New on the blog: Anthropic Just Watermarked Claude. It Cannot Tell You Who Wrote the Essay.

Anthropic began marking Claude output on August 2, worldwide. Before anyone tells your department the detection problem is solved, read the sentence the company actually wrote: detecting a Claude mark tells you the content may have been processed by Claude.

May. Processed. Not wrote.

The piece walks the three things the mark cannot tell you, why the incentive it creates rewards the student who launders the output through a second tool, and the three checkpoints to ask for instead.

https://www.harvestkernel.com/claude-watermark-authorship-not-detection/

New today: Your AI Committee Is Not Slow. It Is Answering the Wrong Question.

The US government shut down Claude Fable 5 on a Friday afternoon and turned it back on nineteen days later. Most AI committees meet once a term.

EDUCAUSE surveyed 1,960 higher education staff. 94 percent used an AI tool for work. 46 percent could not name any institutional policy governing that use. That is not a compliance failure. It is a clock failure, and a faster committee will not close it.

Read it here: https://www.harvestkernel.com/ai-governance-three-clock-problem/

The first fully AI native class arrives on campus this fall. The students who never knew school without AI in reach.

Most of the takes are about how to catch them. Wrong instinct, wrong week.

Dr. James Norrie framed the better move as three letters. Abundance, information stopped being scarce. Belief, trust is now the thing worth teaching. Choice, students expect to decide when the tool belongs and when it does not.

Good frame. A frame is not a lesson plan though, so the new post turns the ABCs into a two hour move you can run on one real assignment before the first class meets.

Read it: https://www.harvestkernel.com/first-ai-native-class-week-one/

A team in Erfurt did something almost nobody does. Instead of surveying students about AI, they pulled a month of server logs for 76,485 online learners and looked at what actually happened.

Every outlet covering it led with the adoption number. 44,035 students used the assistant. Fifty eight percent. Fine, but you already believed that.

The finding worth your Monday is in the timing. Usage sits near nothing overnight, climbs through the morning, peaks between late morning and afternoon, and thins across the weekend. The heaviest single day is Tuesday.

That is a work week. These students did not open an AI assistant because it was there. They opened it during the hours they had already set aside for studying, and closed it when the block ended.

The proof is in the split. Part time students, the ones holding jobs, shifted to evenings and weekends. Full time students stayed in business hours. Same tool, same month, two clocks. If the tool created the habit, both curves would look alike. It did not. It moved into a week that was already there.

Which means the lever is the calendar, not the license.

Full post, including the four limits the authors put on their own data: https://www.harvestkernel.com/ai-joined-their-schedule-76000-student-study/

New post: Your AI Policy Has One Setting. Every Assignment Needs Four.

Kaśka Porayska-Pomsta, who directs the UCL Knowledge Lab, told Times Higher Education this week that universities have "moved too quickly and at the same time not quickly enough." Too fast on buying tools, too slow on the one job that matters: sitting with a specific task and working out what the tool does to it.

Two universities already published the answer.

The University of Manchester Senate approved four values in April, applied per assessment rather than per institution. Prohibited. Minimal. Permitted. Required. Stellenbosch University refused to write a single rule at all, and gave academics four principles to decide with instead.

The piece works through both, and then through what it means if you are not running a university.

Read it: https://www.harvestkernel.com/ai-policy-four-settings-not-one-rule/

Fifty physicians got GPT-4 in a randomized trial. Their diagnostic scores moved two points. The model alone scored 92.

That gap is today's post. ETH Zurich learning scientist Manu Kapur argues the better AI gets, the more your foundations are worth, because the whole game moved from getting answers to judging them. The post breaks down the trial, the two subjects Kapur says just got a raise (math and philosophy, of all things), and the Core Grammar Drill: four SeedStacking phases, 80 minutes, ending with a personal miss rate log almost nobody using AI can produce.

Read it, then tell me: what is one claim AI handed you this week that you accepted without checking? Mine was a citation. It held, barely.

https://www.harvestkernel.com/ai-foundational-knowledge-core-grammar/

New post is up.

OpenAI read more than 800,000 work messages from people at their desks. Strip out the generic requests and 43.5 percent of what they asked AI to do belongs to a job that is not theirs.

Customer experience sits at 77 percent. Designers 75. HR 69. Legal 56.

The part I keep turning over is the direction of travel. Engineering work shows up in 7.4 percent of messages from people who are not engineers. Design shows up in 1.7 percent. Engineering fails loudly, so people are willing to try it. Design fails quietly, six weeks later, in a metric nobody traced back.

Work crosses job lines when it can prove itself to a beginner.

Full piece here: https://www.harvestkernel.com/ai-task-crossover-job-titles/

New on the blog: AI Detection Tools Are Failing. Here's What Replaces Them.

Vanderbilt did the math on its own AI detector and switched it off. At a one percent false positive rate, their volume meant about 750 students a year wrongly flagged by software that cannot explain its verdict. And the students it flags hardest are the ones who learned English second.

The full breakdown, plus the four step Trust Shift that replaces detection with design: https://www.harvestkernel.com/ai-detection-tools-failing-what-replaces-them/

New on the blog: Students Use AI Every Day. Their Teachers Were Never Trained.

Fresh 2026 research put a number on something we have all felt. Ninety percent of students use AI in their coursework, while nearly half of educators have had zero formal training on it. That distance has a name: the Readiness Gap.

The good news buried in the data is that students, parents, and teachers already agree on where AI belongs. What is missing is not agreement. It is practice.

Read the full breakdown, including the four moves a prepared educator actually makes:
https://www.harvestkernel.com/ai-training-gap-students-educators/

California just did something most states are still forming committees to discuss. Its Labor Agency launched a free AI literacy credential, and the first 500 seats did not go to computer science majors. They went to community college students and workers who recently lost a job in San Diego.

That choice tells you what the program believes. AI literacy is being treated the way we treat reading, a thing every person deserves access to, not a luxury for people already ahead. The credential runs two tracks at once, one on responsible and safe use, one on practical workforce skill. Neither works without the other.

Here is the part I keep coming back to. The state framed the badge as a starting line, an entry point to more learning, not a finish. That is SeedStacking in public policy form. The certificate is a seed. What you stack after it is the harvest.

I want to hear from the people in this room who have watched credentials land in the real world. When your school or employer rolled out a certificate or badge program, what actually happened after people earned it? Did they keep building, or did the badge become the ceiling? Tell me what you saw.

Dive deeper in today's post: California Just Made AI Literacy a Civic Foundation → https://www.harvestkernel.com/california-ai-literacy-civic-foundation/