Flashbulb Flashbulb guide
Are AI-generated flashcards accurate, and how can I check them?
Mostly, but not always. In our blind tests on notes the AI card writer never saw while being tuned, about 1% of cards made from typed notes were wrong, about 3–5% from photographed notes, and 4–7% from photographed foreign-language vocabulary (October 23, 2026). Check every card against your notes before studying: is it in your notes, does it have one clear answer, and does the question give the answer away?
Last checked November 3, 2026 by Lyravant.
The mistakes to look for
- Not in your notes. The model adds a fact from its own memory. It may even be true, but it isn't what your course says.
- More than one right answer. "Name an example of a mammal" has thousands of answers; your exam wants the one in your notes.
- A giveaway. "Boyle's law was discovered by…" answers itself.
- Misread source. A photo misread (Δ as A, 3% as -3%) becomes a confidently wrong card.
How Flashbulb keeps cards right
- Only your notes. The writer is told to use only the text you give it, and every card keeps the line it came from ("From your notes").
- A second check. A second pass rejects cards with a named problem: not in the notes, more than one answer, a giveaway.
- Hard rules on top. A card with a number that isn't in your notes is dropped.
- Marks to check. On cards made from photos, the review screen highlights the words most likely misread: accents, stress marks, subscripts and Greek letters ("check these marks"). On a fresh test set it flagged 23 of 26 such mistakes.
- You decide. Every card waits on a review screen (keep, edit, delete) before it joins your deck.
How we measured it
Numbers checked October 23, 2026, for the version of the card writer in the app now; re-measured on new notes whenever the prompt or model changes.
| What you give it | Wrong cards | Fresh test sets |
|---|---|---|
| Typed or pasted notes | about 1% | 1.4% and 0.0% (0 of 265 cards) |
| Photographed notes (the photo goes with the text) | about 3–5% | notes pages 3.2% and 3.3% (4 of 121 cards); formula and worked-example sheets 2.7% (5 of 185); phone photos, handwriting, low light and diagrams |
| Photographed foreign-language vocabulary | 4–7% | 4.1% and 7.4% (20 of 272 cards); handwritten pages in Russian, Ukrainian, Greek, Arabic, Hindi, Chinese, Japanese, Korean, Vietnamese and Hebrew. Most misses are dropped stress marks and look-alike letters |
| Lecture notes from a recording (lines with a serious error) | about 3.6% | 48 of 1,338 lines (95% range 2.7–4.7%), four runs on 9 recorded MIT lecture clips transcribed on the iPhone; most misses are misheard names or terms |
| Pattern finder on the phone (offline fallback, typed and photographed) | 14.1% | 538 cards |
Notes: each test set was written after the card writer was tuned and run once. Reviewers graded every card blind (they didn't know which method made it) with a strict rubric: wrong means incorrect, not in the notes, no single clear answer, or the front gives the answer away. Photos are harder than typed notes: the text is read off the photo first, and a misread letter, accent or symbol can become a wrong card. That's why the photo itself goes to the card writer with the text: on an earlier set of fresh photos, sending only the text read from them gave 12.4% wrong cards. Mistakes cluster by page: one misread symbol can spoil every card from that page, which is why the app highlights marks to check. The numbers vary from set to set, so we show each one rather than a single best figure.
That's why the offline pattern finder is a fallback: it's free and private, but you should check its cards more carefully.
Questions
Do AI flashcards make things up?
They can. Use an app that builds cards only from your notes, shows the source line, and lets you review each card before studying.
Is the AI trained on my notes?
No. Your notes are sent to Anthropic's Claude model to write the cards; Lyravant doesn't store them and Anthropic doesn't train on them.
How accurate is Flashbulb?
In blind tests on notes it hadn't been tuned on, about 1% of cards from typed notes were wrong, about 3–5% from photographed notes and 4–7% from photographed foreign-language vocabulary; about 3.6% of lines in lecture notes had a serious error (checked October 23, 2026). On photo cards the app highlights marks to check. Check each card before you study it.
Flashbulb for iPhone and iPad
Snap your notes. Study the cards.
Photos, PDFs, slides or pasted text become flashcards from your own words. Then a few minutes a day, paced to your exam date.
Coming soon to the App Store
Free. No account. No ads. Needs iOS 26. Privacy policy