AI text is hidden in exams and anchored to law in rental notices
Universities are turning to offline exams and oral defences to measure students' own knowledge. India's Recourse instead retrieves state law before the model drafts a tenant notice, completing Karnataka cases in under five minutes.
Artificial Intelligence··Midday
AI use is widespread and its exam answers are hard to spot
About 90 percent of Harvard undergraduates say they use AI for coursework, and roughly 25 percent say they have used it instead of assigned reading. That prevalence is not a shift that can be measured simply by asking students or judging the style of a response. At the University of Reading, human graders identified AI-written exam answers only 6 percent of the time. In a Brown economics course, the average on take-home midterms was 96 percent, compared with 48.6 percent on the in-class final. These two results put numbers behind universities' decision to redesign the setting in which students demonstrate their own knowledge.[1]
Sydney creates two lanes while Chicago adds oral defence
The University of Sydney now divides assessment into two lanes. Secure, proctored exams sit beside coursework in which AI is used openly. The University of Chicago Law School is piloting device-free first-year classes, with offline exams and oral defences of written work. These arrangements do not try to remove AI from every assignment. They separate work in which the tool may be used from the moment when a student has to show what they know without it. The oral defence also creates a second channel of assessment, one based not only on the submitted text but on how the student presents the reasoning behind it.[1]
Recourse retrieves state law before the model starts writing
Siddharth Agrawal, a product manager in Bengaluru, built Recourse over a weekend after a deposit dispute with his own landlord. The system first retrieves state-specific rental law, filters the relevant material and only then asks Gemini 2.5 to draft the notice. Its retrieval layer uses Redis and a vector database. A Karnataka notice takes less than five minutes, while retrieval for Delhi still fails. After launch, the site drew more than 3,000 visits and over 1,000 registrations, and 51 percent of those who began the process completed it. Here the generated text rests not on the model's general knowledge but on the currently applicable state rule selected before drafting begins.[2]
