This week's theme is provenance and presence. The first half of the issue is about the machinery of proof: Anthropic starts invisibly watermarking Claude's text, universities retire AI detectors in favor of redesigned assessment, and Gallup measures just how routine student AI use has become. The second half is about what no watermark can certify: an engineer's regimen for protecting slow thought, a diagnosis of the meaning crisis in knowledge work, and an OpenAI researcher's meditation on what humans will still need from each other when the machines listen better than we do.
News
The EU AI Act's transparency rules took effect on August 2, and Anthropic has gone further than Brussels requires. An updated support page confirms that models released on or after that date weave an imperceptible, machine-readable watermark into generated text — applied at the model level, so it is present across the API, the Claude apps, and Claude Code; files are marked using the open C2PA standard. "Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere," the company says, and it "may persist through some editing." Alex Cui, CTO of the detection firm GPTZero, promptly published a skeptic's tour: in his testing, watermarks do not survive intense paraphrasing or wholesale human rewriting, and releasing a public detector invites adversaries to train against it — while keeping it private leaves it untested.
The temptation on campus will be to treat a watermark hit as a verdict. It isn't: the mark can show that Claude touched a text — even a proofreading pass could leave a trace — not that a student didn't do the thinking, and heavy editing may strip it. Before rumors reach students, our academic-integrity language should state plainly what provenance signals can and cannot prove.
Sources: "Anthropic says it will watermark text generated by its AI models" — TechCrunch · spotted via @alexcdot on X (GPTZero's CTO on whether text watermarks can be defeated)
Assessment
Kathryn Palmer reports for Inside Higher Ed that the detector era is quietly ending: Yale, Vanderbilt, Johns Hopkins, and Indiana now discourage or ban AI-detection output as sole evidence of cheating, and at least a dozen universities — Northwestern, Georgetown, and NYU among them — have disabled Turnitin's AI detector outright, citing false positives and bias against non-native English writers. With 73 percent of faculty saying they have personally dealt with AI-related academic integrity issues, the burden shifts to assessment design: process-focused assignments, blue-book and oral exams, and transparency requirements that ask students to show how they used AI rather than pretend they didn't. Yale's teaching center puts the rationale plainly: "We want to avoid the inevitable cat and mouse game created by AI detection tools."
The false-positive problem lands hardest on exactly our population: multilingual students writing in English as a second or third language. And the recommended alternatives — oral exams, process portfolios, relationships strong enough that students want to learn — are far more feasible in our small seminars than at a state flagship. Worth circulating before fall syllabi are finalized.
Source: "AI Detectors Are Out, New Assessments Are In" — Inside Higher Ed
Data
Numbers for the policy conversation. In the Lumina Foundation–Gallup State of Higher Education study (6,010 adults surveyed October 2–31, 2025), 57% of U.S. college students say they use AI in coursework at least weekly, and about one in five use it daily — 27% of men versus 17% of women. Institutions haven't caught up: 42% of students say their school discourages AI use and 11% say it is prohibited entirely, while only 7% say it is freely encouraged. What students actually do with it is mostly unglamorous: 64% use it at least weekly for help with coursework they don't understand, 60% to check answers, 54% to edit writing — and 36% to write papers. Among monthly-plus users, nearly nine in ten cite understanding complex material as an important reason.
Gallup's release includes clean charts of use frequency, campus policies, and specific uses — see the figures in the original.
If a majority of students arrive using AI weekly while roughly half hear only discouragement or prohibition, the gap itself is the risk: use goes underground and the teachable conversation never happens. With orientation weeks away, that argues for naming the reality to incoming students early — where AI genuinely helps learning, and where it hollows it out — rather than leaving policy to be inferred from syllabus boilerplate.
Source: "AI Is Routine for College Students, Despite Campus Limits" — Gallup
Practice
Sean Goedecke, an engineer who works with frontier models daily, describes what that work does to attention: agentic coding turns the day into a game show of rapid verdicts — six or seven agent sessions open at once, an endless stream of output to skim and judge — with little room left for slow thought. Slowing down is both miserable ("it's just such a miserable experience to spend your day close-reading LLM output") and competitively irrational when everyone else is pressing the solve-it-ten-times-faster button. His countermeasures are deliberately old-fashioned: write in your own words, because writing "forces you to articulate your thoughts. In a very real sense, it forces you to think"; and read dense books, "the antithesis of AI slop." After some weeks of the regimen: "I can feel parts of my brain stretching again."
Our students are entering exactly this workplace, and the capacity at risk — sustained, self-directed thought — is the one a liberal-arts curriculum is best positioned to protect. Goedecke gives us a working professional's testimony that reading- and writing-heavy seminars aren't nostalgia; they are maintenance for a skill his industry is actively eroding. It also names something many of us now feel in our own AI-assisted research and grading.
Source: "How to keep thinking" — Sean Goedecke
Work & meaning
Aaron Horwath, who runs AI operations at a creative-technology firm, asks in Noema why the mood among knowledge workers is so grim when the tools are so impressive. His answer is existential rather than economic: building on Derek Thompson's "workism" and Guy Debord's Society of the Spectacle, he argues that AI adds a final layer of abstraction between white-collar workers and their labor, exposing how much of it was already hollow — David Graeber's "bullshit jobs," now with better software. He divides colleagues into outcome-first workers, who happily hand tasks to the machine, and experience-first workers, who valued the collaborative "messy middle" and may walk away from corporate life entirely. The essay opens with a finance professional on a train who lights up only when knitting a hat for his niece.
Career advising should absorb this one: our graduates overwhelmingly aim at knowledge work, the very category now having its meaning crisis. A liberal-arts education has an honest answer — purpose, judgment, and an identity not wholly fused to a job title — but only if we make that case explicitly in advising and alumni programming, not just in commencement speeches.
Source: "Why Is Everyone In Tech So Sad?" — Noema
Essay
Houda Nait El Barj, an OpenAI researcher working on AI and human flourishing, writes the most personal essay of the batch, moving between her grandmother's courtyard in Marrakesh and Silicon Valley. Her central distinction: AI can interpret human experience but cannot participate in it. "AI can interpret my grief. It cannot grieve. It can model hunger, but it has never been hungry." She refuses the easy comfort that AI companionship is fake — Harvard Business School studies found AI companions eased loneliness "at a level on par with interacting with another person," and an OpenAI–MIT Media Lab trial of nearly a thousand people found the heaviest users were also the loneliest. Precisely because it works, it lets us skip the friction through which meaning forms. "The great scarcity of the future may be reality raw enough to form us."
A residential liberal-arts campus is, in her vocabulary, a presence machine: shared meals, common rooms, advising conversations — friction on purpose. As AI companions become the default listener for lonely students, our comparative advantage isn't better chatbots; it's unoptimized human encounter. We should program for it deliberately, and watch for the students who have quietly stopped seeking it.
Source: "How AI Will Change Us" — Noema
We are collecting examples for AI experimentation at DKU.
— Yisu