myofficehours.ai — Issue 2

13 August 2026 · A weekly read on AI for faculty

The same tool is no longer the same tool

Three weeks before term, the word ChatGPT in your syllabus quietly stopped naming one product. On Tuesday OpenAI extended its advertising test to the United Kingdom, Mexico, Brazil, Japan and South Korea, and the ads appear on the Free and Go tiers only. Pro, Business, Enterprise and Education accounts stay clean. The assistant your students have and the assistant your institution licenses are now different things.

Two days ago this briefing led with the observation that the price of admission had dropped to zero. That was true and it is still true, and this week is the invoice. Ads on the free tier are matched using the topic of the conversation and the history of previous ones, which is a reasonable arrangement and also a material difference between two products that share a name. The same sorting appeared elsewhere. Anthropic loosened its biology filters enough that ordinary clinical and teaching questions now get answered, while keeping the capability a working virologist would want behind a research pathway it has promised rather than built. Access did not stop being free. It started being tiered.

What this changes in your syllabus

If your course policy names a tool, it now has to say which version, because what a student actually experiences depends on what they are paying. That is not an argument for requiring a subscription, which would be the worst possible response to a widening gap. It is an argument for writing the policy around what you want done rather than around a product name, and for checking one thing before term: whether your institution's licence covers the students in your room or only the staff teaching them. If it covers only staff, that is the difference between a class working with the same assistant and a class where some of them are being advertised to while they work.

The other kind of access, and who pays for it

Underneath the consumer tiers sits a harder version of the same problem. Steven Goodman at Marquette points out that a rack of modern AI hardware draws sixty to a hundred kilowatts where a conventional academic rack draws five to fifteen, which is more than most university data centres were built to supply, and argues that researchers now follow infrastructure the way they once followed libraries. Fisk University is testing what happens when an institution decides to buy its way in: a four-hundred-million-dollar data centre inside a nine-hundred-million-dollar plan, eighteen thousand signatures against it, and a city council that has stopped issuing permits. Compute has become a question about institutional finance and about neighbours, not about IT.

The counter-current worth noticing

It would be easy to read all of that as capability being rationed by whoever sells it, so the week's evidence pointing the other way is worth flagging. Marc Watkins describes three teaching tools built by academics with no software team behind them: a searchable index of prize-winning nonfiction, a Homer reader setting the Greek beside several translations, and his own workshop in which students judge AI-generated anatomy against curated images. None went through procurement. Separately, Yale, Vanderbilt, Johns Hopkins and Indiana have switched off or discouraged AI detection software, which is a rare instance of the sector de-escalating rather than buying something.

One thing to actually try this week

Take one assignment and add a push-back step, borrowing the protocol Tawnya Means published on Monday. Students may use a model, but what they submit includes a record of the interrogation: the evidence they demanded, the assumption they surfaced, the claim they tested against a real case, and what they rejected. You grade the interrogation, not the output. It takes about twenty minutes to write, it works on a free account, and it is the only version of this that survives a student whose model is better than yours. The dashboard has the full route under Teaching AI literacy.

What someone who studies this thinks

Watkins spent this week arguing that the interesting change is not what the models can do but who is now able to build with them. His three examples are all faculty who wanted a tool that did not exist and simply made it, iterating rather than prompting once, with no vendor and no budget line involved. His reading of the moment is that universities have spent three years choosing between adopting an AI product and banning it, and have barely noticed that a third option appeared: owning the thing outright. He is careful not to oversell it, and warns against unexamined adoption in the same paragraph, but the point stands. In a week when almost every other story was about capability being metered by the people selling it, the cheapest capability on offer turned out to be the one you can build yourself.

— Marc Watkins, University of Mississippi, where he runs the AI Institute for Teachers · read the piece

Where the experts actually disagree

Should an institution act now on what AI appears to do for learning?

Tawnya Means (enthusiast) — Yes, and stop waiting for the sector to settle it. Evaluation is a teachable competency, it has a mechanism, and a class can run it next week: treat every output as a first draft, demand the evidence, and be graded on the interrogation rather than the answer. Their argument

Jeppe Klitgaard Stricker (pragmatist) — Act locally, but refuse to let a local result authorise a systemic decision. A benefit claim is a snapshot and a harm claim is a trajectory, and an institution that rewrites its assessment regime on the strength of one strong course-level study has mistaken one for the other. Their argument

Ben Williamson (skeptic) — Look at who is producing the evidence before you act on any of it. Government testbeds, a Department for Education agreement with Google DeepMind, and OpenAI's own measurement suite and twenty-thousand-student trial mean the sellers increasingly design and score the studies their buyers cite. Their argument

The three are not actually contradicting one another, which is the useful part. Means is telling you what to do on Tuesday, Stricker is telling you what that Tuesday does not entitle anyone to conclude, and Williamson is telling you where the slide deck in the provost's meeting came from. A department that holds all three at once will make better decisions than one that has picked a side.

Threads we have been following

Issue 1, 11 August 2026 — we said: the price of admission had dropped to zero, and your students no longer hit a usage wall on the free tier.

Both halves of that are still true, and two days later OpenAI extended its advertising test to five more countries. The free tier has become the advertising tier, targeted on the topic of the conversation and on past chats, while the Education, Enterprise, Business and Pro tiers stay ad-free. The correction is not that access closed again. It is that free and licensed have stopped being the same product with different limits, and a course policy written against a brand name no longer describes what every student in the room is using.

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

• OpenAI expands its ChatGPT advertising test to five more countries — The free tier your students use is now ad-supported and targeted on their own chat history.
• Yale, Vanderbilt, Johns Hopkins and Indiana turn off AI detection — Peer institutions are switching detectors off, and a detector score will not survive an appeal.

Watch

• Anthropic loosens Claude Fable 5's biology filters but keeps professional research locked — Everyday biology questions stop being refused; the capability a working lab wants is still gated.
• Purdue makes AI fluency a graduation requirement, and MIT's Justin Reich argues for slowing down — The first US university has made AI fluency a graduation requirement; expect your curriculum committee to ask.
• Fisk University bets four hundred million dollars on a data centre, and Nashville stops issuing permits — Compute has become a campus finance question, and at Fisk the neighbours have a vote.
• The research computing gap is becoming a hiring problem — Power and cooling have become recruitment arguments, because the hire follows the hardware.
• Who gets to decide whether AI works in education — Ask who ran the study before accepting the finding, because increasingly the vendor did.
• A university sues an AI company over the architecture rather than the training data — The first university patent suit against an AI company, aimed at the architecture rather than the training data.

3 more stories ran this week and are waiting on the dashboard.

Three worth your time

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

[Napkin AI: Turn Text into Diagrams and Visuals](https://www.napkin.ai/) — Images and figures · 15 min · Napkin AI Paste in existing text, such as a paragraph from a lecture or a paper, and Napkin auto-generates an editable diagram, flowchart, or infographic that exports as PNG, SVG, or PowerPoint. You will be able to turn a wall of bullet points into a clean slide visual without prompting or design skill.

[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.

Who we read this week

This issue drew on Inside Higher Ed, Tawnya Means, MIT Schwarzman College of Computing, US Copyright Office, OpenAI, Anthropic, Ben Williamson, Chronicle of 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.