Everyone quoted the revenue number in OpenAI's August 31 announcement. The line that changes your assignment is in the last section: the company says it will explore more native ways for businesses to interact with people inside ChatGPT.

Native is a term of art. It means the commercial content stops looking like a separate block.

Which puts the problem somewhere unexpected. Not in your AI policy. In your assignment prompt.

A prompt that asks which of three approaches is best accepts a recommendation. A recommendation is a product. Nobody, including you, can tell how it was shaped.

The fix is three added requirements, all gradeable. The alternative they rejected and the exact criterion that ruled it out. The criteria written before the conclusion, in their own words. And the sentence where they changed their mind.

Full argument: https://www.harvestkernel.com/chatgpt-ads-assignment-decision-path/

What is the one assignment in your course that would still be satisfied by a confident answer with no reasoning attached?

It is week two on most fall calendars, which means add and drop, which means your inbox stopped being under control somewhere around Friday.

Today's session names something I do not think we have ever named properly.

Writing the reply is not the expensive part. You know your own policy. You know the date. Once you understand what a student is asking, most replies take under two minutes.

The expensive part is the reading. Once through and you get nothing. Again, more slowly. Sometimes you open the assignment to work out which one they mean. Only on the second or third pass do you start writing.

That is the reread tax. It lands on nearly every unclear message and it appears on no workload document in higher education.

The whole method is three moves. Name the shape of the email. Write in one line what it is really asking, in your own words. Answer that, plus the thing they were going to ask next.

The free working session takes the ten oldest messages in your inbox and runs all three on them. About twenty five minutes, nothing saved, nothing signed up for.

https://www.harvestkernel.com/the-reread-tax-audit/

One question, and one line is a complete answer.

What shape does your inbox actually take this week? Permission asks, status checks, deadline negotiations, lost students, or logistics?

1,655 fabricated author records. Real DOIs. Backdated. Sitting on a CERN operated repository right now.

The uncomfortable part is not that they exist. It is that they pass the source check most of us hand to students: does it have a DOI, is it indexed, does the author have other work, does the author have collaborators.

All four come back green on a record no human wrote.

The five questions that still work are all about what is outside the record. Does the venue exist independently. Does the author exist off the aggregators. Does the timeline hold. Has anyone independent engaged with it. Can you get the artifact behind the claim.

One question for you, and a single sentence is a complete answer:

What is the last source you assigned that you have never personally opened the full text of?

Full piece: https://www.harvestkernel.com/fabricated-authors-real-dois-source-check/

A rubric is a claim about what matters.

Every row says the same thing underneath: this is something I can see from the outside, and seeing it tells me something happened on the inside.

That claim held for a long time. Producing a clear, well structured, logically coherent answer used to require the understanding underneath it. Strong enough correlation to grade on.

OpenAI and Bocconi published a randomized experiment yesterday that tested it directly. More than 1,000 undergraduates, four groups, one five point rubric plus separate automated text analysis.

ChatGPT access raised the rubric score by almost a full point.

Causal reasoning training raised the originality and variety of ideas, measurably, and raised the rubric score by nothing.

The rubric asked whether the recommendation addressed two standard goals. It never asked whether the idea was one nobody else had. So a documented improvement in student thinking produced no change in the grade, because the grade was never looking there.

Full write up on the blog, and there is a free twenty minute audit that walks one assignment rubric row by row until you find the criterion a polished answer cannot earn:
https://www.harvestkernel.com/the-missing-row-rubric-audit/

Here is the question, and a number is a complete answer.

How many rows in the rubric you are actually using this term could a well organized but entirely conventional answer earn full marks on?

Every class you are taking this term has an AI paragraph somewhere in its syllabus. They do not agree with each other, and nobody is going to walk you through the difference.

Two numbers from the Lumina Foundation and Gallup 2026 State of Higher Education study, and they describe the same students.

57 percent of US college students use AI for coursework at least weekly. About one in five use it daily.

53 percent say their school discourages AI or bans it outright.

Here is the part that reframed this for me. Gallup also asked the students who mostly stay away from AI why they stay away. The top answer was ethics. School policy came second. The least cited reason on the entire list was not knowing how to use it.

Students who avoid AI are not confused about the tool. They are confused about the rule. That is a completely different problem, and it is solvable in one sitting.

Today's teaching session walks through it: read the actual policy language, sort your real habits into allowed, ask first and never, and write the one email that settles the gray area. The free working session is linked in the resources space.

One question, and a number is a complete answer.

How many of your classes this term have told you, in writing, what you ARE allowed to do with AI, rather than only what you are not?

MIT's committee report has a line in it that is going to be quoted at faculty meetings for a year.

Students notice when instructors are clearly using AI for slides, feedback, or grading while imposing restrictions on student use, and they perceive it as a double standard.

That is not a warning about the future. The committee heard it in listening sessions this year, alongside a student refrain they printed in full: why should I bother coming to class or doing the work if the teacher is just going to give an AI-generated lecture.

The uncomfortable part is that the audit has already been run on all of us, informally, by people who sit in the room and read our comments. Writing a disclosure line does not create that judgement. It just stops pretending it is not happening.

Open the syllabus for a course you are teaching right now. Does it say anything at all about where you use AI? One word is a complete answer: yes or no.

Teaching Day, faculty. Week one of classes.

You sent the welcome announcement. It answered every first week question. By Thursday your inbox had those same questions in it.

Here is the thing that reframes it. You wrote that message on a laptop and could see the whole thing at once. The ones who opened it on a phone got it in a hallway, between two classes, with a thumb over the next thing. What they received was the top of it and a show more link.

So the message you sent and the message they got are two different documents. That is a design problem, and design problems have fixes.

Today's session does the rewrite live, and then hands you the part that makes it proof rather than advice: run it on your own two sections this week and count opens and replies on Thursday. Your numbers, not somebody else's.

The working session is here, six fields, nothing saved, nothing transmitted:
https://www.harvestkernel.com/the-two-version-announcement-test/

Open your welcome announcement on your phone right now. Before you scroll, can you see the one thing a student has to do next? One word answer is fine: yes or no.

A guideline landed at one university this month that asks every department for an AI policy naming three separate kinds of AI: generative, assistive and agentic.

Generative writes the draft. Assistive supports the person doing the work, which is your spellcheck, your read aloud, your captions. Agentic takes the actions, opens the sources, runs the code, and reports back.

Almost every syllabus AI paragraph written since 2023 is about the first one only.

Open your syllabus and read your AI paragraph. Which of the three does it actually name: generative, assistive, or agentic?

One word is a complete answer. If it names two, say which two.

Friday. Something shipped this week that changes what a syllabus is.

Google's new Gemini student hub reads a syllabus and adds the test dates and deadlines it finds to a student's Google Calendar. Google Search will now take a syllabus as a source for a study notebook. Both went live on August 19.

That means your syllabus now has two readers, and only one of them can ask you a follow up question.

Here is the part worth sitting with. A parser finds everything with a date attached to it. It does not find the reasons, the standards, or the policies you wrote as prose. Those are the parts of the course you probably care most about, and they are exactly the parts that do not survive extraction.

Nobody has to change a policy this week. But it is worth knowing which half of your own document travels.

Open your syllabus. How many things in it matter and carry no date? One number.

Syllabus week. Four PDFs landed in two days and you scrolled to the grading table on each one and closed it.

Here is the thing about those four documents. Every one of them was written as if it were the only class you take. Not because your instructors are careless. Because none of them can see the other three. I write syllabi. When I build mine I am looking at my course, my outcomes, my fourteen weeks. I have no idea who else is on your schedule.

So four people, working in good faith, in four different buildings, picked four sets of dates. Nobody stacked them.

The stack already exists. Somewhere in October there is a week holding three things, maybe four. It has been scheduled since before you arrived. The only reason it feels like a surprise in October is that the information was spread across four PDFs and nobody put them side by side.

Two things worth knowing before you look.

Your collision week is almost never the week you would guess. You guess from fear, so you predict the week your hardest course peaks. But a collision is not made by your hardest course. It is made by three ordinary courses landing in the same seven days. The class you are not worried about is the one that ruins the week.

And the move that fixes it expires. Taking one item out of that week and starting it early costs about two hours in August. It costs two hours you actually wanted in late September. In the collision week itself it costs nothing, because there is no early left.

The Collision Map is today's free working session. Every dated item from every syllabus onto one page, the collision weeks marked, and one decision written down. By hand, no account, nothing saved, nothing sent.

https://www.harvestkernel.com/the-collision-map/

Then tell me, because I actually want to know:

Which week on your calendar has three things in it? Month and week, one line. That is the whole question.