New on the blog: AI Grading Makes a Harsh Mark Harder to Fix. Grade Blind First.
In a randomized experiment with 1,339 teachers, published this year in PNAS Nexus, a grade that was three points too harsh survived review more often when it was labeled as coming from AI. A too generous grade got fixed either way. England's qualifications regulator names the same risk in its January working paper, anchoring, and notes that marking in parallel avoids it.
The post walks through the evidence, why academic judgment depends on who actually formed the view, and a five step way to mark first on this week's stack.
Read it here: https://www.harvestkernel.com/ai-grading-anchoring-grade-blind-first/
New on the blog: How Students Use ChatGPT for Essays, and What Cut the Outsourcing.
A Virginia Tech team recorded every request 77 college students made to ChatGPT while writing an essay. The paper reports that asking for the entire essay was the single most common request. Students who used it to plan or to review still felt the essay was as much theirs as students who never used it.
What cut those requests in a later classroom deployment was not a detector. It was knowing the instructor could see the AI history.
Read it here: https://www.harvestkernel.com/how-students-use-chatgpt-for-essays/
New on the blog: Alaska's Draft AI Policy Says a Human Decides. Your Syllabus Should Say Where.
The University of Alaska's regents take up systemwide AI principles in November. The draft defines human in the loop in one sentence: AI could recommend, but a human decides. AI cannot act alone.
Most syllabus AI policies cover what students do with AI. This post is about the other half, the places where your own tools recommend and you decide.
Read it here: https://www.harvestkernel.com/university-ai-policy-human-decides/
New on the blog: Dartmouth Just Authorized an AI Detector. Its Own Rules Say the Score Is Not the Case.
On September 29 Dartmouth gave its faculty seven guidelines on AI detection. Most coverage stopped at the permission. Guideline six is the part worth reading: before any finding, the standards committee looks at a "constellation" of information beyond the detector result.
So where does the rest of the constellation come from? Your assignment, if you build it in before the paper is due.
Read it here: https://www.harvestkernel.com/dartmouth-ai-detector-guidelines/
New post. Proof of Work Just Became an Admissions Standard. Your Transcript Did Not.
Andreessen Horowitz launched a school yesterday. Everyone quoted the no grades, no tests, no homework part. Almost nobody quoted the paragraph near the bottom, where the Academy says it intends to seek regulatory approval for a two year program in Fall 2028 and, if approved, expects to charge annual tuition similar to an elite private university.
The free year is the demo. The filing is the product.
The post covers what was actually announced, what the admissions page asks a seventeen year old to submit instead of an essay, the Berkeley working paper that explains why the opening was there, and the three things that follow for anyone who grades work.
https://www.harvestkernel.com/proof-of-work-admissions-standard/
New on the blog: Microsoft Called It a Doom Loop. Your Reading Assignment Is Already Inside It.
A summary judgment brief unsealed on September 17 in the news organizations' lawsuit against OpenAI and Microsoft quotes an internal Microsoft document describing a "doom loop" for the web, and says Satya Nadella agreed under oath that chatbots give people the information "versus needing to go to the underlying source."
The post is about what that means for the one assumption every reading assignment rests on: that the student opens the source. A summary is usually decent. It drops the qualification, the example, the turn and the voice, which are the four things we assign readings for. The same week, MIT relaunched MIT Reads around reading whole books together.
Read it here: https://www.harvestkernel.com/reading-assignment-doom-loop/
New on the blog.
A state supreme court held a lawyer in contempt this month for filing a brief containing testimony from four witnesses who do not exist. Most coverage made it a story about ChatGPT. Read the actual order and the chatbot is a single neutral sentence in paragraph three.
The finding is in paragraph five. He did not verify the factual claims and legal authority in the brief before signing it and filing it.
That is a workflow with a produce step and a filing step and nothing in between, and it is the same shape as almost every assignment in higher education right now. Your students hit submit, and submit is a signature, and nobody has ever told them what it covers.
The piece works through what the order actually says, who really paid for it, and what a verification step looks like when you build one.
https://www.harvestkernel.com/ai-verification-step-signature/
Free AI Curriculum Has a Byline. Read It Before Your Department Votes.
Natasha Singer spent years documenting how the technology industry got into classrooms the first time, and her new book Coding Kids lays out the mechanics. Apple and Microsoft each wrote their own AP Computer Science curricula around their own tools. Google put Chromebooks everywhere and then put Classroom on top of them. Where a company did not supply curriculum directly, it funded a nonprofit that did.
Today's post traces that pattern forward into the AI cycle, with one verifiable checkpoint: page two of the AILit Framework names the European Commission, the OECD and Code.org. That is not a scandal. It is a reason to read page two before a vote rather than after a cohort graduates.
Read it here: https://www.harvestkernel.com/free-ai-curriculum-byline/
The free working session that goes with it is The Curriculum Byline Audit. One unit, four questions, twenty minutes, and you keep the page.
New on the blog.
Nobody ran the ninety year proof through a detector. They asked for the file.
On Tuesday OpenAI published what it says is a resolution of the Navier Stokes problem, one of the seven Millennium Prize Problems. Within a day the mathematics community was fighting about it in public.
None of that fight was about whether a machine wrote it. Both sides had already shipped a formalization in Lean, which is a document a computer can check line by line without caring who produced it. So the argument moved to the only place left: the trail. Who started the program, whose drafts were where, what the model had access to, and on what dates.
The translation for your course is short. Your integrity language asks for a promise. The field that just lived the hardest version of this asked for three things instead: an artifact somebody else can check, a record of how the work was made, and a line saying what the author is not claiming.
Read it here: https://www.harvestkernel.com/millennium-proof-receipts-not-detectors/
Open your next assignment. What is the one thing in it you are taking on trust?
Today's post: Google Docs Now Talks Back. Your AI Policy Still Names a Website.
On Wednesday, September 3, Google shipped conversational voice features into Gmail, Docs and Keep. A student drafting an assignment can now talk to the document, and the document will structure the draft and, with permission, pull material from their mail, their files and the open web. No second tab.
The line that matters on a campus is the last one in Google's own announcement: these features are "coming soon for Google Workspace business customers." Colleges run Workspace. So the campus account did not get this. Consumer accounts did.
That is the shape of the term. Two students, same section, same rubric, same syllabus paragraph, different tools, because of which account they signed into.
There is a second thing in the post worth more than the feature itself. The machine written summary sitting on top of Google's own announcement contradicts the announcement underneath it, on exactly the availability fact a university needs. That is the best argument for review first in a while.
Read it here: https://www.harvestkernel.com/google-docs-ai-policy-names-a-website/