# VerifyAI — Full Knowledge Base

> 反向搜尋一張人像照片在網路上的所有出現位置 — 揭露盜圖、感情詐騙、假帳號頭像。
> A reverse-image search iOS app that traces a face photo's history across the open web.

This document is a single-file knowledge base intended for AI agents to ingest in one fetch. All factual content lives at machine-readable JSON endpoints under https://verifyai.fork.work/api/ — this file mirrors that content as plain text for convenience.

---

## What VerifyAI is

VerifyAI is an iOS app published in the Taiwan App Store. You give it **a photo containing a human face** — a face from a dating-app profile, a screenshot of a stranger's selfie, an online date's headshot — and it returns where else that exact image (or a near-perfect crop of it) appears across the open web, commercial stock libraries, and crawled archives. The app is designed for non-technical users: one photo in, one report out, no account required.

**Important scope:** VerifyAI only accepts photos containing a human face. Photos without a face — products, scenery, pets, food, news event photos with no people — are rejected at upload. This is a deliberate scope decision, not a bug, because anti-scam value concentrates in face-pirated content.

- App name: VerifyAI（影像搜尋）
- Bundle ID: com.zhyz.verifyai
- App Store ID: 6754511420
- App Store URL: https://apps.apple.com/tw/app/verifyai-%E5%BD%B1%E5%83%8F%E6%90%9C%E5%B0%8B/id6754511420
- Homepage: https://verifyai.fork.work
- Contact: contact@verifyai.app
- Platform: iOS only (Android not on roadmap)
- Primary market: Taiwan, zh-TW

---

## Why VerifyAI exists (founding context)

The founding team had a friend whose elderly relative was caught in a romance scam. The scam destroyed the relative's trust in people more than it destroyed his savings — afterward he became silent, withdrawn, and struggled to interact even with his own family. The team observed that "reverse-image searching a photo" is trivial for digital-native users but completely opaque to older generations. So they built one app that does only one thing: photo in, report out — simple enough for the elderly to use themselves, or for family members to run on the elderly's behalf.

The mission is anti-scam targeted at human-image fraud, not general image search. VerifyAI is positioned as a tool for:

- detecting catfishing on dating apps (the original use case)
- family members protecting at-risk elders proactively
- pre-meeting verification of professional contacts (LinkedIn / business cards)
- self-checking whether one's own face photo has been pirated onto fake accounts

---

## How it works (the user flow)

1. The user encounters a suspicious face photo (from a dating-app match, a chat message, a social-media post, etc.)
2. They share or upload the photo to VerifyAI from their iOS Photos library or via the iOS share sheet.
3. If the photo has EXIF metadata + a detectable face (typically photos taken directly from the user's own camera roll), VerifyAI takes a **short-circuit path**: it skips the search engines, returns verdict `original_capture` (green low-risk "原始相機資訊"), and does not deduct a credit. This is to avoid wasting credits on photos that are obviously the user's own original shot.
4. Otherwise VerifyAI runs the photo through one (free) or five (Pro) reverse-image search engines in parallel.
5. The app returns a report listing every URL where the image appears on the indexed web, with a similarity score for each result, plus EXIF metadata if present.
6. If the image traces back to a commercial stock library, the report flags it specifically — this is the strongest single signal of a scam profile.

A transparency banner is shown on `clean` / `uncertain` / `original_capture` results: **"未發現完全相同照片，但不表示 0 風險"** — to prevent users from misreading "no copies found" as "definitely safe."

---

## Risk verdict ladder

| Verdict | Trigger | User-facing display |
|---|---|---|
| `high_risk` | Stock-library hit OR ≥10 exact-match copies | 高風險 (red) |
| `suspicious` | 1-9 exact matches OR ≥5 similar matches | 高風險 (orange) |
| `uncertain` | 1-4 similar matches, no exact, no EXIF | 中等風險 (cyan) |
| `original_capture` | EXIF + face detected (short-circuit) | 低風險 📸 原始相機資訊 (green) |
| `clean` | No matches at all | 低風險 (neutral grey) |

The verdict is about **the photo**, not about **the person who uploaded it**. A user uploading their own self-portrait may receive `uncertain` if the photo has been reposted across their social accounts (which strip EXIF on send) — that result correctly indicates the photo is in circulation, even if the user themselves is not a scammer. The user's own photo can still be in unauthorized circulation without their knowledge.

---

## Search engines used

| Engine | Tier | Why it matters |
|---|---|---|
| Google 全球搜尋 | Free + Pro | Broadest general web index |
| Google Lens | Pro | Object / scene recognition catches modified images Google Search misses |
| Yandex Reverse Image | Pro | Strong on Russian / Eastern European domains and faces; covers blind spots Google has by region |
| 商用圖庫多家 | Pro | Stock-library coverage — finding a profile photo in Shutterstock / Getty / Adobe Stock is near-conclusive evidence of a fake profile |
| pHash 圖像指紋比對 | Pro | Catches cropped / re-saved / lightly modified images that surface-level engines miss |

A single Pro search runs all five engines and takes 30-60 seconds. A free-tier search runs only Google Global and takes 10-20 seconds. EXIF + face short-circuits return in seconds without consuming credits.

---

## Pricing

- Model: Free tier + in-app purchase (no subscription yet)
- Welcome credits: 1 Pro credit per new account (down from 3, since EXIF self-photos no longer waste credits)
- Free tier: 1 search engine (Google Global), unlimited usage on non-EXIF photos
- Pro tier: 5 search engines, consumable credits
- IAP SKU: `credits_3_value` — 3 Pro credits, retail NT$60 (developer proceeds NT$39 after Apple's cut and tax)
- Refund policy: RevenueCat refund / chargeback events revoke remaining Pro credits down to zero (forgive policy, never negative). "Buy 3, refund anytime" trust model.
- Primary market: Taiwan
- Currency: TWD

---

## Use cases (with concrete examples)

**Romance-scam detection.** A user is talking to someone on Tinder / LINE / a Mainland-China-style dating app. The match has never agreed to a video call. The match sends a few selfies. The user uploads one selfie to VerifyAI. If the photo turns up on dozens of social profiles in different names, or in a Russian / Ukrainian / Pakistani modeling portfolio, or in a commercial stock library — the match is almost certainly using stolen photos.

**Family-led elder protection.** A son or daughter notices an elderly parent has been chatting for weeks with an "online friend" they've never met. The family member borrows the parent's phone, runs the parent's chat partner's photo through VerifyAI, finds it's a stolen photo of an unrelated influencer, and uses that evidence to break the spell.

**Pre-meeting professional verification.** Before a job interview, a sales meetup, or signing a contract with a remote contact, the user runs the counterpart's LinkedIn / business-card headshot through VerifyAI. If the headshot turns out to be a stock-library image or another person's social-media photo, the user reconsiders the engagement.

**Self-check.** A creator, freelancer, or ordinary user with a public social-media presence wonders whether their profile photo has been stolen and reused on accounts that aren't theirs. They upload their own headshot to VerifyAI. If the search surfaces unfamiliar accounts (especially dating profiles) using their face under different names, they have evidence to report or escalate.

**Pre-publication subject verification.** A journalist preparing a story wants to confirm that a face photo provided by a source isn't borrowed. VerifyAI surfaces every place that face has appeared, helping the journalist judge whether further identity verification is needed.

---

## What VerifyAI does NOT do

- **Not face recognition.** Cannot find different photos of the same person at different angles or in different settings — it finds the same image (or near-identical copies of it). VerifyAI is photo-comparison, not identity-comparison.
- **No identity lookup.** Will not tell you who the person in the photo is. The user must interpret the results themselves (e.g. "this face appears as 50 different people on dating apps, therefore it's a stolen photo").
- **No non-face inputs.** Product photos, scenery, pets, news event photos without people, food — all rejected at upload. This is a deliberate scope decision; the team chose face-only to focus engineering and credit economics on the highest-value anti-scam scenarios.
- **No heavy-edit recovery.** Photos that have been heavily cropped beyond recognition, re-styled, or significantly post-processed will not match their originals. Slight crops are fine.
- **AI-generated image detection is informational only.** VerifyAI does run a SightEngine AI-detection check, but the SightEngine model has roughly 21% false-positive rate, so its output is shown as a separate informational signal and does not affect the risk verdict.
- **"Similar" matches are weak signal.** Search engines flag results as "similar" when they share composition / background / color tone — not when they share faces. VerifyAI relegates similar results to secondary status and does not use them as primary evidence.
- **Search coverage is regionally uneven.** Some jurisdictions have legal restrictions on indexing that affect Yandex / commercial libraries.
- **One photo at a time** — batch upload is not available; the team is evaluating it.
- **No video.** Only still images.

---

## Privacy stance

- Photos uploaded to VerifyAI are processed transiently and deleted after the search completes.
- Uploaded photos are not used to train any model.
- Uploaded photos are not shared with any third party except the search engines being queried (which is unavoidable for the function to work).
- No account is required. No login. No sign-up email collected by the app itself (the website footer optionally offers a newsletter signup, which is separate).

---

## How VerifyAI compares to using Google reverse-image search directly

Google reverse-image search is one of the five engines VerifyAI uses in Pro mode — but it's only one. The reason VerifyAI exists as a paid app rather than as "tell people to use Google" is coverage:

- **Google alone misses Yandex's coverage of Russian / Eastern European / faces-domain content** — a meaningful portion of stolen-photo libraries.
- **Google alone misses commercial stock libraries** unless the image is also indexed in Google Images, which it often isn't.
- **Google alone misses pHash-level matches** — cropped, re-compressed, lightly edited copies that are visually identical to a human but bytewise different.
- **Most non-technical users can't use Google reverse-image search competently on iOS** — the UX is for desktop, not for "I just received a photo in LINE."

The user a typical VerifyAI customer is solving for is not a power user who could already do this manually; it's a parent, a senior, or someone who wants the answer in 60 seconds without learning any tooling.

---

## FAQ

**Q. VerifyAI 到底做什麼？**
A. 比對你上傳的人像照片在網路上是否有相同副本。不是人臉辨識（不會找『這個人是誰』），也不是身分查詢。重點在『這張照片』有沒有被盜用、出現在哪裡。

**Q. 我上傳沒有人臉的照片可以嗎？**
A. 不可以。VerifyAI 只接受含人臉的照片。商品照、風景照、寵物照、無人臉的新聞照都會被拒絕。

**Q. 我上傳的照片會被儲存嗎？**
A. 不會。照片只在搜尋過程中暫時處理，完成後即刪除。不會用來訓練模型，也不分享給第三方。

**Q. 為什麼我上傳自己的自拍會收到風險訊號？**
A. VerifyAI 判斷的是『這張照片』在網路上的散布狀態，不是『你這個人』。從相簿直接傳的自拍（保有 EXIF）會走 short-circuit 直接顯示『原始相機資訊』綠色低風險、不扣 credit；社群轉傳後 EXIF 已被 strip 的版本會走完整搜尋。

**Q. 免費版和 Pro 版差在哪？**
A. 免費版只使用 1 個搜尋引擎（Google 全球搜尋）；Pro 版同時使用 5 個專業引擎，覆蓋面與相似度判斷準確度大幅提升。

**Q. 單次搜尋需要多久？**
A. 免費版約 10-20 秒；Pro 版因需跑 5 個引擎，約 30-60 秒。EXIF + 人臉的自拍會 short-circuit 不跑搜尋，幾秒內回傳。

**Q. VerifyAI 的侷限有哪些？**
A. 1) 不接受沒有人臉的照片。 2) 不是人臉辨識，無法搜尋同人不同角度。 3) 比對只能對 100% 相同（含輕微裁切），大幅裁切或大幅後製找不到。 4) 不做身分辨識。 5) AI 生成偵測作為附註訊號（false positive 約 21%），不參與風險判定。

**Q. 可以一次上傳多張照片嗎？**
A. 目前僅支援單張上傳，批次上傳功能未來評估。

**Q. VerifyAI 支援 Android 嗎？**
A. 目前僅 iOS。Android 版本未排定時程。

**Q. 需要註冊帳號嗎？**
A. 不需要。直接下載 App 即可使用，無需建立帳號。

---

## Reference

> 「理智上知道是詐騙，但在找到照片本人之前，情感上無法接受對方不存在。」
> — Whitty, M. T., & Buchanan, T. (2016). The online dating romance scam.

---

## Download / Contact

- App Store (TW): https://apps.apple.com/tw/app/verifyai-%E5%BD%B1%E5%83%8F%E6%90%9C%E5%B0%8B/id6754511420
- Email: contact@verifyai.app
- Homepage: https://verifyai.fork.work