CAL NEWPORT AND ED ELSON: Schools should admit language-model tools only where they serve the goal of training sustained thought, which means very little use at most levels

The Gist

Newport says school is mainly for training minds to think, read, and write, and AI that writes for students skips the workout, so there should be little AI in most classrooms. Math and reading scores did fall from 2022 to 2025, and heavy AI users scored lower, but that doesn't prove AI caused it, and some well-designed AI tutors have helped students learn. This packet steelmans the arguments made by Cal Newport and Ed Elson in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

Schools should judge language-model tools against the purpose of building students' capacity for sustained thought, and on that test general chat tools that do students' reading and writing for them should be used very little at most levels, because they remove the effort through which that capacity develops.

Premises

1. A central aim of schooling is to train the mind for sustained symbolic thinking through reading, writing, and logic, rather than to teach workplace software.
2. Writing is among the most demanding cognitive tasks and a core of curricula, and language-model tools can do the writing for students, removing the effort through which the learning happens.
3. Using current chat tools requires little skill, so the case for early 'AI fluency' training is weak, much as schools do not teach elementary students workplace software.
4. Average PISA scores across OECD countries fell between 2022 and 2025 in mathematics and reading, a period in which generative AI spread among students.
5. In PISA 2025, students who did not use AI chatbots generally scored higher in science than frequent users.
6. In a randomized trial with Turkish high school mathematics students, unrestricted access to a general chatbot raised practice scores but left students performing worse than controls once access was removed.
7. Each educational technology should be judged by whether it advances or impedes the curriculum's purpose, and New York City adopted a one-year moratorium on student-facing generative AI in grades up to eighth for 2026 to 2027.

Assumptions

• The PISA associations are correlational and do not isolate AI from smartphones, the pandemic, or other causes.
• Purpose-built tutoring tools with guardrails are a different case from general chat tools that produce answers.
• PISA science scores were broadly stable between 2022 and 2025, so the faster decline in the AI period applies to mathematics and reading only, and PISA does not show that AI caused it.
• Randomized and quasi-experimental studies have found learning gains from designed, supervised AI tutoring, so the case for very little use applies to general tools that do the work rather than to tutoring designed for learning.

Analysis

Argument type: Inductive.

Potential Fallacies

• Scope extrapolation / hasty generalization (P6 to Conclusion) — A single randomized trial of high school mathematics students in one country supports a policy claim about reading and writing across most grade levels. The inference is made plausible by the effort-and-learning mechanism, but it is not directly tested in younger grades or in writing instruction. Effect sizes from single studies also tend to shrink on replication.
• Post hoc suggestion (mitigated by stated assumptions) (P4 and P5) — Placing the PISA decline beside the spread of generative AI invites a causal reading, and the usage gap in P5 is open to reverse causation and selection effects. A1 and A3 disclaim causation, so these premises work as context rather than proof. Their persuasive impression may still exceed their evidential weight, particularly because both come from the same dataset and share the same confounders.
• Weak analogy / conflating ease of use with competent use (P3) — Comparing chat tools to workplace software assumes that low entry skill means little need for instruction. Using a chatbot is easy, but verifying outputs, spotting errors, and knowing when not to use it may be teachable skills with real value. The analogy supports at most delaying fluency training in the youngest grades, not the broader claim.
• Appeal to institutional action (P7) — The NYC moratorium confirms that a policy was adopted, not that it is correct. A one-year precautionary measure through grade 8 reflects decision-making under uncertainty. It is also narrower than 'most levels'.
• Unargued value premise treated as common ground (P1 and P7) — The purpose test depends on sustained symbolic thought being a central aim of schooling. This is a reasonable and widely held view, and the premise is hedged as 'a central aim', but workforce readiness, equity, and civic aims are not weighed. Because the conclusion is largely built into the criterion, it matters that this premise is contestable.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Effort removal depends on how a tool is used and how work is assessed, not on the tool itself. With in-class writing, oral defenses, and scaffolded or Socratic uses, general models can give feedback at scale and preserve effort. The trial's own guarded-tutor arm suggests design matters more than tool class, so a flat 'very little use' rule forgoes gains.
• Conclusion (High impact) — School-level restriction is hard to enforce because students have chatbots on personal phones and at home. Restriction may push use into unsupervised settings with no guidance, and it may widen equity gaps if affluent families teach AI use at home while schools abstain.
• P6 (High impact) — One trial in one country and one subject cannot establish effects on reading and writing or on younger students. The result may depend on the specific tool interface, short exposure, and how students were instructed, and single-study effect sizes often shrink on replication.
• Premise 3 (Medium impact) — Effective AI use (verification, error detection, knowing when not to offload) is a non-trivial skill that may benefit from early, structured practice. The software analogy understates this, and schools do teach typing, search, and source evaluation despite their apparent ease.
• Premise 1 (Medium impact) — Schooling also serves workforce readiness, civic participation, and equity. If students will enter AI-saturated workplaces, the purpose test may yield a more permissive verdict, and delay could disadvantage students without home exposure.
• Premise 4 (Medium impact) — PISA declines began before generative AI and coincide with smartphones, social media, pandemic disruption, and attendance changes. The stability of science scores cuts against a simple AI story, so these premises add little.
• Premise 2 (Medium impact) — Calculators, spell-checkers, and search engines prompted similar worries about lost effort and were later integrated productively with staged introduction. Without a limiting principle, the effort-removal logic could exclude many tools. The principle needs to distinguish tools that perform the graded cognitive task from tools that scaffold it.

Suggested Improvements

• Define the scope — Specify what 'very little' and 'most levels' mean, for example by grade band (K-8, upper secondary, tertiary), by task (graded writing versus feedback), and by setting (in class versus homework). The vagueness makes the claim hard to test and lets different audiences hear different claims. The evidence (high school math, 15-year-olds, grades through 8) does not cover the stated range.
• State the bridging principle explicitly — Add a premise such as: tools that perform the graded cognitive task should be restricted while the underlying skill is being built, and tools that scaffold it may be permitted. Add a functional test for telling the two apart. This makes the inference from mechanism to policy explicit, supplies a limiting principle against over-restriction, and handles vendors blurring 'general' and 'guarded' tools.
• Strengthen the evidence base for writing and reading — Add direct evidence from writing and reading studies, such as trials comparing delayed unaided performance after AI-assisted versus unassisted writing, and replications of the Turkish trial across ages and subjects. The load-bearing mechanism in P2 is currently supported by inference and by a trial in a different domain.
• Reframe or drop the weak premises — Present the PISA findings explicitly as background context and the NYC moratorium as an example of precaution, rather than as support. Replace P3's analogy with an account of what AI fluency involves and why it can be delayed. Stacking correlational evidence inflates apparent support. Engaging the strongest version of the AI-literacy view makes the case more charitable and more robust.
• Address implementation, enforcement, and equity — Pair the recommendation with assessment redesign (in-class, handwritten, or oral work), a plan for out-of-school use, and provisions for accessibility, language learners, and students lacking outside support. School restrictions alone may leave offloading in place at home, and the policy could widen gaps if these costs are ignored.

Scenario Tests

• A replicated randomized trial in middle-school writing finds that students given unrestricted chatbot access write better with it but perform worse on later unaided writing. (Supports) — This would close the main evidential gap by testing the mechanism in the domain and age range the conclusion names, and would strongly support restricting answer-producing tools.
• A well-designed study finds that teacher-structured use of a general chatbot for draft feedback improves later unaided writing compared with no AI. (Challenges) — The harm would depend on the use pattern rather than the tool class. The argument would then need a functional criterion rather than a blanket 'very little' rule, though the narrowed claim about answer-producing use would survive.
• Causal analysis attributes most of the 2022–2025 PISA decline to smartphones and post-pandemic disruption rather than AI. (Neutral) — The argument's logic survives because it does not rely on PISA for causation (A1, A3). It confirms that P4 and P5 serve only as context, and it removes some of the rhetorical force.
• A school adopts in-class handwritten and oral assessment, so graded work can no longer be offloaded, and then permits wider chatbot use. (Challenges) — If assessment redesign neutralizes substitution, the effort-removal mechanism no longer follows from tool availability. The real issue becomes assessment design, which weakens the case for restricting tools as such.
• Elementary students who are restricted from general chatbots until grade 9 show stronger foundational writing and reading, and are as capable as peers at critical AI use within a year of introduction. (Supports) — This would support both the effort-protection claim and the weak-case-for-early-fluency claim in P3, and would back a staged, calculator-style policy.
• Evidence shows that restricted-school students use chatbots heavily at home with no guidance, and that skills do not improve. (Challenges) — School-level restriction would then fail to protect the capacity for sustained thought and would forgo guided instruction. The case for pairing restriction with assessment and literacy measures would strengthen.

Coherence & Relevance

The argument is coherent at its core: a purpose-based criterion (P1, P7), a mechanism of effort removal (P2), and one causal experiment (P6), with the claim narrowed by A2 and A4 to general, answer-producing tools. The hedged conclusion fits the inductive support. The weaknesses are structural: the bridge from mechanism to policy is left implicit, the experimental evidence is narrower than 'most levels' and reading and writing, and P3, P4, P5 and the moratorium clause in P7 contribute context more than inferential support. Overall it is a reasonable precautionary case that is credible for the narrow claim and under-evidenced for its full breadth.

• A central aim of schooling is to train the mind for sustained symbolic thinking through reading, writing, and logic, rather than to teach workplace software. (Strong) — Supplies the evaluation criterion but is not defended against rival aims. The conclusion depends on it, so its contestability transfers directly.
• Writing is among the most demanding cognitive tasks and a core of curricula, and language-model tools can do the writing for students, removing the effort through which the learning happens. (Strong) — Asserts the mechanism without direct evidence from writing instruction. Assumes students substitute rather than scaffold, and that the removed effort is the productive kind.
• Using current chat tools requires little skill, so the case for early 'AI fluency' training is weak, much as schools do not teach elementary students workplace software. (Weak) — Addresses a related but distinct question (early fluency instruction) rather than whether student use impedes learning. Conflates ease of operation with competent use.
• Average PISA scores across OECD countries fell between 2022 and 2025 in mathematics and reading, a period in which generative AI spread among students. (Weak) — Temporal coincidence only, with no mechanism or dose-response evidence. Aggregate decline does not show individual usage drove it, and A1 and A3 disclaim causation.
• In PISA 2025, students who did not use AI chatbots generally scored higher in science than frequent users. (Weak) — A correlation in a different subject from the argument's focus on reading and writing. Subject to reverse causation and selection, and does not distinguish general tools from guarded ones.
• In a randomized trial with Turkish high school mathematics students, unrestricted access to a general chatbot raised practice scores but left students performing worse than controls once access was removed. (Strong) — The most directly relevant causal evidence, but one study in one country and subject at one age band. Extension to reading, writing, and younger students rests on the mechanism in P2.
• Each educational technology should be judged by whether it advances or impedes the curriculum's purpose, and New York City adopted a one-year moratorium on student-facing generative AI in grades up to eighth for 2026 to 2027. (Moderate) — The criterion is relevant and portable. The moratorium shows precedent, not justification, and it covers fewer grades than 'most levels'.

View this argument on LogicFirst.ai