OpenAI published a claimed Navier Stokes proof and the mathematicians disputing it never once asked for a detector. They asked for a checkable file, a dated record, and a statement of limits. Those three things are assignment design.
OpenAI says ChatGPT ads are labeled and separate today, and that the next phase makes them more native. The response is not a ban. It is one editing pass on the assignment.
A 2026 preprint found 1,655 fabricated author records on Zenodo carrying real DOIs, backdated and indexed. They pass the four source checks most of us teach. Here is the check that still works.
A randomized trial from OpenAI Economic Research and Bocconi University gave more than 1,000 undergraduates ChatGPT, critical thinking training, both, or neither. The rubric caught the ChatGPT polish and could not see the original ideas. Here is the one row to add before your next stack lands.
A PNAS study of NSF and NIH proposals found that heavier language model use did not lower quality. It lowered semantic distinctiveness, pulling proposals toward work the agency had already funded. The fix is not abstinence. It is what you put in the box first.
Two studies this year, one on undergraduate grades and one on federal grant funding, found the same shape. The score went up. The substance did not. The problem is not in anyone’s ethics. It is in the instrument.
Anthropic began marking Claude output on August 2, and the mark travels worldwide. Read the documentation and it says the content may have been processed by Claude. May. Processed. Not wrote. Here is what to ask students for instead.
Ninety four percent of higher education staff used an AI tool for work last year. Forty six percent could not tell you what their institution’s policy says. That gap is not a compliance failure. It is a clock failure, and a faster committee will not close it.
The first fully AI native class arrives this fall, and the instinct is to figure out how to police them. That is the wrong first move. Here is a two hour, week one plan for one real assignment instead.
The largest study of AI use in higher education tracked 76,485 students through a month of server logs. Usage peaked on weekday afternoons and thinned on weekends. AI did not create a study habit. It moved into the one students already had.
AI literacy education built on the SeedStacking™ methodology. Small daily wins that compound into genuine fluency.
We use cookies to improve your experience and analyze site traffic. By continuing, you agree to our Cookie Policy.
The Friday Harvest
One email a week. What changed in AI and higher education, and one thing you can do about it before Monday. Written by a college faculty member who still grades every week.
Get the Friday Harvest