How AI grades answers in your own words

Why free-form answers matter

Flashcards that only accept an exact string teach recognition, not recall. HippoCampus asks you to explain the answer — typed or spoken — so you practice the same skill you need on an exam or in conversation.

Grading free-form text is harder than matching a single phrase. The app sends your answer to an AI grader along with the question and the reference answer (and any rubric the course author supplied). The model returns a structured result: whether you were correct, partially correct, or wrong; what you got right; what was missing; and a recommended rating for spaced repetition.

The path from your voice to a grade

When you answer out loud:

1. Audio is transcribed (speech-to-text) so the grader receives text.
2. The grader compares your answer to the expected content, not necessarily exact wording.
3. The UI shows the label (for example “Mostly correct”), short feedback, and the suggested Again / Hard / Good / Easy interval.
4. You can accept the suggestion or override the rating if you know you guessed or the grader misread you.

On the web and mobile apps, transcription and grading run through the same backend pipeline used for course generation. Content needed for grading is sent to OpenAI under their API terms, as described in the privacy policy.

What the grader is good at

The grader handles paraphrases well: different word order, synonyms, and shorter summaries often count as correct if the key ideas are present. It can flag missing steps in a multi-part answer and summarize the gap in plain language.

For factual questions with a clear reference answer, it is usually reliable enough to drive daily study — especially when you read the feedback and override obvious mistakes.

Where it can be wrong

No model is perfect. Watch for these cases:

• Homophones and transcription errors — speech recognition may mis-hear a term; the grader then grades the wrong text.
• Over-lenient partial credit — you may get “mostly correct” when an expert would want more precision.
• Over-strict wording — rare terms or acceptable alternate names may be marked wrong until you add them to the card’s acceptable answers (when the course allows editing).
• Subjective or open-ended prompts — grading is only as stable as the reference answer and rubric.

Treat the recommended rating as a default, not a verdict. You know whether you recalled the fact or bluffed.

How grading connects to scheduling

The recommended rating feeds the same scheduler as a manual tap on Again, Hard, Good, or Easy. A strong answer pushes the next review further out; a weak one brings the card back sooner. Grading quality affects when you see the card again, not whether the card exists.

Practical tips

• Speak clearly for voice answers; check the transcript if the grade looks off.
• Read the “what you missed” line — it is the fastest way to improve the next attempt.
• Override the rating when you used hints or luck; honest ratings keep intervals trustworthy.
• Edit generated questions if the reference answer is ambiguous; the grader can only enforce what the card says.