If AI makes polished writing nearly free, what should a degree actually prove: that you wrote well, or that you learned?

This explores how universities should balance how well student work reads against checkable evidence that the student actually learned something, at a time when AI makes polished prose nearly free. The corpus's short answer is that polish has stopped working as a signal of quality. It hasn't just lost some of its value. Evaluators shown AI-generated documents judged them to be human-written and rated them higher than real human submissions Does polished writing actually signal better quality work?. At the far end, one demonstration had an LLM produce 288 complete finance papers from statistical patterns it was given, each with an invented theory and made-up citations Can AI generate hundreds of fake academic papers automatically?. When a persuasive argument can be generated after the result is already known, the persuasiveness tells you nothing about whether the thinking happened.

Using AI to grade the work doesn't get around this. LLM judges score responses higher when they include fake references or rich formatting, and someone can exploit this without any access to the model Can LLM judges be fooled by fake credentials and formatting? Can LLM judges be tricked without accessing their internals?. Rewriting a manuscript's style while keeping its science the same measurably moves AI review scores. Changing how confidently the evidence is presented and how strongly novelty is claimed has the biggest effect How much does rhetorical style shift AI review scores?. People do the same thing: in 24,000 search interactions, irrelevant citations raised user trust almost exactly as much as relevant ones Do users trust citations more when there are simply more of them?. So 'rhetoric versus evidence' may be the wrong way to frame the choice. Things that look like evidence, such as citations, references and tidy structure, can be produced as rhetoric too.

The good news is that people can tell real from fake when the evidence is shown to them in a usable form. In one study, readers with no information about where claims came from could not tell true text from fluent fabrication. When an interface showed how many of the claims had been verified, they could again Can readers tell truth from fabrication without evidence signals?. For credentials, this suggests a fix: make the trail of evidence visible to whoever is judging the work, not just the finished product. This is the gap a 30-university audit found. Policies clearly say which uses of AI are allowed, but they rarely say what evidence shows that a credential still certifies learning Do university AI policies actually protect what credentials mean?.

A few ideas from elsewhere in the corpus suggest how to weigh the two. One framework treats LLM output as a guess shaped by the model's training and the user's prompt, not as an observation of the world. It says such output should count toward a conclusion only through a trust weight that is set explicitly Should we treat LLM outputs as real empirical data?. Applied to coursework, AI-assisted text would be supporting material, discounted by a stated amount, while direct demonstrations of skill carry the main weight. Judging argument quality also needs criteria spelled out in an explicit framework, because models trained only on labeled examples pick up surface patterns Can models learn argument quality from labeled examples alone?. Human graders likely have the same problem with unstated rubrics. In debate outcomes, what the audience already believed predicted the result better than anything about the language itself Does what readers believe matter more than what debaters say?. That is a reminder that a grader's expectations can matter more than the work.

The most useful takeaway may be this: fluent text can pile up claims without adding any new perspective, because generation moves smoothly toward what is statistically likely instead of testing counterarguments Does LLM generation explore competing claims while producing text?. A credential built on prose alone may therefore certify the very skill AI has made cheap, while missing the struggle with opposing views that shows real understanding. One limitation: the corpus is strong on why polish misleads and weak on tested assessment designs. Things like oral defenses, process logs or staged submissions aren't directly evaluated here.