AI Deepfakes Impersonated Two Hong Kong Celebrities to Sell Stock Tips — Build a Public Evidence Timeline Before You Report It
In one line: AI deepfakes cloned the face and voice of Hong Kong feng shui master Su Man Fung and YouTuber Ming Zai to push stock tips — and their platform complaints got nothing but auto-replies. Before you file a formal report, you can use LIHKG-scraper’s search mode to turn the public discussion already happening on Hong Kong’s largest forum into a dated, linkable evidence timeline.
Platform complaints go nowhere, because proving a deepfake is hard to begin with
Su Man Fung, a well-known Hong Kong feng shui master, found out two to three months ago — tipped off by his own supporters — that AI-generated videos impersonating him were spreading across at least eight different YouTube channels (weekendhk). Ming Zai, a YouTuber with over 900,000 subscribers, separately discovered that his voice had been AI-cloned to narrate a series of history and current-affairs commentary videos (i-CABLE). Both had already tried reporting individual videos to the platforms directly, but Su described the process as “extremely tedious” — every report drew only an automated system reply with no human follow-up, while the offending channels kept dodging detection by renaming themselves (weekendhk).
It gets worse: these deepfakes weren’t just parody. Scammers hijacked Su Man Fung’s actual YouTube channel name and used his real photos to produce highly convincing videos impersonating him to sell stock-investment tips, then paid to promote them via social media ads that funneled victims into WhatsApp groups for further exploitation. Reporting also identified several other Hong Kong celebrities, financial commentators, and stock pundits hit by the same tactic — this isn’t an isolated incident (gotrip).
Lawyer Chan Pak Ho points out that voice recognition is a harder evidentiary bar than facial recognition, because naturally similar-sounding voices already exist in real life, so proving AI synthesis “faces an objective difficulty” (i-CABLE). He also notes AI-impersonation claims have no direct Hong Kong court precedent yet, but if the deepfake’s purpose is fraud — for example, impersonating a stock analyst to push investment tips — it may still trigger fraud-related criminal charges (weekendhk). In other words, before formal evidence-gathering (a police report, or a complaint to the platform or a consumer body), what a target can already start doing is documenting how long and how widely the scam has been discussed in public — and a large public forum like LIHKG is exactly where that trail tends to survive.
Scrolling LIHKG by hand vs. letting LIHKG-scraper build the timeline
| Scrolling LIHKG manually, thread by thread | LIHKG-scraper’s search mode, scheduled | |
|---|---|---|
| Find the relevant discussion threads | ✅ (but you re-search each keyword by hand) | ✅ automated |
| Each thread’s post time and last-reply time | Have to open every thread and copy it yourself | ✅ returned directly as created_at / last_reply_at |
| Build a running, complete log over time | Easy to miss threads or forget what you’ve already checked | ✅ every run’s results can be logged straight into a spreadsheet |
| Labor required | Re-search and re-open threads every time | Runs on a schedule; a human only needs to review |
| Cost | Free but time-consuming | Billed per row actually returned (pay-per-event) |
How to build a public discussion timeline with LIHKG-scraper
Step 1 — Use search mode, with the target’s name plus deepfake-related terms as keywords
{
"mode": "search",
"keywords": ["Su Man Fung AI", "蘇民峰 深偽"],
"order": "desc_create_time",
"maxItems": 100
}keywords is sent directly to LIHKG’s site-wide search and is not limited to any one sub-forum. order: desc_create_time sorts by original post time, which is what you want for a timeline; if you’d rather see which thread got the most engagement, switch to desc_reply_time. To track only newly-appearing discussion, add dateFrom (e.g. 1w) to limit results to threads active within the past week — this also keeps re-runs cheap when you schedule it.
Step 2 — Log the key fields from every result into a table
Search mode returns a list of threads (ThreadRow), each with thread_id, title, reply_count, created_at, last_reply_at, thread_url, and scraped_at. Logging those fields gives you a running record of how many times the topic has come up on LIHKG, when it started, and when it last got a reply — with a link back to the original thread so anyone reviewing it can verify the content directly.
Step 3 — Re-run on a schedule to keep the timeline current
Use Apify’s Scheduler to re-run the same keyword set every few days; newly-appearing threads accumulate into your existing log. If you need the full content of a specific thread — say, someone posted a screenshot or link to the deepfake clip in a reply — switch to mode: "thread" with that thread’s thread_id to pull the original post and every reply.
What this method can’t do: search mode only returns a list of threads discussing the topic on LIHKG — not the deepfake video itself, and not legal advice. This timeline only helps demonstrate how long and how widely the scam has been circulating in public discussion; an actual police report or platform complaint still requires you to independently preserve the deepfake clips and the scam ad’s screenshots or links as primary evidence.
What it actually costs
LIHKG-scraper is billed pay-per-event, on the rows actually returned:
| Event | Price | Applies to |
|---|---|---|
search-listing (per thread-list row) | $0.002 | Search mode, sub-forum listing mode |
product-detail (per thread) | $0.008 | Full-post mode’s original post |
review-item (per reply) | $0.003 | Full-post mode’s replies |
A first pull of 100 rows in search mode to build the initial timeline costs roughly $0.2. After that, a weekly re-run scoped with dateFrom to only new threads typically returns a handful to a few dozen rows, so a month of monitoring is very likely under $1 — cheap insurance against having to reconstruct the timeline from scratch after a complaint has already gone nowhere.
Three ways to start
From least effort to most freedom
FAQ
How did Su Man Fung and Ming Zai find out they’d been deepfaked? Su Man Fung, a well-known Hong Kong feng shui master, learned from his own supporters two to three months earlier that AI-generated videos impersonating him were circulating across at least eight different YouTube channels. Ming Zai, a YouTuber with over 900,000 subscribers, found that his cloned voice had been used to narrate a series of history and current-affairs videos. Both said the fake video and audio were convincing enough to fool casual viewers.
What happened when they reported the fake videos to the platforms? Su Man Fung said the complaint process was extremely tedious — every report only triggered an automated system reply, with no human follow-up, while the offending channels kept dodging detection by renaming themselves. Scammers also hijacked his real YouTube channel name, ran paid social ads promoting the fake videos, and funneled victims into WhatsApp groups to pitch stock tips.
What do lawyers say about the legal difficulty of proving AI deepfakes? Lawyer Chan Pak Ho said voice deepfakes are harder to prove in court than face swaps, because naturally similar-sounding voices already exist, making it objectively difficult to prove AI synthesis — and Hong Kong has no direct case precedent yet. But if the deepfake’s purpose is fraud, such as impersonating a stock analyst to push tips, it could trigger fraud-related criminal charges.
How do I configure LIHKG-scraper’s search mode to build this kind of evidence timeline? Set mode to search, put the public figure’s or brand’s name plus AI/deepfake-related terms into keywords, and pick desc_create_time for order to sort by original post time. Each result returns thread_id, title, created_at, last_reply_at, reply_count, and thread_url, which you can log into a table to see how long the discussion has been running on LIHKG and how many separate threads are talking about it.
Can this replace filing a police report or the actual deepfake evidence? No. Search mode only returns a list of threads discussing the topic on LIHKG (titles, timestamps, reply counts) — it is a supporting public-discussion timeline, not the deepfake video itself and not legal advice. For an actual police report or platform complaint, you still need to keep the deepfake clips and ad screenshots or links as your primary evidence.
How is LIHKG-scraper priced? Search mode (including keyword search) bills under the search-listing event at $0.002 per thread-list row returned; thread mode for a full post bills under product-detail at $0.008 per thread; each reply pulled inside a thread bills under review-item at $0.003 per reply. All pay-per-event based on the actual rows scraped — no monthly fee.
Chad runs 40+ published Apify actors, including the LIHKG-scraper used here. Every date and detail above is checked directly against the news reporting cited; details that couldn’t be verified (such as either victim’s specific financial losses) are left out of this piece.