Every round of Fill in the Blank shows you a real sentence with one word missing, and you have to supply it. The sentence itself has always mattered for how well the round teaches you anything — a garbled or oddly translated example is a bad round no matter how correct your answer is. What's new is which sentence gets picked for a given word, out of however many candidates exist for it.
The idea behind i+1#
Stephen Krashen's comprehensible input hypothesis is one of the most cited ideas in second-language acquisition: people acquire language from input that is almost entirely understandable, with a small stretch — one notch beyond current ability — built in. He called that stretch i+1: your current level, i, plus a single increment. Too far below your level and there's nothing left to learn; too far above and there's nothing to anchor the new piece to. The zone that actually works is narrow, and it's defined by how much of the surrounding material you already understand, not by how hard the new item is in isolation.
Applied to a single sentence, i+1 has a very literal shape: every word is known except one, and that one is exactly what you're being asked to produce. Fill in the Blank already had the "one word missing" part built in by design — the blank is always the tested word. What it didn't have, until now, was any sense of whether the rest of the sentence was actually i+1 for the person reading it, or whether it happened to be stuffed with three other words they'd never seen.
What actually changed#
Fill in the Blank has always chosen its example sentences for quality — natural length, a usable translation, the target word appearing cleanly. That quality bar hasn't moved. What's new is that, among sentences that clear it, the app now also considers how much of the rest of the sentence you already know — using the same mastery and review data behind mastery-aware blending and the vocabulary graph — and favors the one that leaves the blank as your only gap.
The result is that two learners studying the exact same word can be shown two different sentences, each i+1 for them specifically:
| Candidate sentence for "aeropuerto" | Who it fits | Why |
|---|---|---|
| El ___ está muy lejos del centro. | A beginner who knows estar, muy, lejos, centro | Every surrounding word is already familiar — only the blank is new |
| Perdimos el vuelo porque el tráfico cerca del ___ era terrible. | An advanced reader comfortable with perdimos, vuelo, tráfico, cerca | Richer, more natural sentence — still fully comprehensible for this reader |
Both sentences are real, human-written examples that existed in the pool already. Nothing was generated for either learner — the app simply picked the one that leaves them with a single, isolated gap instead of a sentence that's difficult in three places at once.
The preference is bounded, on purpose#
i+1 fit is a preference on top of sentence quality, not a replacement for it. A well-formed, well-translated sentence where you know all but one word still loses to nothing — but a sentence that's otherwise weak (odd phrasing, no translation, an unnatural length) doesn't win just because it happens to be i+1 for you. If the best i+1 match for a word is a poor-quality sentence and a good-quality sentence is available with one extra unfamiliar word in it, the app can still choose the good-quality one. The point is to steer selection among already-decent candidates toward the one that puts you in the comprehensible-input sweet spot — not to chase a perfect i+1 fit at the expense of a sentence that actually reads well.
It also isn't static. Your vocabulary changes every week — words graduate from learning to mastered, new words get saved — so the same word pool for "aeropuerto" can rank differently next month than it does today, even though the sentences themselves haven't changed. There's nothing to configure: the ranking simply reflects your review history as it stands each time a session loads.
Known-ness also follows you through inflection, not just exact word matches. If your review history shows you know comer, a candidate sentence containing comí or comiendo is treated as containing a known word around the blank, not an unfamiliar one — the same inflection resolution that lets mastering despertar also credit despierta elsewhere in the app.
Where the sentences actually come from#
It's worth being precise here, because it would be easy to assume this means AI-generated sentences tailored to you — it doesn't. The sentence pool comes from Tatoeba, an open, community-built corpus of real, human-written sentences with translations, the same source Fill in the Blank has always used. What changed is entirely on the selection side: which existing, real sentence gets shown to which learner for a given word. No sentence is written or rewritten per user.
There's one narrow exception, and it predates this feature entirely: on words where Tatoeba genuinely has no usable sentence at all, Pro users fall back to a sentence generated by Gemini at an intermediate level. That fallback only fires when the corpus has nothing to offer — it isn't part of how i+1 selection works, and it hasn't changed.
What it costs#
Nothing. Your known-word set is assembled from data the app already has — the same spaced-repetition history that drives your review schedule — and the ranking happens on-device. No word you've saved or reviewed is sent anywhere for this feature, no extra network calls are made, and it's available to every learner regardless of subscription. This is one of the places where personalization is genuinely free: the intelligence is in how existing data gets used, not in any new AI call.