We Must Stop Using AI to ‘Level-Down’ Our Students  🔗 **Source:** [We Must Stop Using AI to ‘Level-Down’ Our Students](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) > [!note]+ Document Notes > tags: [ai-in-education, literacy-instruction, text-leveling, science-of-reading, edtech-procurement, special-education] > author: [MISSING] > source: EdSurge > reference: https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students > date created: > > # 🏷️ Title: We Must Stop Using AI to ‘Level-Down’ Our Students > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [EdSurge — We Must Stop Using AI to ‘Level-Down’ Our Students](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) - [Timothy Shanahan — Leveled Reading: Leveled Lives](https://www.shanahanonliteracy.com/) - [Natalie Wexler — The Knowledge Gap](https://naturalwexler.com/) - [Diffit](https://web.diffit.me/) ## 📝 Summary - A New Jersey literacy coach adopted an AI text-leveling tool (Diffit) to rewrite chapters of "All Quiet on the Western Front" down to third-grade summaries for eighth-grade special education students, then reversed course when results flopped. - It matters because AI makes the leveling shortcut nearly free, right as many states legislate away from leveling under Science of Reading policy. Cheap automation can re-institutionalize a practice the research already rejected. - Stakeholders: classroom and special education teachers, literacy coaches, district curriculum buyers, LDTCs, edtech vendors, and the students on the receiving end of a simplified text. - Risks and tradeoffs: leveling buys short-term teacher capacity (differentiation is the most time-expensive part of the job) and pays for it with student growth, engagement, and equity of access to the canon. The counter-approach — scaffolding a hard text — costs more teacher time and demands more instructional skill. - Bottom line: the tool isn't the problem. Point AI at scaffolds (think-aloud scripts, chunking, vocabulary sets, fluency passage selection) instead of at the text itself. ## 📖 Main Content - **The core argument.** Text leveling means matching students only to texts at their measured "instructional level" — a seventh grader reading at third-grade level gets third-grade books. The author argues that decades of reading research show this caps growth, and that what students actually need is a complex text plus a teacher modeling how a strong reader attacks it. - **The new failure mode.** Leveling used to be labor-intensive, which limited it. Generative AI removes the labor. A teacher can paste a chapter and get a simplified summary in seconds, so the practice spreads through classrooms faster than policy can push it out. - **The lived case.** The author, a literacy coach in 2025, was told by a respected Learning Disabilities Teaching Consultant to run "All Quiet on the Western Front" through Diffit for eighth-grade special education students. He deferred to seniority against his own training. The tutored student — dyslexic, ADHD — showed no interest, waited to be told answers, and retained nothing across sessions. - **The reversal.** He returned to re-reading the real text, sequence and character mapping, theme discussion, vocabulary work, background knowledge building, and spaced/interleaved retrieval practice. The student joined general-education peer discussions of Paul Bäumer's psychology. Three years on, the same novel is a student favorite in those classrooms. - **The equity frame.** He names leveling as an inequity generator: some students get the literary canon, others get a machine summary of it. That reframes an accommodation decision as an access decision. - **The Faust structure.** The essay is organized as Faustian bargain → fallout → redemption, with AI cast as Mephistopheles — a powerful servant, dangerous master. - **Terms defined.** *Text leveling*: assigning texts by measured reading level. *Scaffolding*: temporary supports that keep the original text intact. *Science of Reading*: the body of evidence on how reading develops, now driving state legislation. *Fountas and Pinnell / Accelerated Reader / Guided Reading*: elementary leveling systems named as the popularizers. - **Uncertainties and contested claims.** The evidence here is one coach's practice narrative with n=1 tutoring plus classroom impressions — no control group, no assessment data. "Decades of research" is asserted by link, not summarized. The piece also stops short of saying when leveled text is legitimately appropriate (decoding-stage readers, newcomer English learners), which is where the real professional disagreement sits. And the article both criticizes Diffit's core function and praises it for scaffolds — defensible, but it does soften the critique of the vendor whose primary marketed feature is the thing being condemned. ### 🗒️ Notes - **Opening move.** The Ponyboy rewrite ("Keep being good and special, Ponyboy. Don't change.") is a demonstration, not an anecdote — it shows the loss rather than describing it, in one line, before any argument starts. - **The purchase moment.** "In my own district, just two years ago, educators raved about a new AI tool that helps teachers level text and encouraged the department to spend thousands of dollars on it." Procurement enthusiasm preceded pedagogical review. - **Why the shortcut is seductive.** The author names differentiation for students who don't get it after the third teaching as the most time-consuming part of his job. The tool promised to absorb exactly that cost. - **The tradeoff stated plainly.** "This program was going to do it all for me, but the caveat was that it would also do all the thinking for the kids." - **Policy vs. market.** States legislate toward Science of Reading and away from leveling programs; edtech vendors ship leveling inside new AI models. The two vectors point opposite directions. - **Where the pushback came from.** Eighth-grade teachers, not administrators — "It's too hard and too boring. Our dyslexic and ADHD students will never be able to read Shakespeare, Chaucer, or Remarque." - **Structural cause, not teacher blame.** Special education teachers had "too many preps and too large a caseload to adequately understand these new mandates." The leveling shortcut filled a capacity gap the district created. - **Authority deference.** He questioned the practice out loud ("Isn't that like…frowned upon?"), got overridden by a veteran LDTC's decades of experience, and complied. That's the mechanism by which a bad practice survived a correct instinct. - **Failure signature.** No interest, waiting to be told answers, no cross-session retention, no classroom improvement reported by the special education teacher. Four independent signals, all soft. - **What replaced it.** Re-reading the real text, sequence/character mapping, theme discussion, vocabulary exploration, background knowledge, all under spaced and interleaved retrieval practice. - **The five sanctioned AI uses.** (1) Identify passages that will trip students up and teach syntax/vocabulary from them. (2) Chunk, not change, passages — use "/" to separate phrases as embedded fluency scaffolds. (3) Identify passages suitable for fluency practice. (4) Generate think-aloud scripts for teachers who aren't strong at modeling. (5) Build word sets for morphological or connotative study. - **Licensing and funding disclosure.** Part of an EdSurge series on educator experiences, supported by the Learning Commons, editorial control retained by EdSurge, published CC BY-NC-ND 4.0. Reusable with attribution, no derivatives. ## 📚 References - [We Must Stop Using AI to ‘Level-Down’ Our Students — EdSurge](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) - [Shanahan on Literacy (Timothy Shanahan)](https://www.shanahanonliteracy.com/) - [Diffit](https://web.diffit.me/) - [Creative Commons BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/) - Wexler, Natalie. *The Knowledge Gap.* (book — no canonical primary URL cited in source) - Remarque, Erich Maria. *All Quiet on the Western Front.* (novel referenced throughout) - Hinton, S.E. *The Outsiders.* (source of the "Stay gold, Ponyboy" line) ## 🗣️ Paraphrase - Simplifying a famous sentence strips what made it worth reading. That's what leveling does to whole texts — it trades the experience for readability. - A district bought an AI leveling tool on enthusiasm. The tool worked as advertised, which was the problem: it did the thinking the students were supposed to do. - Research says students grow on complex texts paired with a teacher who models and scaffolds. Leveling gives them neither the complexity nor the modeling. - The author was told by a senior colleague to run a war novel through an AI leveler for special education eighth graders. He doubted it, deferred anyway, and watched a student disengage completely. - Going back to the real text with mapping, discussion, vocabulary, background knowledge, and spaced retrieval turned that same student into a participant in general-education discussions. - The student never needed an easier book. She needed a different route through the hard one, and more time to walk it. - Handing some students the canon and others a machine summary is an equity failure dressed as an accommodation. - AI belongs on the scaffolds — hard-passage detection, chunking, fluency selection, think-aloud scripts, word sets — not on the text. ## 📰 Newsletter Here's a story that should sound familiar to anyone who's ever watched a plugin solve the wrong problem beautifully. A New Jersey literacy coach bought into an AI tool that rewrites hard books into easy ones. It worked. His dyslexic eighth-grader stopped caring, stopped remembering, and started waiting for him to hand over the answers. He pulled the plug, went back to the actual novel with real scaffolding, and three years later "All Quiet on the Western Front" is the favorite read of the year in those classrooms. The line worth stealing: use AI to chunk, not change. That's a design principle far beyond the classroom. Every time you point automation at the artifact instead of the path to the artifact, you delete the part where someone learns something. Same reason a docs site that answers every question with a generated summary produces developers who can't read your source. Worth your time if you build learning content, write documentation, or ship anything with a "simplify this for me" button. The piece is CC BY-NC-ND 4.0, so quote it freely with attribution. - [Read the article](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) - [Shanahan on Literacy](https://www.shanahanonliteracy.com/) — the research backdrop - [Diffit](https://web.diffit.me/) — the tool in question, praised for its scaffold features ## ✅ Action Items #### DevRel - Developer advocacy: build a talk on "scaffold, don't simplify" — AI that supports comprehension of real source vs. AI that replaces it. Measure by post-talk adoption of doc-reading behaviors, not satisfaction scores. - Education: audit every workshop for the leveling pattern — where do you hand learners a simplified stand-in instead of the real API response, real error message, real config file? - Developer experience: keep quickstarts pointed at real artifacts. Add chunking, annotation, and glossary layers around raw output rather than paraphrasing it away. - Developer marketing: stop shipping "we make it simple" as the only value claim. Test "we make hard things reachable" as the alternative message. - Community management: run a survey on where community members feel talked down to by generated content. - Internal/external/cross-project: agree a shared rule with docs and support teams — AI summarizes navigation, never replaces the canonical text. #### OSPO - Policy & Governance: write an AI-content policy that separates generation-of-scaffolds from generation-of-substitutes. Decision rights sit with maintainers, not tooling vendors. - License Compliance: note this source is CC BY-NC-ND — no derivatives. Any internal reuse quotes and links; it does not remix. - Security Management: not applicable to this source — the article raises no vulnerability, disclosure, or supply-chain concern. - Software Inventory: add procured AI tools to the inventory with a "what does this replace?" field. The district bought a leveler before anyone asked that. - Developer Support: offer consultation hours on AI tool evaluation so individual contributors aren't overridden by seniority alone, as the author was. - Community Engagement: fund upstream documentation work rather than downstream summary generation. - Strategic Alignment: add a KPI for comprehension depth, not just time-to-first-success. Time-to-first-success rewards leveling. #### Community - Maintainers: run a README clinic — is the README a scaffold into the code, or a substitute for reading it? - Contributors: build first-issue paths that require reading real source, with annotation support rather than pre-digested explanations. - Sponsors: fund the expensive path (mentorship, pairing) since that's the analog of scaffolding, and it's what gets cut first. - Ecosystem: coordinate on shared guidance for AI-generated docs across adjacent projects. #### WordPress - Core: triage AI-adjacent proposals against the "supports thinking vs. replaces it" test. - Core Performance: not applicable to this source — no performance dimension. - Core AI: adopt an explicit guardrail — AI features scaffold user comprehension of what WordPress is doing, never hide it. Build an evaluation set from that principle. - **Core Program Team:** add an AI-use retro question to cross-team rituals: what did the tool do *for* people that they needed to do themselves? - Mobile: review any AI summarization in the app for the same substitution risk. - Playground: build a reproducible lab that shows a scaffolded vs. simplified learning path over identical source code. - Community: add "AI in teaching" to meetup kits with the chunk-don't-change framing. - Training: audit Learn WordPress lessons for over-simplification. Where a lesson paraphrases core code, consider annotating the real code instead. - Marketing: not applicable to this source — no campaign hook worth building on an education op-ed. - Support: check whether AI-drafted replies teach the asker or just hand over an answer. - Design: apply the same test to UI copy — does it explain or does it obscure? - Documentation: examples-first patterns already match the scaffolding model. Add hard-passage annotation to DevHub gaps. - Photos: not applicable to this source. - TV: caption quality is the scaffold equivalent for video — hold the line on real transcripts over generated summaries. - Polyglots: translation is not leveling. Keep register and complexity faithful to the original in glossary alignment. - Accessibility: important nuance — this article argues that simplifying content is *not* an accessibility accommodation. Keep WCAG conformance (structure, contrast, navigation) separate from content dilution. - Meta: not applicable to this source. - CLI: keep `--help` output complete rather than trimmed; add examples as the scaffold. - Hosting: not applicable to this source. - Tide: not applicable to this source. - Openverse: education tie-ins could carry the scaffolding framing. - Themes: not applicable to this source. - Plugins: not applicable to this source. #### The WPOCC - Programs: fund a maintainer-education fellowship built on scaffolded reading of real codebases. - Projects: run a documentation sprint that annotates rather than rewrites. - Funding experiments: milestone payouts tied to comprehension outcomes, not artifact count. ### FAIR - Technical independence: local metadata and provenance let you verify what an AI tool actually changed — the same audit need the article implies for leveled text. - Policies and covenants: anti-capture language should cover vendor-driven pedagogy, where the product's default becomes the practice. - Productized governance artifacts: a scorecard field for "does this tool replace user cognition?" belongs in operator agreements. ## 🧪 Hypotheses - Students taught a grade-level complex text with structured scaffolding outperform matched peers given AI-leveled versions on delayed comprehension and retention measures, even when the leveled group scores higher on immediate readability-matched checks. - The availability of one-click AI leveling increases the rate of leveling in districts whose written policy prohibits it, because tool default beats policy text. - Deference to seniority is the primary mechanism by which evidence-contradicting practices persist in schools — coaches who doubt a practice comply anyway when the recommender has more years of experience. - Special education caseload size predicts AI leveling adoption better than teacher beliefs about reading science do. - Students who receive AI-simplified texts show measurably lower rates of self-initiated question-asking than students on scaffolded complex texts. - Teacher-facing AI use (think-aloud scripts, word sets, passage selection) produces larger student gains than student-facing AI use of the same tool. ### 📊 Methodology - Cluster-randomized trial at the classroom level: complex-text-plus-scaffold vs. AI-leveled text, same novel, same instructional minutes. Primary outcome is delayed (6- and 12-week) comprehension on the original text; secondary outcomes are vocabulary retention and discussion participation coded from recordings. - Difference-in-differences on district AI-tool procurement dates vs. observed leveling frequency in lesson plans, using states that passed Science of Reading legislation as the policy contrast. - Survey plus vignette experiment on authority deference: present coaches with a practice recommendation attributed to a novice vs. a veteran and measure compliance rates. - Administrative-data regression: caseload and prep counts against tool telemetry for leveling requests, controlling for teacher credentialing. - Discourse analysis of classroom transcripts for student-initiated questions across the two conditions. - Within-teacher crossover: same teacher uses AI for scaffolds in unit A and for leveling in unit B, with student outcomes compared to control for teacher quality. ### ❓ Questions - Where is the legitimate floor for leveled text — decoding-stage readers, newcomer multilingual learners — and who decides it in a district? - If the vendor's primary marketed function is the harmful one, does praising its secondary features actually change district purchasing behavior? - What assessment would have caught this failure faster than "she didn't remember the content from one session to the next"? - How do you write a procurement rubric that distinguishes scaffolding features from substitution features when both ship in the same product? - Does the same substitution risk apply to adults using AI summaries of technical documentation, and if so, what's the professional equivalent of the scaffold? - What happens to the teacher-capacity problem that leveling was solving? Scaffolding costs more time, not less. ### 🔮 Future Research - Longitudinal tracking of students taught under leveling-heavy vs. scaffolding-heavy regimes into high school reading outcomes. - Instrument design: a validated observation protocol that distinguishes "AI produced a scaffold" from "AI produced a substitute" in lesson artifacts. - Cost modeling of scaffolding at scale — what staffing ratio makes complex-text instruction sustainable for special education caseloads? - Vendor behavior study: do edtech companies change default features in response to state Science of Reading legislation, or only marketing language? - Transfer study on whether the substitution effect appears in adult professional learning (documentation, compliance training, onboarding). - Equity analysis: which student populations receive AI-simplified content at disproportionate rates? ### 💡 Implications - **Practice:** the differentiation decision moves from "which text?" to "which route through this text, and how long?" That's a scheduling and staffing question as much as a pedagogical one. - **Policy:** legislation that names banned *programs* will lag tools that ship the same practice as a feature. Policy needs to name the practice, not the product. - **Security:** not applicable to this source — the article raises no security dimension. - **Education:** teacher training has to include AI-use judgment, since the article's failure came from a well-intentioned expert recommendation, not from ignorance. - **Funding models:** grant-funded editorial series (Learning Commons here) can carry practitioner reversal stories that vendor-funded content structurally won't. Preserving that channel matters. - **Procurement:** add a mandatory question — what human cognitive work does this tool absorb, and is that work the point of the activity? ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid Grade Level roughly 11–12; Flesch Reading Ease roughly 45–52. Long compound sentences, education-domain jargon (instructional level, Science of Reading, LDTC, interleaved retrieval practice), and literary allusion (Faustus, Mephistopheles, The Matrix) push it to college-entry difficulty. Mildly ironic for an article about access. #### 👀 Visual learners - Present: one stock news image; no diagrams, charts, or screenshots. - Improve: a before/after side-by-side of the original Remarque passage and the leveled version; a flowchart of the five sanctioned AI uses; a chunking example showing the "/" phrase separation the author describes but never displays. #### 🔊 Aural learners - Present: no audio version, podcast, or narration. - Improve: a 3-minute author-read audio summary; TTS on the article body; a short podcast segment pairing the author with the LDTC he disagreed with. #### 🗣️ Verbal learners - Present: strong — the Ponyboy rewrite is a definition by demonstration, and the Faust/Mephistopheles frame carries the whole argument. Terms like text leveling get explicit examples. - Improve: a glossary for LDTC, Fountas and Pinnell, interleaved retrieval practice; a plain-language callout distinguishing scaffolding from leveling in one sentence each. #### 🧩 Kinesthetic learners - Present: the five-item list of AI uses is actionable but not demonstrated. - Improve: a copy-paste prompt library for each of the five uses; a downloadable chunked passage to try in class tomorrow; a one-lesson pilot protocol. #### 🧠 Logical learners - Present: clear narrative causality — bargain, fallout, redemption — with a stated mechanism (effortful thinking builds learning). - Improve: a table comparing leveling vs. scaffolding across cost, evidence, growth, and equity; a decision tree for when a support is a scaffold vs. a substitute; the actual research citations rather than a single "decades of research" link. #### 🧑‍🤝‍🧑 Social learners - Present: EdSurge series framing invites practitioner response; no comments or discussion prompts in the text. - Improve: department-meeting discussion guide; a peer-review rubric for AI-supported lesson plans; moderated Q&A with special education leads. #### 🧘‍♀️ Solitary learners - Present: implicit reflection through the author's own reversal. - Improve: journaling prompts ("where have I deferred against my training?"); a self-audit checklist of your last five AI-assisted lessons. #### 🌿 Naturalistic learners - Present: not applicable to this source — no environmental or sustainability content appears. - Improve: frame instructional capacity as an ecosystem — what gets depleted when scaffolding time is cut, and what regrows when it's funded. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: EdSurge uses standard article headings ("The History of Text Leveling," "The Faustian Bargain," "The Fallout," "The Redemption") and a news landmark structure. - Improve: the extracted text shows section headings rendered as plain paragraphs, which suggests visual styling may not map to real `h2` elements. Verify heading levels programmatically. The trailing "News image" needs descriptive alt text, not a filename. #### Keyboard navigation - Present: standard editorial page, no custom widgets in the article body. - Improve: confirm visible focus on the many inline links and a skip-to-content link ahead of site navigation. #### Contrast and readability - Present: standard EdSurge editorial styling; body text appears to meet 4.5:1. - Improve: check the caption and byline text, which are typically the failure points at small sizes; verify line-height at 1.5 for body copy given long paragraphs. #### Text resizing and reflow - Present: responsive news template. - Improve: verify 200% zoom without clipping on pull quotes and the embedded link-heavy paragraphs, several of which break mid-sentence in the extracted text. #### Media and captions - Present: no video or audio, so no caption obligation. - Improve: if the series adds audio, ship downloadable transcripts and avoid autoplay. #### Forms and interactions - Present: not applicable to this source — the article carries no forms or interactive controls in the content body. - Improve: newsletter signup and share controls elsewhere on the page need labels and 24×24 CSS pixel targets. #### Standards reference - Likely gaps map to WCAG 2.2 AA: 1.3.1 Info and Relationships (headings that look like headings but may not be), 1.1.1 Non-text Content (the "News image" alt), 1.4.3 Contrast (byline/caption text), 1.4.12 Text Spacing, 2.4.7 Focus Visible, and 2.5.8 Target Size (Minimum) for share controls. - Content note: this article argues against simplifying text as an accommodation. That's consistent with WCAG — 3.1.5 Reading Level is AAA and recommends *supplementing* difficult content, not replacing it. ### 🌍 DEIB Assessment #### Cultural breadth - Present: single-district, single-state (New Jersey), U.S.-only frame. The literary canon named is Western and largely European (Shakespeare, Chaucer, Remarque, Hinton). - Improve: examples from multilingual-learner contexts and non-U.S. systems; acknowledge that "the great literary canon" is itself contested ground; link translated and global case studies. #### Inclusive language - Present: person-adjacent phrasing is inconsistent — "our dyslexic and ADHD students" is quoted from teachers, but the author also writes "a dyslexic, ADHD student I tutored" in his own voice. - Improve: identity-first vs. person-first language varies by community preference; state which convention you're using and why. Define LDTC and Fountas and Pinnell on first use — both are insider jargon. #### Neurodiversity support - Present: chunked into four named sections with a numbered list of recommendations; consistent narrative structure. - Improve: the paragraphs are long and the headings are far apart. More frequent subheadings, a summary box up top, and icon+text labels on the five recommendations would help. The article's own advice — chunk, don't change — applies to itself. #### Representation - Present: the student at the center is a dyslexic eighth grader with ADHD who succeeds. Special education teachers are treated sympathetically as structurally overloaded rather than blamed. The LDTC is credited with expertise even while being disagreed with. - Improve: the student never speaks. Quoting her, or the special education teacher, would balance a piece that currently runs entirely through one coach's perception. No racial, linguistic, or socioeconomic dimension is named, though leveling's equity effects are heavily patterned along those lines. #### Psychological safety - Present: strong — the author indicts his own judgment rather than the teachers he coached, and names the systemic cause (caseloads, preps) before any individual failing. That makes it safe to admit the same mistake. - Improve: a content note for the war-trauma discussion; explicit guidance for teachers who disagree with a senior colleague, since that's exactly the situation the article describes without resolving procedurally. ### ✅ Logical Validity #### Strong reasoning - **Causal, "The Faustian Bargain" through "The Fallout":** the author states a mechanism (effortful thinking produces learning), predicts a consequence, applies the intervention, observes the predicted failure, reverses, and observes recovery. That's a coherent causal chain with a named mechanism rather than a bare correlation. Clarify further by reporting what was measured and when. - **Analogical, "The Redemption":** the Faust/Mephistopheles analogy does real work — it names the specific failure mode (losing control of a servant force) rather than decorating the argument. "Faustus' tragedy was not that he sought power, but that he ultimately lost control of the force meant to do his bidding." Strengthen by extending it to procurement: Faustus signed a contract too. - **Deductive, opening paragraphs:** if simplification destroys what makes a text worth reading, and leveling is simplification, then leveling destroys what makes a text worth reading. The Ponyboy demonstration makes the first premise near-undeniable in one line. - **Inductive, policy section:** state legislation moving away from leveling plus edtech shipping leveling in AI models supports the generalization that tooling outruns policy. Modest sample, sound direction. #### Fallacies - **Appeal to authority (accepted, then correctly rejected).** "this was a well-respected educator with decades of experience working with students with disabilities; surely she knew better than I did." The author names this as the error, so it's reported rather than committed — but the essay never proposes a replacement decision procedure. Fix: state the rule (evidence over tenure) and give coaches a concrete escalation path. - **Hasty generalization.** One tutored student and one district's teacher impressions carry the empirical weight for a categorical claim ("We Must Stop"). Fix: pair the narrative with the actual research it gestures at, or scope the claim to "in my experience, and consistent with X." - **Anecdotal evidence in the reversal.** "Three years later, those same teachers... are no longer surprised when 'All Quiet on the Western Front' is their students' favorite read of the year." Favorite-read status isn't a comprehension measure. Fix: add assessment data or label it as engagement evidence only. - **False dilemma (mild).** The framing sets leveled text against complex text with scaffolds, omitting mixed approaches, staged text sets, and the decoding-stage cases where simplified text is standard practice. Fix: name the boundary conditions explicitly. - **Suppressed tension / special pleading.** Diffit is condemned for its "primary function" and then praised — "Diffit does an incredible job of providing these resources." Not strictly a fallacy, but it protects the vendor from the conclusion the argument reaches. Fix: state plainly that the recommendation is to buy the tool and disable or avoid its headline feature. - **Loaded language.** "atrocity," "sterile and lifeless simplicity," "Faustian tragedy" do persuasive work that the evidence doesn't fully fund. Fix: keep the imagery, add the data. #### Tone of voice - Confessional and self-implicating. "My gut and everything I had learned about teaching reading told me this was wrong... so I trusted the veteran educator at the time." He's the one who got it wrong, and he says so first. - Literary and allusive — Hinton, Remarque, Marlowe's Faustus, The Matrix — signaling in-group membership with English teachers. - Urgent and prescriptive in the title and closing ("We Must Stop," "a powerful servant, but a dangerous master"), softened in the middle by structural empathy for overloaded teachers. - Practitioner-to-practitioner, not researcher-to-reader. The five recommendations read like staff-meeting advice. ### 🛡️ DARVO Analysis - **Deny:** no denial pattern. The author admits the harm and locates himself as the cause. The closest thing to minimization is the treatment of Diffit — the vendor's core function is condemned in the abstract while the product is praised by name, which softens accountability for the tool's design defaults. That's deflection toward "misuse," not denial: "The fault lies not in the tool itself, but in how we train ourselves as educators to misuse it." - **Attack:** none directed at people. The LDTC who gave bad advice is described as "well-respected" and never named. Teachers who pushed back are quoted, but their objection is answered with evidence rather than character attack. Edtech companies are criticized as a category, which is the sharpest edge in the piece. - **Reverse victim/offender:** absent, and notably so. The obvious available move — casting himself as the victim of bad district advice — isn't taken. He calls it "the veteran teacher's advice" and then says "After taking the veteran teacher's advice, I recommended text leveling to the teachers I coached." Responsibility stays with him. - **Emotional cues and rhetoric:** conflict is internalized rather than externalized ("The world didn't freeze, per se, but in that moment, it certainly felt like I was living in 'The Matrix'"). The rhetorical heat goes to the practice and the market, not to individuals. - **Related fallacies:** the tool/misuse split is a mild single-cause simplification — it exempts product design from responsibility for what its defaults encourage. Worth flagging because it's the one place accountability gets moved off-stage. ## 🎯 Conclusion - Bottom line: AI made text leveling free, and free is what makes a rejected practice spread. The fix isn't to ban the tool — it's to point it at the scaffolds (hard-passage detection, chunking, fluency selection, think-aloud scripts, word sets) and leave the text alone. Students don't need an easier book; they need a different route through the hard one and more time to walk it. - Next steps for readers: audit your last five AI-assisted lessons or docs for substitution vs. scaffolding. Add a "what human thinking does this absorb?" question to any tool purchase. If you're a coach or lead, write down the rule that evidence outranks tenure, before you're standing in the hallway being told otherwise. - Next steps for builders: if your product's headline feature replaces the reader's work, expect it to be adopted for exactly that, no matter what your best-practice docs say. Defaults are the policy. ## 🔄 Refactor Appendix — We Must Stop Using AI to ‘Level-Down’ Our Students > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. > > Konik claim-header rules: present tense, declarative, ≤10 words when > possible, searchable, unique, mirrors annotation focus if present. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Atoms (H3) ### [H1] Removing effortful thinking removes the learning > [!quote] ID:rw_h1 > When we keep kids from actually doing work that requires effortful thinking, they don’t learn anything at all. **Context & use**: - Section: opening, immediately after the district's AI leveling tool purchase. - This is the mechanism the whole essay rests on. It generalizes past classrooms — any tool that absorbs the cognitive work also absorbs the growth. Applies directly to AI-generated documentation summaries. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) <!-- atomicity: 1 idea ✓ --> ### [H2] Blame educator training, not the tool itself > [!quote] ID:rw_h2 > The fault lies not in the tool itself, but in how we train ourselves as educators to misuse it. **Context & use**: - Section: end of "The Fallout," transitioning into "The Redemption." - The hinge of the argument, and its weakest joint — it exempts product defaults from responsibility. Useful as a claim to argue *with*, since defaults shape practice more than training does. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) <!-- atomicity: 1 idea ✓ --> ### [H3] Struggling readers need a route, not an easier text > [!quote] ID:rw_h3 > My student didn’t need a different text; she just needed a different roadmap and timeframe to get there. **Context & use**: - Section: "The Fallout," after reverting to scaffolded instruction on the original novel. - Reframes differentiation from content substitution to path and pacing. This is the transferable design principle — same artifact, different route. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) <!-- atomicity: 1 idea ✓ --> ### [H4] Chunk the passage; never change the passage > [!quote] ID:rw_h4 > Use AI to chunk, not change, the passages. **Context & use**: - Section: "The Redemption," recommendation two of five. - The single most portable rule in the piece. Six words that separate scaffolding from substitution, and they apply to docs, onboarding, and training content unchanged. **Tags**: #xref #share **Highlight**: [view highlight](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) <!-- atomicity: 1 idea ✓ --> ### [H5] Leveled text creates inequity of canonical access > [!quote] ID:rw_h5 > why would we give some students access to the great literary canon, and others a robot-created summary? **Context & use**: - Section: "The Redemption," first paragraph on responsible AI use. - Converts a pedagogy debate into an access-equity claim. The rhetorical question form makes it quotable but also unfalsifiable — pair with data when reusing. **Tags**: #xref #addendum **Highlight**: [view highlight](https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Automation that absorbs the effort absorbs the learning **Claim** (one sentence, Courtney's voice): When a tool does the difficult cognitive work an activity exists to produce, it deletes the outcome while appearing to deliver it. **Evidence pointer**: [[#H1]] > [!quote] > When we keep kids from actually doing work that requires effortful thinking, they don’t learn anything at all. **Scope & conditions**: - Holds when: the difficulty *is* the point — reading comprehension, learning an API, debugging, writing to think. - Breaks when: the difficulty is incidental toil with no learning payload — boilerplate, formatting, dependency bumps, transcription. Automating those is pure gain. **Links**: - [[Desirable difficulty]] — because extends - [[AI-assisted documentation]] — because applies-in developer education - [[Text leveling failure case]] — because evidence-for <!-- atomicity: 1 idea ✓ --> #### [P2] Scaffolds change the route; substitutes change the artifact **Claim** (one sentence, Courtney's voice): A legitimate support alters how someone reaches the original material, never what the material is. **Evidence pointer**: [[#H4]] > [!quote] > Use AI to chunk, not change, the passages. **Scope & conditions**: - Holds when: the original artifact is the learning target — a novel, a codebase, an API response, a spec. - Breaks when: the artifact is itself a poor proxy for the goal, in which case replacing it is the correct move, not a substitution failure. **Links**: - [[P1 — Automation that absorbs the effort absorbs the learning]] — because extends - [[Progressive disclosure]] — because contrasts on whether the full artifact stays reachable - [[Docs quickstart design]] — because applies-in developer experience <!-- atomicity: 1 idea ✓ --> #### [P3] Product defaults outrank stated best practice **Claim** (one sentence, Courtney's voice): A tool's headline feature becomes the practice regardless of what its guidance recommends. **Evidence pointer**: [[#H2]] **Scope & conditions**: - Holds when: the harmful path is faster and the correct path costs more human time — exactly the leveling-vs-scaffolding gap. - Breaks when: strong enforcement sits outside the tool (audit, review gate, compliance check) and actually bites. **Links**: - [[Defaults as policy]] — because extends - [[Educator training as the fix]] — because contrasts on where responsibility sits - [[Procurement rubric]] — because applies-in tool evaluation `proof_state: working` ## Tests / falsifiers - Find districts with written anti-leveling policy plus an AI leveling tool in place; if leveling frequency stays flat after adoption, defaults aren't driving practice. - If vendor-published best-practice guidance measurably shifts feature usage away from the headline feature, the claim weakens. <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Where does my own work substitute a generated summary for the real artifact? Audit the last five AI-assisted docs pages. — https://www.edsurge.com/news/we-must-stop-using-ai-to-level-down-our-students #fleeting #inbox - [F2] Read Timothy Shanahan's "Leveled Reading: Leveled Lives" — cited as the eye-opener that reframed the author's practice. #fleeting #inbox - [F3] Natalie Wexler, "The Knowledge Gap" — named as the prior benchmark for reading-education writing. #fleeting #inbox - [F4] Article seed: "Chunk, don't change" as a documentation design principle — the leveling failure mapped onto developer docs and onboarding. #articleseed #fleeting #inbox - [F5] Open question: what's the professional-learning equivalent of leveled text, and are AI meeting summaries it? #fleeting #inbox - [F6] Tension worth writing about: WCAG 3.1.5 recommends supplementing hard content, not replacing it — accessibility and this article agree, but "simplify for accessibility" is widely misread. #xref #fleeting #inbox - [F7] Procurement rubric line item to draft: "what human cognitive work does this tool absorb, and is that work the point?" #share #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Audit the last five AI-assisted docs or lesson artifacts for substitution vs. scaffolding — content strategy #things - [ ] Draft the "chunk, don't change" post applying text-leveling research to developer documentation — writing #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/We Must Stop Using AI to ‘Level-Down’ Our Students.md|We Must Stop Using AI to ‘Level-Down’ Our Students by edsurge.com]] ## Highlights - This program was going to do it all for me, but the caveat was that it would also do all the thinking for the kids. When we keep kids from actually doing work that requires effortful thinking, they don’t learn anything at all. ([View Highlight](https://read.readwise.io/read/01ky53p53k3zzjn0nkjj13w7qh)) - Note: Central thesis: removing effortful thinking removes the learning. - Tags: ai-in-education, desirable-difficulty, atom/literature/h1 - Once my student began working to understand the dense text, the payoff was immediate. Our discussions of Paul Bäumer’s psychology and the trauma of war were profound. This dyslexic, eighth-grade student with ADHD, whom many had already given up on, started to join her peers in these discussions back in the general education classroom. My student didn’t need a different text; she just needed a different roadmap and timeframe to get there. ([View Highlight](https://read.readwise.io/read/01ky53kqcmhmebvdft3h9ghpvv)) - Note: The case supports preserving the destination while changing the roadmap and timeframe. - Tags: educational-equity, scaffolding, atom/literature/h1 - Once my student began working to understand the dense text, the payoff was immediate. Our discussions of Paul Bäumer’s psychology and the trauma of war were profound. This dyslexic, eighth-grade student with ADHD, whom many had already given up on, started to join her peers in these discussions back in the general education classroom. My student didn’t need a different text; she just needed a different roadmap and timeframe to get there. ([View Highlight](https://read.readwise.io/read/01ky53pam4v93xy5qg9g70vzk6)) - Note: The load-bearing evidence: accommodation as different path, not lower ceiling. - Tags: scaffolding-vs-substitution, accessibility, atom/literature/h1 - The fault lies not in the tool itself, but in how we train ourselves as educators to misuse it. ([View Highlight](https://read.readwise.io/read/01ky53p6ygtra33cj48gsz8m6a)) - Note: Zettel candidate — transfers directly to DevRel arguments about AI tooling. - Tags: atom/zettel/z1, tool-neutrality - A recurring theme in the use of AI in education is that it must be used critically and responsibly, especially given its impact on the student experience and its interactions with current pedagogy. In this use case, irresponsible AI use actually leads to inequities; why would we give some students access to the great literary canon, and others a robot-created summary? ([View Highlight](https://read.readwise.io/read/01ky53pdfrp9dfzpw04dbb4gpf)) - Tags: equity, ai-in-education, atom/literature/h2 - - Use AI to chunk, not change, the passages. Take it a step further by using “/” to[separate phrases](https://www.researchgate.net/publication/247597151_Spacing_printed_text_to_isolate_major_phrases_improves_readability) and create embedded fluency scaffolds that actually prove comprehension. ([View Highlight](https://read.readwise.io/read/01ky53krb5tbfpmq8fq8nbhe9b)) - Note: AI can expose structure without replacing the source task. - Tags: ai-in-education, reading-instruction, atom/literature/h1 - To avoid a Faustian tragedy, we must remember that AI is our Mephistopheles: a powerful servant, but a dangerous master. It should support thinking, creation, and access — not replace them. ([View Highlight](https://read.readwise.io/read/01ky53ksbyqhk5gnmkdh1jc9kr)) - Note: The closing image is rhetorically strong; the scaffold-versus-substitute test carries the actual argument. - Tags: atom/literature/h3, fallacy/loaded-analogy, ai-in-education