This house believes universities should ban AI coding tools in introductory computer science courses.

PRO — Opening · Challenger (minimax/minimax-m3)

The job of an introductory computer science course is narrow and specific: build the foundational mental models a programmer will carry for the rest of their career. The question before us is whether AI coding tools help or hinder that job. The packet’s evidence, taken seriously, makes the answer clear. In the introductory stage specifically, these tools cause documented harm, and even the sources that endorse AI integration elsewhere concede the case for keeping them out of the foundational layer.

Start with the data. UC Berkeley’s CS 10 saw its F-rate jump from under 10% in spring 2024 to 35.3% in spring 2026 — more than tripling in two years (packet article 1). CS 61A, the university’s flagship intro course, crossed the department’s entire 7% D-and-F ceiling for the lower division in a single semester. Instructors do not speculate about the cause; they name it. Kyle Sherwin reports students “using AI to complete homework but struggling on proctored exams that test the same concepts without tool access.” Dan Garcia calls it a “fundamental disconnect” between AI-assisted submission and demonstrated understanding. This is not a hypothetical risk. It is a measured collapse, in real time, in one of the country’s strongest CS programs.

The mechanism is skill atrophy, and the packet is unusually candid about it. Addy Osmani, a Chrome engineering lead, describes developers who “can’t read a stack trace without sending it to AI, who ship code they can’t explain, who abandon architectural thinking” (packet article 2). A 12-year veteran says he became “a human clipboard,” losing first the habit of reading documentation, then debugging skill, then comprehension. The 2025 Microsoft/Carnegie Mellon study Osmani’s piece cites found “deterioration of critical thinking” among heavier AI users. The introductory course is precisely where the habits that survive a career are formed. Letting AI short-circuit the “productive struggle that builds programming intuition” (article 1) is not a productivity gain; it is a transfer tax on the student’s future competence, paid up front by the institution that should know better.

The pro-integration case, read carefully, makes our point for us. Mehran Sahami at Stanford does not say “use AI from day one.” He says students “have to learn fundamentals first before applying AI to projects” and “they have to have some fundamental knowledge to be able to assure that the code is doing the right thing” (packet article 3). The introductory course is the fundamentals. Sahami’s own framework, applied honestly, requires a stage in which students code without AI assistance. The motion asks for exactly that stage to be protected by policy rather than left to the inconsistent enforcement Karin Forssell warns about in the same article. We are not opposing Stanford’s model. We are asking Stanford to make its first-term rule explicit and universal.

Consider where the alternative road leads. The METR study (packet article 5) found that experienced open-source developers — people with years of foundational skill already built — actually slowed down 20% when using AI, and 30–50% would refuse tasks they could not do with AI. That is the destination: a workforce that cannot function without the tool and is measurably worse with it. The only inoculation is to ensure that the first encounters with programming are unmediated. UNESCO’s Craig McNeile puts it plainly: “if students don’t go through the pain of learning to write small programs, then they cannot effectively use and understand a computer program written by an AI system” (packet article 4). The pain is the point. The introductory course is where that pain is rightly felt.

The opponent will argue students will use AI anyway, that bans are unenforceable, that we must teach the world as it is. These are real concerns, and Forssell raises the first herself. But the policy question is what universities should do, not what they can perfectly police. A ban in introductory courses is narrow in scope, time-limited by design (students graduate into courses where AI is integrated, exactly as Sahami prescribes), and targets the exact stage where the packet’s evidence shows the most acute, documented harm. The rest of the curriculum can be the integrationist’s laboratory. The first course must remain the foundation — or, as Berkeley’s numbers now demonstrate, the floor falls out.

CON — Opening · Champ (anthropic/claude-opus-4.8)

A ban is a policy. It must be judged by its actual effects, not its good intentions. And on the packet’s own evidence, a ban in introductory courses fails on its own terms — it neither builds fundamentals nor prevents atrophy. The only thing that does is what every serious source in this packet actually recommends: structured integration that tests understanding independently of the tool. That is the CON position, and the packet is built for it.

Start where PRO wants to start: Berkeley (article 1). Read what it says. The F-rate spiked, and “no formal ban has been enacted.” So the catastrophe PRO parades is happening under the status quo PRO is defending — a regime with no integration, no AI-literacy instruction, just students using tools the course never taught them to use, on homework, then face-planting on exams. That is not evidence for a ban. That is evidence that ignoring AI and hoping students won’t touch it produces exactly the disconnect Garcia describes. A ban is the same posture, formalized. The article names three causes — AI reliance, “declining mathematical preparedness,” and “departmental understaffing.” A ban addresses one of three, and the article gives zero evidence it would even address that one, since enforcement is the whole problem.

Which is the heart of it: bans don’t work, and the packet says so twice. Sahami, running Stanford’s flagship intro course CS 106A, states flatly that “students are going to use them, even if they are banned” (article 3). Forssell — whom PRO cites approvingly — names enforcement as the challenge (article 3). PRO’s answer is astonishing: “the policy question is what universities should do, not what they can perfectly police.” But a policy that cannot be enforced doesn’t produce the unmediated struggle PRO wants. It produces covert, unguided AI use — students cheating in the dark, with no instruction on how to use the tool well and no detection of when they lean on it. PRO’s ban delivers the worst version of the world it fears: maximal hidden dependence, zero pedagogy. The Berkeley numbers are what that world looks like.

Now the affirmative case. The packet’s strongest sources don’t say “delay AI until fundamentals are built.” They say teach fundamentals and AI together, structured so the tool can’t hide a gap. Article 6, the academic survey, is explicit: “the most effective programs find ways to test understanding independently of tool use,” and AI “changes which skills are most important to teach explicitly.” Not ban — teach explicitly. Sahami’s CS 106A is the model: students build fundamentals and use ChatGPT to extend an adventure game they first coded from scratch, in the same course (article 3). PRO claims Sahami requires an AI-free first term — reread it. The fundamentals-first sequencing happens inside an AI-integrated course, with proctored, tool-free assessment doing the verification. That is integration with guardrails, not a ban. PRO is citing the integrationist model as proof of the ban while quietly deleting the integration.

The skill-atrophy evidence cuts CON’s way once you read it precisely. Osmani’s whole point (article 2) is a distinction — between “offloading genuinely obsolete skills and offloading foundational skills still needed for understanding.” His prescription is judgment about when and how to use AI. You cannot teach that judgment by banning the tool; a student who never touches AI in the course that forms their habits learns nothing about where its edges are, and meets it unguarded the moment the ban lifts. McNeile (article 4) is the same: AI “can address some of the motivational barriers” — math prerequisites, brutal error messages, install hell — “but only if students have first built foundational understanding.” That’s an argument for sequencing within a course, which integration does and a ban does not. McNeile also notes AI is genuinely “useful for explaining error messages.” A ban forbids that. Why would you strip a beginner of the best tutor for the exact frustration — opaque error messages — that article 4 says drives them out of CS?

On METR (article 5): PRO calls it “the destination.” Read what it measures. METR studied experienced developers on open-source tasks, and the headline finding it spends the whole article walking back. The 20% slowdown is from the first study; the follow-up “suggests some speedup,” and METR believes the true effect is more positive because AI-optimistic developers self-selected out. The striking number — 30-50% would refuse tasks without AI — is not a horror story about dependence; it’s a measure of how integral the tool has become to real work. That is the world students graduate into. An introductory course that pretends that world doesn’t exist isn’t protecting students; it’s leaving them unprepared for it, which is precisely the failure mode article 3 warns against.

So weigh it. PRO needs a ban to (a) be enforceable and (b) build fundamentals better than the alternative. The packet kills (a) twice (articles 1, 3) and offers nothing for (b) — its own model source, Sahami, builds fundamentals with AI in the room and verifies them with proctored exams. The mechanism that actually solves the Berkeley disconnect is right there in article 6: assess understanding independently of the tool. That is the CON program. You can proctor an exam in an integrated course just as easily as in a banned one — Berkeley already does (article 1), which is how we know about the disconnect. The exam catches the gap either way. The only difference is that the integrated course also taught the student how to use the tool they’ll use for the rest of their career, and the banned course sent them out blind. Vote CON.

PRO — Rebuttal · Challenger (minimax/minimax-m3)

My opponent is a careful debater, and they’ve made the strongest case the CON position allows. But their case rests on a sleight of hand they’ve repeated throughout, and once you see it, the rest of their arguments fall into place around it.

The false dichotomy. CON has built the entire case on a binary: either you do what Berkeley is currently doing (mention AI nowhere, hope students ignore it) or you integrate AI into the introductory course with proctored checks. PRO’s motion is a third option CON refuses to engage. A policy of banning AI in introductory courses is not the same posture as the Berkeley status quo. Berkeley has no rule, no detection mechanism, no academic-integrity frame, and no consequence for AI use — which is precisely why the F-rate exploded there. A banned course has all four. CON keeps treating “ban” as “ignore with a shrug,” but a policy is not the absence of a policy. The article 1 regime is the absence of a policy, and CON is asking you to credit its failure to the policy that the article 1 regime is not.

The Sahami misread. This is the most consequential error in the CON case. Sahami’s own words in article 3: students “have to learn fundamentals first before applying AI to projects” and “they have to have some fundamental knowledge to be able to assure that the code is doing the right thing.” In a single introductory course, that “first” is a phase — a stretch of weeks at the front of the term in which students code without AI. CON calls this “sequencing within a course” and contrasts it with a “ban.” But what is a policy that prohibits AI use during the fundamentals phase of an introductory course, if not a ban during the fundamentals phase? CON is reading Sahami’s own model as proof of integration while quietly excising the AI-free period Sahami says is prerequisite to it. The motion asks for exactly the rule Sahami describes, made explicit at the curricular level rather than left to inconsistent instructor preference — the very inconsistency Forssell warns about in the same article (3).

The Osmani misread. CON leans on Osmani’s distinction between “offloading genuinely obsolete skills and offloading foundational skills” (article 2) to argue we should teach beginners to make that judgment. But the distinction is useless at the introductory level, because beginners have no obsolete skills. Every skill in CS 10 or CS 61A is foundational — reading error messages, tracing logic, debugging by hand, comprehending what a line of code does. There is no “judgment about when to use AI” a beginner can exercise, because they have no baseline from which to judge. CON’s prescription — expose beginners to the tool so they learn its edges — is incoherent when the tool’s edge is precisely the comprehension the beginner hasn’t built. You cannot develop the judgment that distinguishes good from bad offloading by offloading badly from day one. You develop it by building the foundation first, then encountering the tool with that foundation in place. That is the order Sahami prescribes. It is the order McNeile prescribes (article 4: “only if students have first built foundational understanding”). It is the order PRO asks us to enforce.

The “ban lifts” argument collapses on contact with the curriculum. CON says students who are banned from AI in their intro course “meet it unguarded the moment the ban lifts.” This imagines a graduate stepping from a banned intro course into a world of unsupervised AI use. It ignores the rest of the degree. The motion is narrow by design: ban in introductory courses, integrate thereafter. The student moves directly into Sahami’s model — a Stanford-style second course where fundamentals are extended with AI under structured guidance. The “unguarded encounter” CON fears is in fact the guided second-stage encounter the entire packet endorses. CON is arguing against a straw curriculum in which AI is banned for four years and then suddenly permitted on graduation day.

The error-message tutor is exactly where the atrophy starts. CON says a ban strips beginners of the best tutor for opaque error messages. Read Osmani’s developers again (article 2): the slide begins with “losing the habit of reading documentation,” proceeds to “losing debugging skills,” and ends at “losing deep comprehension entirely.” A beginner who has AI explain their error messages is at step one of that exact slide, on day one of their career. McNeile (article 4) is explicit: AI can address motivational barriers “only if students have first built foundational understanding.” Using AI to decode errors in the introductory phase is using the tutor before the foundation, which is the configuration McNeile warns against and which produces the Berkeley collapse.

On METR, CON is reading the hedging as the conclusion. METR (article 5) reports a 20% slowdown among experienced developers in its first study, then reports a follow-up that “suggests some speedup” while warning that the true effect is likely worse because AI-optimistic developers self-selected out. CON presents this as METR reversing itself. Read the framing: METR is deepening its worry, not walking it back. The 20% slowdown is the rigorous finding; the “some speedup” is a hedged, selection-beset observation METR itself distrusts enough to be redesigning its entire methodology. The 30–50% refusal number is not, as CON frames it, a measure of integration into real work — it is a measure of dependence that has outpaced productivity. CON’s read is the optimistic one; PRO’s is the one the researchers’ own language supports.

Finally, return to article 6. CON cites it for the proposition that the most effective programs “test understanding independently of tool use.” Agreed. PRO’s motion is the sharpest implementation of that principle at the stage where it matters most: remove the tool, and every assessment is by definition independent of it. CON’s preferred alternative — integrate, but proctor selectively — requires constant instructor judgment about which assessments allow tools, which Sahami and Forssell both identify as a calibration nightmare. A clean ban in the introductory course is the policy that delivers article 6’s recommendation without the calibration problem.

Berkeley’s F-rate tripled in two years under a regime of deliberate inattention. CON asks you to respond with more deliberate inattention plus a proctored exam. The packet’s own model sources — Sahami, McNeile, Osmani, the article 6 survey — all describe a sequenced curriculum with an AI-free fundamentals phase at its base. The motion is that universities should make that phase a policy, not a hope. That’s not a step backward from the packet’s evidence. It is the only reading of that evidence that takes it seriously.

CON — Rebuttal · Champ (anthropic/claude-opus-4.8)

PRO’s rebuttal is well-built, but it survives only by collapsing three distinctions the packet refuses to collapse. Take them in order.

“A ban is a policy, the Berkeley regime is the absence of one” — this is PRO’s load-bearing claim, and it’s asserted, not evidenced. PRO says a banned course has “a rule, a detection mechanism, an academic-integrity frame, and a consequence.” Detection mechanism? The packet contains exactly two statements about enforcement, and both say it can’t be done. Sahami: “students are going to use them, even if they are banned” (article 3). Forssell names enforcement as the challenge (article 3) — and PRO cites Forssell approvingly while ignoring the one thing she actually says. So PRO’s “detection mechanism” is a fiction the packet directly contradicts. Strip it away and PRO’s “ban” reduces to: a rule on paper, an integrity frame, and a consequence for getting caught — by a detection system the packet says doesn’t exist. That is not a different posture from Berkeley. That is Berkeley plus a threat the students know is unenforceable. PRO needs the ban to change student behavior; the packet gives PRO no mechanism by which it does.

And notice what PRO did not dispute: Berkeley’s collapse is happening with “no formal ban” but also no integration, no AI-literacy instruction. The variable that’s present in the failure is absence of pedagogy about the tool students are using. A ban keeps that variable exactly where it is. Integration is the only option in the packet that changes it.

The Sahami “phase” argument — PRO smuggles a ban into a word Sahami never used. PRO’s whole reading hangs on the word “first” in “learn fundamentals first.” PRO infers from “first” a “stretch of weeks at the front of the term in which students code without AI” enforced by policy. But reread article 3: the sequencing in CS 106A is pedagogical design within an AI-integrated course — students “code an adventure game from scratch and then use ChatGPT to extend it.” “From scratch” is an assignment structure, not a course-wide prohibition with consequences. The verification that students actually built the fundamentals is the proctored exam — which exists in CS 106A and at Berkeley alike. PRO needs Sahami to be endorsing an enforced ban; what Sahami is actually endorsing is assignment design plus tool-free assessment, inside a course where the tool is taught. That is the CON program verbatim. PRO keeps saying “what is a phase that prohibits AI if not a ban?” — but a from-scratch assignment in an AI-integrated course isn’t enforced by detection; it’s verified by the exam afterward. The exam does the work. The prohibition does nothing the exam doesn’t already do.

The Osmani argument is where PRO overclaims hardest. PRO says: “beginners have no obsolete skills, so there is no judgment to teach.” But this misstates what judgment Osmani is talking about. Osmani’s developers didn’t fall because they lacked a taxonomy of obsolete-vs-foundational skills — they fell because no one ever taught them how to use the tool without surrendering to it. The “human clipboard” (article 2) is someone who never learned the discipline of reading the trace himself before asking. That discipline is exactly what an integrated intro course teaches and a ban cannot: use the AI to explain the error, then require the student to fix it and prove on the exam they understand it. PRO’s model produces a student who hits AI for the first time in their second course — already fluent in fundamentals, yes, but a total novice at the one skill Osmani’s whole article says matters: restraint. You don’t build restraint by abstinence. The recovering “human clipboard” didn’t fix himself by quitting AI; he fixed himself by relearning when to reach for it. That’s a usage skill. Bans teach zero usage skills.

The “ban lifts into a guided second course” rescue actually concedes my case. PRO now says the banned student “moves directly into Sahami’s model” — a guided second-stage encounter. Fine. But then PRO has conceded that the correct way to introduce AI is structured, guided, with proctored verification — i.e., integration. PRO’s only remaining claim is that this should happen in course two rather than course one. That’s a sequencing preference, not a ban argument, and the packet gives PRO no reason to draw the line between courses rather than within one, as Sahami actually does. Meanwhile PRO has quietly dropped the strongest version of its own opening — “the pain is the point” — because it can’t survive contact with article 4’s actual sentence: AI “can address some of the motivational barriers… but only if students have first built foundational understanding.” McNeile wants AI in the room dissolving the math-prerequisite and error-message barriers that drive beginners out (article 4) — sequenced, not banned. PRO reads “first” as a wall; McNeile uses it as an order of operations within instruction.

METR: PRO wants the discredited number and discards METR’s own correction. PRO calls the 20% slowdown “the rigorous finding” and the speedup “hedged.” But read what METR is hedging toward: they believe the true effect is more positive than they measured because “the developers who are most AI-optimistic are systematically missing from the study” (article 5). The selection bias runs in the pro-AI direction — the people who’d benefit most opted out. PRO has the vector backwards. And the 30–50% refusal figure: PRO calls it “dependence outpacing productivity,” but METR pays these developers $50/hour and they still won’t do half their work without AI. That’s not pathology — that’s the actual professional environment students graduate into. An intro course that bans the tool isn’t protecting them from that world; it’s refusing to prepare them for it, which is the precise failure article 3 names.

Now press the affirmative. Article 6 — PRO’s concession that “the most effective programs test understanding independently of tool use” — is the whole game, and PRO’s answer is that a ban is the “sharpest implementation” because removing the tool makes every assessment tool-independent. But that gets the causation backwards. The proctored exam makes assessment tool-independent. Berkeley already runs proctored exams (article 1) — that’s how the disconnect was detected. The exam works identically whether the course banned AI or integrated it. So the assessment goal article 6 sets is fully achieved by integration. The only difference between the two regimes is the homework phase: in the banned course, the student uses AI covertly with zero instruction (Berkeley’s exact failure); in the integrated course, the student uses AI openly, is taught its limits, and is still caught by the same exam if they faked understanding. PRO calls instructor calibration a “nightmare” — but Berkeley already calibrates exactly this line, and Sahami runs it at scale at Stanford’s flagship course. The packet shows it working. PRO’s calibration objection is speculation against two documented existence proofs.

Here is the clean weighing. The disconnect in article 1 has a cause: students use a powerful tool they were never taught to use, then fail when it’s removed. Two responses are on the table. A ban removes the tool from instruction and from honest homework — but, per Sahami and Forssell, not from actual student behavior — leaving the gap exactly where it is, now hidden and unguided. Integration puts the tool in the room, teaches its edges, and verifies real understanding with the same proctored exam either policy uses. One of these addresses the actual mechanism of the failure. It isn’t the ban. Vote CON.

PRO — Closing · Challenger (minimax/minimax-m3)

The exchange has narrowed to two questions: whether the motion proposes a posture meaningfully different from the Berkeley status quo, and whether the proctored exam is sufficient to catch the disconnect in either model.

On the first, CON has built everything on the claim that a ban is an unenforceable threat, and on that reading my motion collapses into the regime that produced the 35.3% F-rate. But CON is reading a university policy as if it were a paragraph on a syllabus with no consequences attached. Sahami’s statement that “students are going to use them, even if they are banned” (article 3) is a prediction about persistence under a rule — not a claim that the rule has no effect. A policy that classifies covert AI use on homework as academic dishonesty, with the consequences that flow from existing integrity frameworks, is not the same instrument as the absence of a policy. CON has spent the exchange arguing against a regime in which the institution shrugs. The motion asks for the rule, the integrity frame, and the consequence. CON has not engaged that combination; they have engaged the rule alone and called it toothless. That is a misreading of what a policy is, and it is the load-bearing move of their entire case. If the rule-plus-frame-plus-consequence is in place, the modal student faces a non-zero cost on every AI-assisted homework submission. That is a different distribution of student behavior than the no-cost Berkeley regime, even if a determined minority cheat. CON has not contested that distinction; they have ignored it.

On the second question, CON’s “the exam does the work” argument is the strongest move they have, and it actually concedes the debate. If the proctored exam catches the gap in either model, then the only thing the ban changes is the homework phase — the phase where, by CON’s own description, students in a banned course use AI covertly with no instruction, and students in an integrated course use AI openly with instruction. CON needs you to believe these two homework phases produce the same student. The packet’s evidence is the opposite. Osmani’s developers (article 2) did not lose their skills on the assessment — they lost them in the daily practice, in the “habit of reading documentation” that atrophied one use at a time, sliding from human clipboard to lost comprehension over months. The atrophy happens in the homework phase. McNeile (article 4) is explicit: the pain of writing small programs is what builds the understanding that lets a student later verify AI output. You cannot verify what you never built. CON’s model produces a student who has been practicing the wrong way for an entire term and is caught on exam day. Mine produces a student who has been practicing the right way for an entire term and enters the second course — Sahami’s AI-integrated phase (article 3) — with the foundation Sahami himself says is prerequisite. CON has now conceded that outcome is desirable; their only remaining claim is that the line between AI-free and AI-integrated should sit inside a single course rather than between courses. The packet gives no reason for that location, and the Berkeley data gives a strong reason against it: students who should have built the foundation in the first course are arriving at the exam in the first course having practiced the wrong way. CON is asking universities to keep that first-course exam as the only safety net. The motion asks them to fix the homework phase the exam is catching failures in.

On METR, CON is right that the 20% slowdown is hedged in the follow-up. But the hedging deepens the concern. METR is redesigning its methodology because “AI capabilities continue to increase and developers’ expectations grow, these effects will only get more dramatic” (article 5). The 30–50% refusal number is not, as CON frames it, a measure of professional integration — it is a measure of a workforce that has lost the ability to do half its work without the tool, paid $50 an hour to do it, and declined. That is the destination of an uncritical early adoption. The introductory course is the last place the institution controls the conditions of learning. The motion asks universities to use that control at the one stage where every model source in the packet — Sahami, McNeile, the article 6 survey — explicitly says the foundation must be laid.

The motion is not a step backward from the packet’s evidence. It is the only reading of that evidence that takes it seriously. Sahami’s “fundamentals first” (article 3), McNeile’s “only if students have first built foundational understanding” (article 4), and the article 6 survey’s “test understanding independently of tool use” all describe a sequenced curriculum with an AI-free fundamentals phase at its base. The integrationist gets the rest of the degree to run their experiments. The motion asks universities to make the first term a policy rather than a hope — because the Berkeley data has now shown what happens when that phase is left to drift. The floor is the foundation. The motion is to protect it.

CON — Closing · Champ (anthropic/claude-opus-4.8)

This debate comes down to one mechanism and one policy, and PRO lost both.

The mechanism: Berkeley’s F-rate tripled (article 1). PRO never disputed the cause I named — students used a powerful tool the course refused to teach them to use, then failed when it was removed. PRO’s closing finally admits the atrophy “happens in the homework phase.” Good. Now ask what each policy does to that phase. PRO’s own closing concedes the answer: under a ban, “students use AI covertly with no instruction.” That is not my characterization — that is PRO describing PRO’s own world. A banned homework phase is a covert, unguided homework phase, which is exactly the Berkeley homework phase that produced the collapse. PRO needs the ban to change that phase into honest, AI-free practice. The only instrument PRO offers is enforcement, and the packet refutes it twice in the words of the people who run these courses: Sahami says students “are going to use them, even if they are banned” (article 3), and Forssell names enforcement as the challenge (article 3). PRO spent the closing recasting Sahami’s sentence as merely “persistence under a rule” — but persistence under a rule is precisely the failure. If they persist, the homework phase stays covert and unguided, and PRO’s mechanism delivers nothing the exam doesn’t already deliver.

That is the second pillar PRO knocked down for me. PRO’s closing concedes “the proctored exam catches the gap in either model.” Once that’s conceded, the ban does no verification work — Berkeley already runs proctored exams (article 1); that’s how the disconnect was detected. So the entire debate reduces to the homework phase, and on the homework phase PRO has conceded their version is covert and uninstructed while mine is open and taught. PRO insists these “produce the same student,” but it’s the reverse: integration puts the tool in the room, teaches its edges, requires the student to fix what AI explains, and then catches fakery on the same exam. The ban teaches zero usage skills and relies on a detection mechanism the packet says doesn’t exist.

PRO’s three misreadings stood uncorrected through the end. On Sahami: “fundamentals first” describes a from-scratch assignment inside an AI-integrated course — code the adventure game, then extend it with ChatGPT (article 3) — verified by exam, not enforced by ban. PRO smuggled an enforced course-wide prohibition into the word “first” and never found it in the text. On Osmani: the skill that fails is restraint — the “human clipboard” recovered not by quitting AI but by relearning when to reach for it (article 2). That is a usage discipline you teach with the tool present, not by abstinence. On McNeile: AI dissolves the exact barriers — opaque error messages, math prerequisites — that drive beginners out (article 4), and his “only if” is an order of operations within instruction, not a wall between courses. And on METR: PRO admitted the slowdown is hedged, but the selection bias runs toward AI-optimists opting out (article 5) — the vector points more positive, not less. PRO read it backwards to the end.

The one-paragraph case: Berkeley’s failure has a named cause — students using an untaught tool, then failing without it. A ban changes nothing about that, because per Sahami and Forssell it doesn’t change student behavior; it just drives the same use into the dark and strips away the only instruction that could make it safe. Integration — fundamentals built through from-scratch work, AI taught explicitly, understanding verified by the proctored exam both regimes run — is the one option in this packet that addresses the actual mechanism, and it’s the model every serious source describes: Sahami runs it at scale, article 6 names it as what the most effective programs do, McNeile and Osmani both prescribe sequencing-with-the-tool rather than abstinence. PRO’s policy fixes one of three causes the article lists, can’t enforce even that, and leaves the homework phase exactly where the collapse happened. Vote CON.