myofficehours.ai
This is issue 3, from 14 August. Read the latest issue →

Issue 3 · 14 August 2026

myofficehours.ai — Issue 3

14 August 2026 · A weekly read on AI for faculty

Detection moved to the vendor, and the vendor is not ready

Two days ago the news was that four universities had switched their AI detection software off. This week the thing meant to replace it arrived in two forms: a watermark Anthropic will build into Claude's writing, and a legal duty on the large AI companies to publish a detector anyone can use. The first independent test of the second one found most of it missing.

Anthropic set out how its watermark works. The model's low-stakes word choices are nudged using a secret key, so whoever holds that key can calculate how likely it is that Claude wrote a passage. Nothing is added to the text and nothing identifies a person. The reason it is happening is the European Union's AI Act, which several other providers are also now answering. That Act and the California AI Transparency Act began requiring covered companies to offer a public detector on 2 August. Indicator and the human rights group WITNESS went and looked: of thirteen companies large enough to be covered, six had published one, and all but one of those could be fooled.

What this means for your syllabus

A watermark is a better instrument than the detection software the sector spent three years arguing about, because the evidence is placed in the text as it is written rather than guessed at afterwards from the surface of the prose. It is also an instrument you will never operate. The key belongs to the company, the detector is theirs to publish, and this week's census says most have not published one. So if you are finalising a policy this fortnight, stop writing sentences that promise a check. Say what you want a student to do, what you want them to show you about how they did it, and what happens when they show you nothing.

The counter-current worth noticing

Two studies this week complicate the assumption underneath all of it, which is that the tools reliably help. Northwestern researchers examined 5,700 confidential grant proposals to the National Science Foundation and the National Institutes of Health and found that proposals showing heavy language-model use were measurably less distinctive from work already funded. At the National Institutes of Health they were more likely to be funded, and the extra papers that followed were the less-cited ones. Separately, a study of 13,037 students compared three ways of giving AI feedback: 26 per cent engaged when the workflow made them act on a suggestion, 14 per cent when comments simply arrived, and 0.1 per cent when it was optional. Making a tool available is not the same as it being used, and using it is not the same as it helping.

The decisions being made while you are away

California's legislature sent the governor a bill stating that the instructor of record for a state university course must be a person, and cleared a second that would oblige campuses to set AI procurement standards, require vendors to disclose what their models were trained on, and keep four years of training records. Colorado's attorney general opened comment until 26 October on rules covering advising chatbots and algorithmic decisions. Underneath all of it sits the problem Stateline reported this week: there is almost no federal guidance on buying these tools, the providers know far more than the purchasers do, and the burden of deciding falls entirely on the institution. That burden is lightest where there are staff to carry it, which is how a gap opens between two institutions paying the same licence fee.

One thing to actually try this week

Take one assignment you already teach and rewrite it so a model cannot complete it alone. The dashboard has a twenty-minute route under Teaching and course design: put the thinking somewhere a model cannot reach, which usually means your own data, this term's discussion, or a decision the student defends out loud. If you would rather work with the tools than around them, Teaching AI literacy now has a session plan where students critique what a model gave them and are graded on that judgement.

What someone who studies this thinks

Interviewed by Joshua Kim this week, Howell says something administrators rarely put in print: she has yet to see compelling evidence of large gains in teaching and learning from AI, and thinks it premature to know whether the net effect will be positive. What worries her is not cheating but what she calls the ubiquitous temptation of cognitive surrender. The part worth copying is what she does next, which is to spend money this autumn on grants that pay faculty to redesign residential courses, on the reasoning that the redesign is worth doing whichever way the evidence eventually lands.

— Sonia Howell, Director of digital learning, University of Notre Dame · read the piece

Where the experts actually disagree

Now that verification is becoming a vendor service, what should an academic integrity policy rest on?

Anthropic (neutral) — Put the evidence in the text as it is generated. A watermark keyed to the provider is a stronger answer than inference from style, and regulation now requires it of everyone. Their argument

Alexios Mantzarlis (skeptic) — Measure the machinery before relying on it. Seven of thirteen covered companies have published no detector at all, in apparent breach of a duty already in force, and all but one of those that exist were fooled in testing. Their argument

John Warner (skeptic) — Verification is the wrong lever. Students outsource work because the transactional model rewards the grade rather than the learning, and Michigan's pass or no credit first semester attacks the incentive instead of the symptom. Their argument

Nicole Brownlie (pragmatist) — Change the question the assessment asks, from whether the student wrote this to how the student was thinking, with short guidance attached to each reading rather than a module completed once. Their argument

The one position with no support this week is the one many institutions are quietly holding, which is that a reliable check will arrive shortly and the policy can wait for it.

Threads we have been following

Issue 2, 13 August 2026 — we said: Yale, Vanderbilt, Johns Hopkins and Indiana had switched off or discouraged AI detection software, which we called a rare instance of the sector de-escalating rather than buying something.

Within two days the replacement showed itself, and it is not a purchase either. Anthropic published the design of a watermark it will build into Claude, driven by the European Union's AI Act, while California and Colorado moved on statute and rules. None of that gives an institution an instrument it can run itself. Indicator and WITNESS tested the detectors those same laws require and found six of thirteen companies had published one. The universities that stopped buying detection were right, and the thing meant to replace it is not yet working.

What just became possible

Not things to do this week. Things that can now be done at all.

Agent-driven reproduction of a whole conference's published claims, at about twenty dollars of compute per paper (a demonstration)

A researcher can now have a coding agent re-derive and re-run a published paper's central claims for roughly twenty dollars of cloud compute, and doing this across 2,226 papers from one machine-learning conference falsified or contested at least one claim in 496 of them.

What to do now: Point a coding agent at a paper you are refereeing or building on and have it try to reproduce the result. Every logbook, verdict and agent trace from the exercise is public, so you can copy a working reproduction rather than start cold. The research route on the dashboard is the free starting point for reading with a model. Hugging Face

Open optical character recognition good enough for eighteenth and nineteenth century print, priced per thousand pages (already shipping)

A library or a humanities department can re-read its digitised eighteenth and nineteenth century books with freely licensed models at forty-six cents to under two dollars per thousand pages, at 96.9 to 97.6 per cent character accuracy measured against 2,165 pages transcribed by hand.

What to do now: Run one of the leading open models over a sample of your own scans and score it yourself, since the ground-truth dataset and the scoring harness are both published and the smallest model fits on a laptop. Read the authors' two warnings first: the models quietly modernise the long s, and the evaluation covers roman typefaces only, not Fraktur, non-Latin scripts or handwriting. Hugging Face

The rest of the week, briefly

One line each, ordered by how much it should change what you do. The full account of any of them is on the dashboard.

Act on this

Watch

Three worth your time

156 resources went into the library this week. These three are the ones to open first.

[Ten Simple Rules for Using AI in Grant Writing](https://medicine.stanford.edu/news/stories/2025/07/10-rules-for-ai-in-grant-writing.html) — grants · 10 min · Stanford Medicine Ten concrete rules for using AI responsibly in grant proposals, including checking funder-specific AI policies, never pasting unpublished data into public chatbots, and verifying every AI-suggested citation before submission. You will be able to use AI as an editing and brainstorming aid without risking intellectual property leaks or fabricated references.

[AI-Resilient Assignments](https://ctl.wustl.edu/resources/ai-resistant-assignments/) — Teaching and course design · 15 min · WashU Center for Teaching and Learning Six concrete strategies for making assignments harder to complete with AI alone, such as requiring authentic real-world tasks, oral exams, or references to specific in-class discussions the model cannot access. You will be able to pick at least one strategy and apply it to an existing assignment this week.

[Semantic Scholar Tutorials](https://www.semanticscholar.org/product/tutorials) — Research and literature · 20 min · Allen Institute for AI Three short official tutorials with video and transcript teach how to navigate the citation graph by type, filter search results by field, date, and author, and use paper pages to find related code, figures, and citing works. Afterward you can build a targeted reading list and trace how a key paper has been used since publication.

Who we read this week

This issue drew on Anthropic, Indicator, Inside Higher Ed, Times Higher Education. The full watchlist, with what each source is good for and where they stand on AI in education, is on the dashboard under "By voice".

The whole library lives on the dashboard, sorted by what you are trying to get done and by the tools you already have. Each task runs from a twenty-minute start to something you could spend a weekend on.

myofficehours.ai is assembled automatically: a daily sweep for new tutorials and a weekly edition on Monday mornings. Every link is checked before it ships. Reply with anything broken, missing, or worth adding.

← Every issue