A warning page before a block
Taiwan's Ministry of Digital Affairs (MODA) announced a new online anti-fraud alert mechanism on August 4, 2026, built with Chunghwa Telecom, Taiwan Mobile, Far EasTone, Trend Micro and Gogolook. When a reported site is confirmed as fraudulent, participating providers show users a warning page. That happens before the domain is blocked through formal administrative procedures, according to MODA's announcement. MODA describes the design as closing the gap between discovery and takedown. The ministry's pipeline used to run "discovery, notification confirmation, and take down". It now adds simultaneous alerts.
The scale explains the urgency. MODA's Online Fraud Reporting and Inquiry Portal, launched September 30, 2024, has received more than one million reports in two years, and ISPs have taken down more than 500,000 fraudulent sites, Taipei Times reported. MODA says May 2026 alone brought 33,201 reports, more than ten times the initial monthly volume.
The strongest case for the mechanism
The case for acting at network level is real. Scam sites live for hours or days, and takedown workflows take longer than that. Every hour between a confirmed report and a block is an hour in which a victim can be defrauded. Minister Lin Yi-ching put the stakes in arithmetic: if an average scam costs NT$10 million, 10,000 reports would represent NT$100 billion in prevented losses, per the Taipei Times. Individual users cannot verify a fake delivery-service or payment page. The ISPs and security firms that see traffic at scale can.
A warning page is also a lighter instrument than the alternative. It does not delete anything, does not stop a user from proceeding, and gives the person the information to decide. That fits a pro-speech, pro-innovation view better than the default tools available to the state, which are silent DNS blocking and platform takedown orders.
Why the warning-first design is the better trade-off
Proportionality asks whether the state chose the least restrictive effective means. On that test, a visible warning beats an invisible block for three reasons.
- Transparency to the user. A warning tells the user that intervention occurred and why. Silent blocking looks like a broken internet.
- Reversibility. A wrongly flagged site can be cleared and the warning lifted without the damage of a completed takedown.
- Graduated escalation. The warning bridges the window before a formal administrative block, so the harsher step still requires the formal process.
Taiwan has also already legislated the harder tools. The Fraud Crime Hazard Prevention Act requires online advertising platforms to restrict access to or remove content when notified of suspected fraud (Article 33). Article 42 lets authorities order ISPs to suspend analysis service or restrict access to fraudulent websites, as the statute's English text reads. Advertising platforms face a 24-hour removal duty with fines from roughly USD 16,000 to USD 3.2 million. Against that backdrop, a warning is the mild end of the toolkit.
Where the proportionality risk sits
The risk is not the warning page itself. It is the machinery that decides what gets a warning. Three gaps deserve attention.
First, the legal basis is not stated. MODA's announcement refers to blocking through "formal administrative procedures" but does not name the provision that authorises the pre-block warning. The statute has clear appeal language for telecom penalties (Article 26 sends disputes directly to administrative litigation). Nothing in the material I reviewed says how the operator of a wrongly flagged site challenges a warning. A warning that appears before any formal decision effectively sanctions a site with no decision to appeal.
Second, the flagging pipeline runs on volume. A portal fed by more than 33,000 reports in a month, with automated case assignment and verification by public and private partners, will make errors. Impersonation cases are checked with the impersonated party, which is sensible. But fraud detection also involves commercial security vendors. A warning from a vendor-informed blocklist is a private judgment given public force. Reputational harm to a small merchant or a legitimate but unusual site can arrive within minutes and last much longer.
Third, over-blocking accumulates. A takedown count of more than 500,000 sites is a measure of activity, not accuracy. MODA has not published, in the sources I could verify, a false-positive rate, a count of reversed flags or a median time to clear a wrongly flagged site. Without those, no outsider can tell whether the mechanism is precise or merely busy.
What proportionate implementation looks like
None of this requires abandoning the mechanism. It requires publishing the rules under which it operates:
- Public criteria. State what evidence makes a site "confirmed" fraudulent, and which body signs off, rather than which vendor flagged it.
- A fast counter-notice route. Give site owners a named contact and a stated clearing time, with the warning lifted while review is pending for low-risk categories.
- Transparency reporting. Publish flags issued, flags reversed and time-to-clear at least quarterly. The platforms regulated under the Act already owe annual transparency reports, so ISPs and MODA can meet the same standard.
- Scope discipline. Keep the warning to sites confirmed as fraud. Do not stretch it to content that is merely misleading or disliked.
MODA's record shows it can iterate. Its October 2025 release on digitising takedown workflows described replacing paper notices with an electronic reporting platform across local governments, after Taipei's city government found Meta had not removed fraudulent ads within 24 hours on two occasions. The same discipline should now apply to error correction.
The bottom line
A warning page is a defensible answer to a fast-moving harm, and it is better than the silent block it precedes. Its legitimacy will rest on whether Taiwan lets the public see how often the system gets it wrong and how quickly it fixes those errors. Without that, a proportionate design becomes an unaccountable one.