📄 Full text: Stanford SCALE · local
The single most consistent finding in the 2026 Stanford SCALE review: pedagogically designed, tutoring-specific AI consistently outperforms general-purpose chatbots on durable learning outcomes.^stanford-evidence-base-ai-k12-2026
The Core Distinction
| Dimension | General-Purpose AI (e.g., ChatGPT, Gemini) | Tutoring-Specific AI |
|---|---|---|
| Interaction model | Open-ended Q&A; completes tasks on request | Structured hints, Socratic questioning, step-by-step scaffolds |
| Cognitive load | Reduces all load, including germane (productive) load | Reduces extraneous load while preserving productive struggle |
| ZPD targeting | Often operates outside the zone of proximal development | Explicitly calibrated to learner readiness |
| Metacognitive demand | Low — AI does the reasoning | High — learner must reason with guidance |
| Transfer evidence | Mixed to negative when tool is removed | More promising (limited causal data) |
Evidence from the Causal Literature
General-Purpose AI: Mixed or Negative Transfer
- Bastani et al. (2025): High schoolers using a general-purpose chatbot for math practice scored ~17% worse on closed-book final exams than peers with no AI access, despite higher practice grades.^stanford-evidence-base-ai-k12-2026
- Lehmann et al. (2025): General-purpose AI for programming increased topics covered but harmed understanding and widened achievement gaps for low-prior-knowledge students.
- Stadler et al. (2024): General-purpose AI produced lower-quality reasoning and argumentation vs. traditional search.
- Kosmyna et al. (2025): AI essay assistance led to 83% of participants failing to recall a quote from their own essay, vs. 11% for non-AI users.
Tutoring-Specific AI: Better Outcomes
- Bastani et al. (2025): A tutoring-specific chatbot with pedagogical guardrails (hints, step-by-step reasoning, misconception targeting) mitigated the exam score drop observed with general-purpose GPT. General-purpose GPT Base caused the drop; the tutoring variant prevented it.
- Kreijkes et al. (2026): Retention improved when AI use was paired with traditional strategies like note-taking — suggesting that friction-preserving designs matter.
Why This Happens: Learning Science Mechanisms
1. Desirable difficulties — General-purpose AI removes productive struggle; tutoring tools preserve it via graduated hints. 2. Germane load — Effective learning requires processing that feels effortful. General AI short-circuits this. See cognitive-load-theory. 3. Metacognition suppression — When AI completes reasoning, students lose practice in monitoring their own understanding. 4. Expertise reversal — Novices need scaffolding, not answers. General AI gives answers; tutoring AI gives scaffolds.
Important Caveats
- The causal comparison base is tiny (most studies are single-condition AI-access vs. no-access, not head-to-head tutoring vs. general).
- "Tutoring-specific" is not yet a standardized design category — implementations vary widely.
- Long-term transfer data (months or years out) is essentially absent.
Implications for Practitioners
- For tool selection: Favor products with explicit pedagogical guardrails (hints, Socratic mode, step-by-step requirements) over raw LLM access.
- For policy: School/district procurement criteria should distinguish between "AI-integrated" tools (tutoring-specific) and "AI-access" tools (general chatbox).
- For research: Head-to-head RCTs comparing pedagogically designed AI vs. raw LLM access on delayed post-tests are urgently needed.
Related Pages
- difficulty-aware-dialogue-kt — General LLMs reframed as psychometric instruments through IRT mapping
- genai-meta-analysis-programming-learning — Productivity gains vs. learning outcomes across general AI tools
- nsmq-riddles-science-math-benchmark — LLMs underperform top students, reinforcing specialized tutoring needs
- moodle-ai-tutoring-deep-learning — Shows how general LLMs can be scaffolded into tutoring roles
- critical-thinking-genai-scaffolding — Vendrell & Johnston (2026): eight design principles for scaffolding critical thinking with LLMs in higher education.
- ai-learning-companions-framework — three-foundation framework for AI learning companions prioritizing durable learning over performance
- ai-tutor-behavioral-evaluation — behavioral evaluation axis for AI tutors — measuring what students actually do with feedback
- ai-literacy — Students should know whether they interact with a general LLM or pedagogical system
- ai-k12-evidence-base — The broader evidence landscape
- ai-learning-transfer — Durability of gains when AI is removed
- multimodal-ai-tutoring — Structured dialogue as tutoring-specific intervention in STEM
- collaborative-ai-tutoring — ProPACT's proactive dyadic scaffolds (tutoring-specific design)
- knowledge-tracing-irt — Explicit ability/difficulty calibration for tutoring
- metacognition — How tutoring-specific design preserves metacognitive demand vs. general AI suppression
- self-regulated-learning — The SRL framework and AI's role in scaffolding or displacing regulation
- ai-tutor-safety-harms — SafeTutors taxonomy: pedagogical harms from "helpful" general-purpose AI
- llm-fallacy-misattribution — Students misattributing AI-generated outputs as their own competence
- pedagogical-llm-training — Training methods (EduQwen) that align models with tutoring-specific design
- educational-llm-alignment — Why general LLMs fail on tutoring impact despite benchmark success
- educational-vlm-evaluation — Visual tutoring-specific design for STEM work
- agentic-workflows-education — Agentic paradigms that embed tutoring-specific guardrails
- zone-of-proximal-development — (create when second source emerges)
- quantum-education-its — Domain-specific vs. general-purpose tutoring design
- ai-metacognition-stem-review — Domain-specific tutoring (ITS) vs. generic chatbots for metacognitive scaffolding
- hybrid-human-ai-tutoring-differentiated — Human component remains differentiable alongside AI