Why Hebrew copy written by AI usually sounds off
Modern LLMs are excellent at English and merely competent at Hebrew. The difference shows up in three places: overly formal register where conversational would be natural, English sentence structure translated word-for-word (which produces grammatically valid but rhythmically wrong Hebrew), and a small bag of overused phrases — "במסגרת זו", "ברצוננו לציין", "חשוב לציין כי" — that native Hebrew speakers immediately tag as machine-generated.
For Israeli SMBs sending Hebrew marketing emails, WhatsApp broadcasts, or website copy, the cost is reply rate. Hebrew readers are unusually sensitive to register mismatches; copy that sounds robotic gets ignored even when the offer is relevant.
What this analyzer will measure
- Formality register on a 1–5 scale (אנא vs. בבקשה, etc.). Flags mismatch with the apparent context.
- Sentence-length variance. AI tends to produce uniformly medium-length sentences. Native writing has bursts.
- Overused phrase detection. The 30 most common AI-tells in Hebrew business writing.
- Niqqud handling. Detects whether vowel pointing was added programmatically (and is therefore probably wrong).
- Code-switch quality. When Hebrew copy mixes English brand names or technical terms, does it use the transliteration a native speaker would?
- Readability grade. An adapted Flesch-Kincaid equivalent calibrated for Modern Hebrew.
- Direction issues. Mixed LTR/RTL artefacts that survive copy-paste from a web editor.
- Suggested rewrites. One-sentence-at-a-time, showing the "before" and a more natural alternative.
Why we're building it
We needed this internally. Our voice agents (Link Voice) speak Hebrew to thousands of callers a week, and the same quality bar applies to outbound copy from Mailer. The analyzer started as an internal QA tool — we're shipping it publicly because the problem isn't unique to us, and the alternatives (manually copying into GPT and asking "is this Hebrew natural?") are slower and less reliable.
Privacy
The analyzer will run client-side wherever possible. Server-side scoring uses ephemeral inference — your text is not stored, not used for training, not logged beyond standard rate-limit counters. That's the same standard we apply to call transcripts in Link Voice.