A sixth-grade humanities teacher in San Francisco has replaced take-home essays with oral exams, asking students to discuss character, plot, theme and metaphor. Tristen Chang designs her classroom so students cannot rely on chatbots; students can answer the 15 questions in as little as five minutes. The example comes from a Reuters feature by Helen Coster, which describes teachers rethinking how they assess students as generative AI becomes more common.
Chang’s approach is one teacher’s classroom choice, not evidence of a districtwide requirement. That distinction matters in San Francisco, where the status of school AI rules remains in development. SFUSD says its linked AI document is general guidance for internal staff use and should not be treated as official, board-approved policy.
Local rules are still taking shape
In a September 15 report, KALW reported that a Presidio Middle School student said AI rules varied among his teachers. KALW also reported that SFUSD was drafting an AI policy, with a school board vote scheduled for March 9, 2027. The district’s staff-guidance page does not itself establish that such a board-approved policy has been adopted.
The district’s webpage also warns that online ChatGPT plagiarism checkers can be unreliable and says uploading student work to many of them violates SFUSD’s student privacy protection. Those cautions speak to risks in trying to detect AI use; they do not determine how individual teachers should assess students.
More use, but uneven instruction
National figures help explain why teachers are reassessing assignments, without showing that any one classroom method works best. RAND reported that in 2025, 54% of students and 53% of English language arts, math and science teachers said they used AI for school—increases of more than 15 percentage points from surveys one to two years earlier. In a separate measure reported by RAND, the share of middle-school and older students using AI for homework rose from 48% in May 2025 to 62% in December.
That growth has not been matched by consistent instruction, according to RAND: as of spring 2025, 35% of district leaders said their districts provided students with AI training, and more than 80% of students said teachers had never explicitly taught them how to use AI for schoolwork. The figures describe broad patterns, not SFUSD specifically.
What oral exams can—and cannot—show
A live oral exam can make a student’s explanation available to a teacher in the moment, but Chang’s example alone does not establish how reliable or effective the method is across students or subjects. Nor does research yet settle the broader question of AI’s effects on learning. A small MIT Media Lab essay-writing study reported weaker overall brain connectivity among participants using an LLM than among those using a search engine or no tool. The researchers said the preprint had not been peer-reviewed and that its conclusions should be treated cautiously; the study is not proof that oral exams improve learning.
For now, Chang’s choice is a response to a practical assessment problem: how to see students work through ideas when take-home writing may involve AI assistance. The local rules that will govern such choices are still being worked out, while national data show student use rising faster than many districts’ training and guidance.
