CAPIT SCAM
Spot fake listings and scams before your visitors do
SCAM reads the profiles, listings, texts and photos posted on your site and tells your moderator which ones to review first, with the reasons. It learned from the moderation decisions of the dating site Affection.org since 2017.
Private launch: access is granted on request, site by site, after a discussion with us.
The first 50 analyses each month are free. Beyond that, prices are communicated to you before any billing.
- An opinion, never a penalty: SCAM ranks your moderation queue and explains; your moderator decides.
- Readable reasons: a photo already seen in a fraud, an email or network already sanctioned on your site, a request for money in the text, network signals that you send.
- Trained on real decisions: 16,321 frauds and 11,730 legitimate listings confirmed by moderators.
- One site, one key, one queue: your data and your feedback stay separate from those of other customers.
- No appearance criteria: sex, age or physical looks are never a reason.
- A simple HTTPS API, with asynchronous analyses and tracking of your usage.
How it works
- Your site sends SCAM the text, photo or profile to examine, with your site’s key.
- SCAM acknowledges receipt, then analyzes in the background. You fetch the result with the identifier you received: your page does not wait.
- The result gives an estimated risk, the reasons and a recommendation: usual review, examine carefully or examine as a priority.
- Your moderator decides. Their decision can be sent back to SCAM, which stores it separately for your site.
How SCAM learned to recognize fraud
A model learns from examples whose answer is known. SCAM used those that already existed: the moderation decisions of Affection.org, since 2017.
- 160,615 listings reviewed, including 16,321 frauds confirmed by a moderator and 11,730 legitimate listings confirmed.
- The 48,335 automatic deletions of the old system serve as memory (“this email has already been sanctioned”) but not as examples, because no human judged them: SCAM does not imitate a robot’s mistakes.
- For each listing, SCAM sees only what was known before it was created: same email, same IP address, same network, same username or same photo as a listing already sanctioned, IP address reputation, IP country different from the declared country, clues in the text.
- Photos are compared by fingerprint, which recognizes the same image even if recompressed or flipped like in a mirror. The image itself is not kept in the training set.
- The model is judged only on a period it has never seen (March to November 2026): it is never evaluated on what it has already learned.
What this gives
Test on 690 Affection.org listings from March to November 2026, including 39 frauds, never seen during training. Question: by reading only the top 10% of listings in the ranking (69 listings), how many frauds do you find?
| How the 69 listings to read are chosen | Frauds found out of 39 |
|---|---|
| At random | about 4 |
| Ranking by a simple model | 8 |
| SCAM ranking | 25 |
Retrospective evaluation on a single site and on few frauds: the figures need to be confirmed on your own site. Measured on November 18, 2026.
What SCAM cross-checks
- Your site’s history: listings already sanctioned, same email, same network, same photo.
- The network signals your site sends: VPN, proxy, hosting provider, mismatch between the declared country and the country of the connection.
- The text: request for money, bank details, invitation to move to an external messaging service, insult or targeted threat.
- Other reputation databases and reports of known scammers may be added, according to their terms of use.
What SCAM does not do
- It deletes nothing and approves nothing: it gives an opinion.
- At this stage it is used to rank the moderation queue, not to delete automatically: its scores are not yet reliable enough to decide on their own.
- The results above come from a single site and a test of 39 frauds. Your site will have its own measurement.
An API for your developer
You send, SCAM replies with an identifier, you fetch the result. OpenAPI documentation and JavaScript, Python and PHP examples are provided with your access, along with the API base address.
POST /v1/check/text | Analyze a text or a listing description. |
|---|---|
POST /v1/check/image | Analyze a photo (fingerprint, reuse, content). |
POST /v1/check/profile | Analyze a profile: text, network signals and your site’s measurements. |
GET /v1/analysis/{id} | Fetch the result of an analysis. |
POST /v1/feedback | Send back your moderator’s decision. |
GET /v1/usage | Track your usage for the month. |
Your data
Each site has its own key and separate data. Emails, IP addresses and usernames are turned into fingerprints that make it possible to recognize “the same one” without displaying it. Your moderator’s feedback is used to train a model only if you allow it. The precise terms of processing are agreed when your access is opened.
Take part in the launch
SCAM is not yet open to the public. Describe your site and the type of listings or profiles to protect: we will reply with trial access.
