In a recent case of marriage and dating fraud revealed by the Chengdu High tech Police, a gang used AI to mass produce fake images of middle-aged and elderly men, packaging them as "privileged" blind date partners, and then attracting women with marriage and dating needs through live streaming platforms. In less than a year, more than 50 women were deceived, with amounts involved exceeding one million yuan.

AI makes the cost of fraud increasingly low. If the platform wants to block more, it needs to increase its governance efforts. The investment cost is borne by the platform, but the losses from being deceived mainly fall on the users. In this situation, it is difficult for the platform's anti fraud efforts to keep up with the changes in fraud.

AI enhances the replication ability of fraud

In the case disclosed in Chengdu, chatting with the victim is still manually handled, but the photos and videos used to package the "male guests" can already be generated in bulk. After a person's character becomes invalid, they can start over by changing their face, name, and creating a new set of experiences. AI reduces not only the cost of creating a fake photo, but also the cost of starting over after a failed false identity.

After this low-cost production capacity enters the content platform, the economies of scale will rapidly amplify. In April 2026, Tiktok disposed of more than 1300 accounts that misled the middle-aged and elderly by using the image of "AI Bazong", and more than 30000 related content were removed from the shelves. One governance involves over 30000 pieces of "AI hegemony" content, and AI false personas are no longer a sporadic phenomenon.

Criminals may not only use AI to falsify identities, but also use AI to impersonate real identities. In a case disclosed in March 2026, criminals synthesized illegally obtained ID photos into dynamic facial videos with blinking, shaking heads, and other movements for false real name authentication. The police have seized over 50000 such videos.

AI has also entered the interactive stage of fraud. A risk warning issued by Huatai Property and Casualty Insurance Liaoning Branch in 2025 mentioned that an elderly person in Chaoyang City was pulled into a WeChat group promising high rebates and invested a total of 3000 yuan under continuous inducement. The risk warning states that after the elderly reported the incident, the police verified that the WeChat group was operated by an AI account. The interaction between "active atmosphere" and "rebate commitment" in the group can both come from the program. Scammers no longer need to hire a large number of "daycare" to create an atmosphere, and a set of AI programs can achieve their goals.

From creating a face, maintaining a false persona, to continuously interacting with victims, more and more links in the fraud chain can be quickly replicated. Blocking only a single fraudulent account is difficult to curb the entire scam.

Why is the high interception rate not enough

According to the 2026 Anti fraud Governance Report released by Kwai, 98.1% of fraud behaviors were blocked by active interception in the past year, and 17600 fraud related videos and live broadcast rooms were blocked daily. With the help of AI intelligent analysis assistants, the cycle from discovery to online interception strategies for new fraud methods has been shortened by 60%. According to data disclosed by Weibo, by 2025, over 6.35 million fraudulent accounts will be closed, over 5.44 million high-risk conversations will be blocked, and over 68.36 million security alerts will be triggered.

These numbers reflect the risk situation of platform interception, but they are not enough. Looking at the overall interception rate alone, it is difficult to know where the problem lies with those scams that have not been stopped: when new scams first appear, the platform may temporarily not be able to recognize them; Some accounts have already experienced abnormalities and even received user complaints, but they have not been dealt with in a timely manner. These two situations reflect different governance issues. The data that needs to be paid attention to also includes how long it took for abnormal accounts to be discovered and verified, and whether the same type of fraud continues to occur after being dealt with.

The platform needs to identify and intervene earlier and more meticulously. However, for platforms, verifying high-risk accounts and increasing identification rules require continuous investment of more manpower, material resources, and financial resources, and even being too strict may inadvertently harm normal users. The platform can easily calculate how much money was spent; After intercepting a fraud, it is difficult to directly convert the amount of losses avoided and trust accumulated for users into quantifiable platform revenue. In this case, the platform's anti fraud drive may not be as strong.

Anti fraud measures can also affect platform revenue

Platforms are not always pure defenders outside the fraud chain. In some businesses, fraud can even bring revenue to the platform.

In February 2026, the Jiangsu procuratorial organs disclosed a fraud case of "live streaming ranking PK". Two anchors falsely claimed that the loser of the live PK would transfer 200000 yuan to the user who gave the most rewards in the winner's live room, inducing one user to give a reward of over 40000 yuan. The platform extracted more than 20000 yuan from it according to the rules, and after the incident, the official withdrawal was refunded to the victim.

In this case, although the platform did not directly participate in the fraud, its business processes were exploited by the fraud.

The interest relationship in advertising business is more direct, as the advertisers suspected of fraud are also platform customers themselves. According to an internal document from Meta (Facebook's parent company) in November 2025, the platform will only directly ban advertisers when the automated review system determines that the likelihood of fraud is higher than 95%; Some advertisers who have not reached this threshold but are still deemed high-risk will not be immediately stopped, but will only need to pay higher advertising fees to the platform.

Another internal document compares fraudulent advertising revenue with potential regulatory costs: a type of fraudulent advertising with higher legal risks can generate approximately $3.5 billion in revenue within six months, while expected regulatory fines are around $1 billion. When high-risk advertising itself is a considerable source of revenue, it is difficult for platforms to consider this revenue when deciding whether to ban it.

Live streaming tipping and advertising placement are both businesses that platforms make money from. Once fraud is hidden in these businesses, the more thoroughly it is managed, the more likely the platform's revenue will be affected.

Let the losses that could have been prevented enter the platform economy account

When the loss belongs to the category of 'could have been prevented', but the platform has not taken necessary measures, it should bear the part that matches its own degree of liability. Returning victims' funds and disposing of platform revenue are only remedial measures after the fact. If the worst outcome for the platform is only to return the portion that should not have been earned, the cost is still limited. Only by allowing the missed risks to potentially result in additional losses, will the platform calculate these risks as part of its own costs beforehand.

What losses can be considered "preventable" depends on how much risk information the platform had before the incident, whether there were practical and feasible intervention measures at that time, and whether timely measures could have avoided the losses. The boundary of responsibility in reality has begun to form around these three issues. The Anti Telecommunication Network Fraud Law requires Internet service providers to recheck the fraud related abnormal accounts monitored and identified, and take measures such as restricting functions and suspending services. Those who violate legal obligations and cause damage shall bear civil liability in accordance with the law. Article 1197 of the Civil Code further stipulates that if a network service provider knows or should know that a user has infringed upon the civil rights and interests of others by using its services, but fails to take necessary measures, it shall bear joint and several liability with the relevant users.

For 'should know', the Supreme Court has provided examples. In an e-commerce platform case announced by the Supreme People's Court in August 2026, the court determined that the platform knew or should have known that the merchant had long emptied their account and evaded refunds based on data such as business scale, capital flow, and complaint records. However, the platform did not take timely measures such as capital control, and therefore should bear joint and several liability with the merchant for repayment.

The timing of platform intervention in fraud risks is shifting from "post event reminders" to "during the transaction process". In June 2026, the fraud prevention mechanism displayed by Alipay and Hongmeng Xingdun shows that in some payment scenarios, the system can intercept within the user's authorization after identifying risks.

For platforms that have long-term investment in anti fraud and have indeed reduced fraud losses, it is also possible to consider providing clearer positive incentives in regulatory evaluation and other aspects. Missing risks comes at a cost, and early governance also requires returns, so that platforms have the motivation to invest more resources before fraud occurs.