AI study tools: notes, flashcards & spaced repetition from own materials

Students spend hours turning slides and lectures into notes and flashcards, and most never review on a schedule that fits how memory works. AI study tools run retrieval over the learner's own corpus, generate notes, cards, questions and audio summaries with citations, and schedule reviews with a memory model such as FSRS. Typical benefit: less time on card making, more time on active recall. Strong private use; usually learner-driven, outside the institution.

Problem & value

• Today: learners highlight, re-read and hand-write cards in Anki or Quizlet. Card creation is the bottleneck, and uniform review intervals either waste time or let items decay.
• AI value: (1) conversion of heterogeneous material into study objects in minutes; (2) answers and questions grounded in the learner's own sources with page/timestamp citations; (3) a scheduler that estimates recall probability per card and orders reviews to hit a target retention.
• Evidence: spaced retrieval practice is among the best-supported learning techniques in the cognitive-science literature [1]. Anki adopted FSRS as a built-in scheduler option, and the open-spaced-repetition project publishes benchmark comparisons against older SM-2-style schedulers [2][3]. Controlled studies of AI-generated study aids are still limited; do not rely on vendor learning-gain claims.
• Real use: NotebookLM (Google), Quizlet's AI features and Anki with FSRS are widely used by students [4][5].

Variants

• Private student (main case): upload lecture PDFs, get cards and quizzes on phone.
• Exam preparation / certification: medical, law, IT or professional exams, with large decks and deadlines.
• Language learners: vocabulary cards from own reading and audio, with sentence mining (see EDU-011).
• Institution-provided: university licenses a tool for students; instructor-curated course sources (overlaps EDU-001, EDU-004).
• Corporate self-study: employees convert manuals and policies to review decks (see EDU-006).
• Accessibility: audio overviews and simplified summaries for learners with reading difficulties.

How it works

upload (PDF/PPTX/scans/audio/video/web)
  -> parsing: OCR + layout, ASR for recordings (with timestamps)
  -> chunking + embeddings -> private vector index per learner/notebook
  -> generation (RAG): notes, cards, questions, audio script  [every item cites its chunk]
  -> learner review/edit/delete (human-in-the-loop)
  -> scheduler: memory model (FSRS: difficulty, stability, retrievability) per card
  -> daily review queue -> ratings logged -> parameters re-fitted to the learner

• Grounded generation: retrieval restricts the model to the uploaded sources; cards link back to slide/page so errors can be checked.
• Card quality: prompts follow "one fact per card", cloze deletion, and image occlusion for diagrams. Dedupe against the existing deck.
• Scheduler: FSRS is an open-source memory model with a small set of parameters fitted to a user's review history (no deep net required, runs on-device). Knowledge tracing (e.g. deep or Bayesian models) is the research/enterprise alternative.
• Audio: script generation by LLM plus TTS for podcast-style overviews.
• Human-in-the-loop: the learner reviews generated cards before they enter the deck; wrong cards found during review are flagged and fixed.

Data preparation

The data is the learner's own material: lecture slides, PDFs, scans of handwritten notes, textbook chapters, lecture recordings and web pages. It lives on laptops, phones, cloud drives and the learning platform. Typical gaps: scanned or photographed notes without text layer, recordings with poor audio, duplicate versions of the same slides, mixed languages, and copyrighted material that may not be uploaded to third-party tools. For institutions: lecture material scattered over platforms, with unclear licence and access rights.

Preparation steps: - Collect material in one folder per course with clear names, drop duplicates and outdated versions. - Run OCR on scans and photos (PREP-006, DOC-014) and transcribe lectures (SPEECH-001). - Check copyright and privacy before upload: no third-party personal data or exam materials in cloud tools, see PREP-024 for simple classification rules. - For institutions: curate an authoritative course corpus with metadata (PREP-010, PREP-013). - Spot-check the first generated cards against the source and delete bad ones.

Rough share of effort: low, about 5-15 percent for individuals; institutional offers need more for licensing, metadata and single sign-on.

Implementation paths

Which path for whom

• Private (students, self-learners): Path A, because mature study apps (Anki with add-ons, NotebookLM-type tools, Quizlet) cost little and need no setup; choose a local app such as LM Studio or Ollama for confidential material.
• Public (universities, schools offering tools to students): Path B on EU-region or sovereign cloud APIs, because an institution-wide offer needs login integration, data protection contracts and EU hosting.
• Developer: Path C, because FSRS, open embeddings and a small local model give a self-hosted, privacy-friendly app at HW2.

Path A — Buy (recommended)

• Source-grounded notebooks: NotebookLM (Google; summaries, audio overviews, flashcards, quizzes from uploaded sources).
• Flashcard platforms: Quizlet (Magic Notes, AI study modes), Anki (free, FSRS built in, desktop/Android/iOS; AnkiMobile paid), RemNote, Brainscape, Knowt, Mochi.
• Chat assistants with file upload and study modes: ChatGPT study mode, Claude, Gemini guided learning; Mistral Le Chat (EU provider).
• EU options: check vendor location individually; Mistral (FR), DeepL Write for text polish (DE), and open-source Anki (no account needed, local data) are the most EU-friendly. For institutional rollouts prefer tools with a DPA, EU processing and no training on user content.
• Choose A when: an individual or small group wants results within days.

Path B — Build on cloud APIs

• Models: frontier LLMs with structured outputs for card JSON (Claude, Gemini, GPT, Mistral Large via EU endpoints/regions); Whisper-class ASR APIs, Mistral OCR or Azure Document Intelligence for scans.
• Stack: embeddings + pgvector/Qdrant, per-learner namespaces, citation spans, AnkiConnect or .apkg export so learners keep ownership of decks.
• Scheduler: open-source FSRS implementations (py-fsrs, ts-fsrs, fsrs-rs) [3].
• Choose B when: a university, publisher or startup wants a branded tool integrated with an LMS (Moodle, Canvas).

Path C — Self-host (open source)

• Models: open-weight LLMs such as Qwen3, Llama, Gemma or Mistral via Ollama/vLLM; embedding models like bge-m3 or multilingual-e5; faster-whisper / Whisper large-v3 for ASR; Docling or Marker for parsing; Piper, Kokoro or XTTS for TTS.
• Apps: Anki + FSRS add-ons that call a local LLM; Open Notebook (open NotebookLM alternative), Open WebUI knowledge bases.
• Hardware: HW2 (one 24 GB GPU) suffices for a campus-scale service of a ~14-30B quantised model; a student laptop with a small model (HW1) handles single-user notes and cards.
• Choose C when: data must not leave the institution or the learner wants fully local use.

Path D — Custom training

Rarely needed. Fine-tuning pays off for (a) a card-quality or question-difficulty classifier trained on user edits and ratings, and (b) memory models fitted on large review logs (the FSRS-style parameters are already per-user). Needs review logs from many learners, which raises GDPR and consent questions.

State of the art (2026)

• FSRS replaced the older SM-2 default in practice for many Anki users and is benchmarked openly by its community [2][3]. Gains over SM-2 are reported as fewer reviews for the same retention, but results depend on the dataset; verify on your own data.
• Long-context and multimodal models read whole lecture decks and recordings; source-grounded citations are now standard in NotebookLM-type products.
• Audio overviews make multi-speaker spoken summaries from sources; useful for commuting but weaker for precise detail.
• Limits and failure modes: hallucinated or oversimplified cards; mis-OCR of formulas and handwriting; cards that test recognition instead of understanding; "illusion of knowing" when a learner passively reads AI summaries instead of retrieving; over-long decks that create review debt; cards inherit errors of the source slides. Research warns that offloading to chatbots can reduce learning if it replaces effort [6].

Shared resources & combinations

• Shared stack: RAG over private documents (DOC-014, KNOW-018), summarisation (LANG-008), transcription (SPEECH-001).
• Combinations: EDU-002 (question generation, instructor side), EDU-004 (dialogue tutor explains missed cards), EDU-005 (adaptive paths use the same recall estimates), EDU-011 (language flashcards), EDU-001 (instructor-provided source packs).
• Differences: KNOW-018 stores and retrieves knowledge; this use case optimises retention through scheduling.

Effort, cost & KPIs

• Path A: hours; free tiers exist, paid plans are typically a few euros per month (check current prices).
• Path B: 1-3 months for an LMS-integrated product; cost driven by tokens for long documents and ASR minutes.
• Path C: 1-2 months; one GPU server plus maintenance.
• KPIs: time from upload to first review session; card acceptance rate (cards kept vs deleted); factual error rate on a sampled audit; retention (share of reviews rated "good/easy"), review load per day, FSRS log-loss / calibration; exam outcomes (ideally compared to a control group).

People & change

For private use there is little process: the learner decides, reviews cards and owns the data. For institutions, involve management and the department lead (teaching and learning centre) to set a policy on permitted tools, IT for login, hosting and integration with the learning platform, the DPO for the data protection basis, and legal for copyright and contracts. Students and lecturers are the real stakeholders: ask lecturers whether their slides may be processed and involve student representatives early. Works council: no, in general, because students are not employees and the tool does not monitor staff. If lecturers or staff are required to use or monitor it, §87(1) no. 6 and §90(1) no. 3 BetrVG (or the Personalrat) may apply, so check. A DPIA is usually not needed for a plain study tool without profiling; ask the DPO, and conclude an AVV (Art. 28 GDPR) with the provider.

Support AI literacy (Art. 4 AI Act asks for support measures) with a short module on verifying generated cards, hallucinations and honest limits: generated content can be wrong and replaces neither reading nor practice. Pilot with a volunteer course, collect feedback on card errors, and keep learners in control with edit and delete of every card. Typical concerns are cheating and shallow learning; address through clear course rules.

See GUIDE-006, GUIDE-007 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: typically minimal risk when used by learners for self-study; AI literacy (Art. 4) applies to providers and institutions deploying it. Transparency (Art. 50) applies to synthetic audio. It becomes high-risk (Annex III, education) if outputs are used to evaluate learning outcomes, steer learning for admissions/assessment decisions, or monitor students at institution level (see EDU-007, EDU-005). Emotion recognition of students is prohibited (Art. 5(1)(f)); not a part of this use case and must not be added [7].
• GDPR: uploads often contain personal data (lecture recordings with other students' voices, notes with names); review logs are behavioural data. Minors: parental consent / age limits (Art. 8, national age 13-16); education platforms for children need a DPIA. Prefer EU processing, DPA, no training on content, deletion on request [8].
• Copyright: uploading textbooks or licensed lecture material to third-party services may breach licences; recordings of lectures may need lecturer consent; text and data mining exceptions (DSM Directive Art. 3/4) do not cover redistribution of generated decks.
• Academic integrity: check institutional rules on AI use; summaries are not a substitute for own work.
• Mitigations: show citations, let learners edit and report cards, offer local-only mode, time-limited retention, export in open formats (.apkg).

Getting started

1. Pick 1 course, upload its slides and lecture notes to NotebookLM or a comparable tool; generate 30-50 cards and delete weak ones.
2. Import them into Anki with FSRS enabled, set a target retention (e.g. 90%) and a daily review cap.
3. Audit 20 random cards against the source for errors; note the failure types.
4. For institutions: run a DPIA, choose an EU-hosted or self-hosted path, define allowed sources (licence check).
5. Track card acceptance, daily load and exam results for one term before scaling.

Sources

1. Dunlosky et al. — Improving Students' Learning With Effective Learning Techniques (Psychological Science in the Public Interest, 2013) — https://journals.sagepub.com/doi/10.1177/1529100612453266
2. Anki manual — FSRS — https://docs.ankiweb.net/deck-options.html#fsrs
3. Open Spaced Repetition — FSRS algorithm and implementations — https://github.com/open-spaced-repetition/fsrs4anki
4. Google — NotebookLM Help — https://notebooklm.google/
5. Quizlet — AI study tools (Magic Notes) — https://quizlet.com/features/magic-notes
6. OECD — Digital Education Outlook 2026 / Generative AI in education — https://www.oecd.org/en/topics/digital-education.html
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
9. Open Notebook (open-source NotebookLM alternative) — https://github.com/lfnovo/open-notebook