AI moderation
Every listing on bănuț.md passes through an automatic AI filter — before it reaches the site. The goal isn't censorship, it's hygiene: to prevent scams, spam and fake listings while preserving the speed at which a real person can publish.
What the AI checks
Fraud and scams
- Requests for advance payment to foreign accounts, Bitcoin, Western Union.
- "I lost my card, send me the money" schemes.
- Phones cloned from old ads, numbers known as fraudulent.
- Impossible promises ("100% guaranteed profit", "earn without working").
Spam and irrelevant content
- Duplicate listings (same text posted 5+ times).
- Text unrelated to the category (third-party site promotion).
- Referrer links, external promo codes, pyramid schemes.
Forbidden content
- Weapons, illegal substances, controlled medication.
- Legally protected animals, trophies, heritage objects.
- Sexual content or content exploiting vulnerable people.
- Counterfeit products presented as original (iPhone "replica", knockoff perfume, etc.).
The 3 possible decisions
Allow — published directly
If the AI has high confidence (≥90%) that the listing is legitimate and the user has a verified phone, the listing becomes active immediately. ~85% of listings go this route.
Review — sent to manual verification
If the AI isn't sure (ambiguous category, suspicious price, unusual language), the listing lands in our moderator queue. A human reviews within hours and approves/rejects. While waiting, your listing is hidden but stays in your account.
Block — automatic rejection
Only if the AI is very sure the listing breaks the rules (obvious fraud, forbidden content), it's rejected directly. You get a message with the reason. You can edit and resubmit, or appeal at ajutor@banut.md.
What the AI does NOT do
- Doesn't read your messages. The filter only runs on public listing titles and descriptions — not on private chats.
- Doesn't collect data about you. Only the listing text reaches the model. No IP, no history, no identification.
- Doesn't discriminate. Listings are judged on content, not user.
- Isn't infallible. If it rejects a listing unfairly, appeal at ajutor@banut.md — a human reviews the decision.
How it improves
Every decision (allow, review, block) is logged in an internal audit. Our moderators review edge cases weekly — if an error pattern repeats, we tune the prompt. We use Claude Haiku and Sonnet from Anthropic, chosen for the quality/cost ratio and because they don't train on your data.
Other AI filters on the site
- AI fair price — on listings, we show whether the price is below/at/above the category median, as a signal for the buyer.
- Category detection — when you publish, the AI automatically suggests the category based on the title.
- Specific field generation — for "Mercedes W211" the AI proposes the right fields (year, mileage, fuel, transmission), no manual setup.
- Search with AI ✨ — you can type "BMW under 200k lei Chișinău" and the AI applies the filters automatically.
Want more technical detail? We have a blog where we write about behind-the-scenes choices.