review that wants alignment evidence for every one of them. Doing this once is a good afternoon. Doing it across a whole course is the part that never happens, which is why most rubrics in use this fall were written before any of this was true.

The Rubric Builder in the Faculty Toolkit does that one job. It reads your Course Record, the five minute profile of what you teach and how you sound, and drafts precise, measurable criteria in your language rather than generic ones. Review first, always.

Do the first one by hand anyway. There is a free working session that walks you through it, nothing generated and nothing saved: The Missing Row.

THE FORWARD

Forward this to the colleague who is about to spend Sunday grading against a rubric they inherited.

Not because the rubric is wrong. Because they are the only person who knows which row is missing from it, and nobody has asked them that question in years.

WATCH THIS WEEK

Harvest Kernel carries the long teaching sessions, faculty on Tuesdays and students on Thursdays.

The Daily Seed carries the short ones.