Hallucinations & safety

How Spark keeps the AI on a short leash.

If students are going to learn from AI‑generated content, "close enough" is not good enough. Spark stays grounded in your actual lectures and raises its hand when it's unsure instead of bluffing.

note_alt Study notes that are double‑checked, not free‑styled

For each chapter, Spark goes through a two‑step process: first, draft notes based on your lecture; then, a separate check asks "Is this scientifically correct?" and "Does this really reflect what the teacher said?"

If the second pass isn't confident, Spark tries small corrections a few times. If it still can't resolve the issue, the note is clearly marked for review instead of being quietly shipped as perfect.

  • check_circle

    Two-pass system: generation first, then verification for accuracy and alignment.

  • check_circle

    Automatic correction attempts before flagging - fixes what it can, surfaces what it can't.

  • check_circle

    Broken formulas, strange characters, and unclear content get flagged for your review.

quiz Quizzes anchored in real chapters

Spark's quizzes are designed to be answerable from the material your students have actually seen. Each question is tied to one or more specific chapters, not to vague "topics in the air."

There are checks to make sure the correct answer matches the explanation and is actually supported by the relevant chapter. Questions that don't clearly belong anywhere, or that can't be fixed cleanly, are dropped rather than forced in.

  • check_circle

    Every question linked to specific chapters - students know what they're being tested on.

  • check_circle

    Answer and explanation must match and be supported by the chapter content.

  • check_circle

    Questions that can't be fixed are dropped - no forcing content that doesn't fit.

account_tree Chapters and slides that match what's actually on screen

Spark doesn't just trust the first guess about where chapters begin and end. It cross‑checks suggested boundaries against the real transcript, and if something looks out of order, it runs a dedicated fix‑up step instead of silently slicing the video in the wrong place.

The same spirit applies to slides and key frames: visual duplicates are detected and removed, and if Spark can't find any usable slides for a chapter, it surfaces that as a visible issue rather than pretending everything is fine.

  • check_circle

    Chapter boundaries cross-checked against transcripts - fixes applied automatically when needed.

  • check_circle

    Visual duplicates detected and removed - no redundant slides cluttering the course.

  • check_circle

    Missing slides flagged as issues - you see problems before students encounter them.

forum A tutor that knows its place

The course tutor is intentionally narrow‑minded: it answers from your course, not from the open internet. Every reply is based on notes and chapters that were indexed beforehand, and those same snippets are surfaced back to the student as gentle references.

If a question falls completely outside the course, the tutor is encouraged to say so and redirect the student, instead of inventing an answer just to be helpful.

  • verified

    Answers constrained to your indexed course content - no open web hallucination.

  • link

    Source chapters and notes surfaced with every answer - full transparency for students.

  • block

    Out-of-scope questions get honest redirects - no inventing answers to be helpful.