Just after chatting with a friend about a certain product, when you turn on your phone, an advertising recommendation pops up. Many people have had similar experiences, and many people will wonder: is the chat being "overheard" by the phone?
Many third-party organizations have conducted surveys on this, and the results show that from the perspective of economic cost, technical difficulty, and legal risks, "eavesdropping" is not as easy to achieve as everyone imagined. The costs and risks for merchants far outweigh the benefits. The reason why users feel like they are being "eavesdropped" is due to the combined effects of factors such as poor protection of personal information on the Internet and iterative upgrades of algorithm technology.
Algorithms are “smarter”
Ms. Qin, who lives in Shaanxi, has found that her mobile phone seems to "understand" her more and more recently. She posted on social media: "Sometimes just a thought may flash through my mind, and the next second it can accurately push ads to me."
The casual comment attracted hundreds of comments. Netizens posted similar experiences: "I just brushed my dog's teeth, and I thought it was difficult to brush the lower teeth. I turned on my phone and pushed me a dog chew that is convenient for brushing the lower teeth." "When I added rice at noon, the spoon got rice on it and it was difficult to wash. After washing the bowl, I pushed a rice spoon that didn't stick to the rice!"
Such accurate and real-time advertising push made many people say "it's amazing" and "it's scary to think about it". Some people are still wondering: How is it done? Is my phone "eavesdropping" on me?
Searching for keywords such as "mobile phone eavesdropping" on the Internet consumer service platform "Black Cat Complaint", you can see a large number of complaints. The targets of the complaints include short videos, e-commerce platforms, and AI smart assistants connected to large models. Some people even suggested that the elderly at home were being "eavesdropped on their daily chats" and were pushed a large number of highly targeted advertising videos, causing family conflicts, and asked the platform to "stop infringement."
Regarding this question, the reporter asked many people in the industry, and the answer was "three noes": "eavesdropping" on the platform is not economically cost-effective, technically unrealistic, and not legally allowed.
It is understood that the current non-real-time voice transcription services with large data scale are priced by mainstream service providers in the market at about 0.6-1 yuan/hour, and real-time recognition prices are even higher. From a technical perspective, if a user is eavesdropped for 16 hours a day, the daily cost is about 9.6-16 yuan/person, and the maximum annual cost per person is close to 6,000 yuan. Including data storage and analysis and other costs, if an App with a daily active user (DAU) reaches 100 million, the cost of "eavesdropping" all year round will reach hundreds of billions.
"The proportion of effective information with real advertising value in users' daily conversations is extremely low. Compared with 'eavesdropping', companies can obtain more accurate business data at a lower cost through users' active clicks and other legitimate actions." said a relevant person in charge of a short video platform.
At the end of 2024, the China Electronics Technology Standardization Institute’s Cyber Security Center Evaluation Laboratory conducted an experimental test on the possibility of App “eavesdropping”. He Yanzhe, deputy director of the laboratory, said that when a mobile phone is “eavesdropped”, it often consumes more energy and suffers from rapid power loss, hot body, and high memory usage.
That being the case, why have algorithmic recommendations become so accurate and “understand” you better? Cross-platform marketing recommendations are one of the main reasons.
"This is called programmatic advertising in the industry." The person in charge explained, "For example, a user searches for a cat food on e-commerce platform A, opens social platform B, and finds that platform B is also recommending cat food. This is not the result of platform B's monitoring, but that platform A, as the advertiser of platform B, uses platform A's data for targeted display, achieving cross-platform advertising and marketing."
"Survivor bias" is also one of the reasons why users feel "overheard." "For example, 1,000 users mentioned cats, and 999 of them pushed content that had nothing to do with cats after opening the app, but one user was pushed with cat content. This abnormal coincidence will be particularly noticeable," the person in charge said.
"Insure" personal information
"Eavesdropping" may be an illusion, but users' needs for information security are real.
"Algorithms that are 'smarter' will indeed expose personal information to more risks." Wu Shenkuo, a professor at Beijing Normal University Law School and deputy director of the China Internet Society Research Center, analyzed, "Cross-platform marketing allows data to flow across multiple platforms. Once a certain link is attacked, the scope of impact will be larger; the data collection process will be more complicated, and users' informed consent will become more difficult."
This requires the Internet platform, legislation, supervision and other parties to work together to "insure" personal information.
It is understood that the App must obtain the explicit authorization and consent of the user before it has permission to access the mobile phone microphone, and it is not allowed to open and record at will without the user's consent. This is a legal red line.
In 2019, the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation jointly issued the "Methods for Determining Illegal Collection and Use of Personal Information by Apps", which clarified the identification standards for six major categories of violations, including failure to disclose collection and use rules, failure to clearly state the purpose, method and scope of collection and use of personal information, collection and use of personal information without user consent, violation of the principle of necessity, collection of personal information unrelated to the services provided, etc.
In January this year, the Cyberspace Administration of China issued the "Regulations on the Collection and Use of Personal Information by Internet Applications (Draft for Comment)", which requires that when an app is launched for the first time, it must "inform users of the rules for the collection and use of personal information through pop-up windows and other conspicuous means" and "should only use camera, microphone and other permissions when the user actively chooses to use functions such as taking pictures, sending voice, recording and recording."
Starting from 2025, relevant special actions will continue to be carried out, and a batch of illegal apps, SDKs (software development kits), smart terminals, etc. will be notified and investigated in batches. The areas involved include live broadcasting, ticketing, catering retail, financial securities, travel and transportation, etc.
From the perspective of industry regulations, formal mobile phone systems have strict restrictions on collecting sound data, and major app stores have also established strict product review mechanisms. If sensitive permissions such as microphones are used, the calling scenario and purpose of use must be explained through privacy policies, system permission application documents, etc. If there is a violation, the application store may be removed from the shelves and bear corresponding legal liability.
Ordinary users should also increase their awareness of data security and proactively protect their own rights and interests. Faye Wang, a senior partner at Beijing Jingdu Law Firm and an expert member of the China Computer Association Data Security Industry Expert Committee, reminded that when downloading an app, users need to carefully read the relevant privacy policy agreement and then check the box. Authorizations that involve privacy and do not affect normal use can be turned off manually to avoid unknowingly collecting personal information.
"To ensure that the algorithm is always good, we can also take a technical governance path." Wu Shenkuo suggested, "For example, using technology to allow multiple platforms to conduct model training without exchanging original personal data, to achieve 'data is available but invisible'."
More "smart" and more positive
When algorithms become “better at understanding you”, cyberspace governance also becomes “better at understanding algorithms”.
On April 27, the "Social Value of Douyin Algorithm" report was released, introducing the positive significance brought by the accurate distribution of the algorithm, including assisting in rumor control, discovering high-quality content, activating regional economic vitality, etc.
Take rumor management as an example. Recently, a piece of news that "a woman fell from a building in Pudong, Shanghai during a typhoon" was circulated on the Internet. In the context of frequent extreme weather, it caused many netizens to worry and fear. After being confirmed as a rumor by the official department, the rumor monger has been subject to administrative sanctions For users who have browsed such false information, the platform will use algorithms to accurately push rumor-refuting content to their information flow. "I was really panicked when I saw it, but I was relieved after seeing the official refutation of the rumor." Xiaoban, a netizen who lives in Zhejiang, left a message under the rumor-refuting video.
In 2025, Douyin launched a large AI truth-seeking model and used the large model to proactively attack rumors. The exposure of rumor content on the platform dropped significantly throughout the year.
"Previously, the platform's identification and handling of rumors mainly relied on user reports, manual inspections, and collaboration with small models, which is like finding a needle in a haystack." The relevant person in charge of Douyin told reporters, "When staff see the rumor refutation information, they will go back to check and refute the rumors. The rumors have already spread." Now with the help of rapid identification and accurate push, rumor refutation information can be directed to people who have seen the rumors. "From users looking for the truth to the truth looking for users."
More "smart" algorithms also have the potential to activate regional economies. In 2025, the first "Scottish Premier League" league became popular all over the Internet, and Douyin-related content was played more than 37 billion times and shared more than 110 million times. With the help of algorithms, the platform will generate event-derived content More accurately pushed to users in different circles such as sports, food, and tourism, helping the popularity of the "Suzhou Super League" spill over from the stadium to the streets. During the competition, the number of Douyin group buying orders in Jiangsu Province increased by 24% year-on-year, and sales increased by 68% year-on-year.
It can be said that the recommendation algorithm is like a huge funnel that operates with precision. It needs to integrate multiple factors such as user and content similarity, time, etc. to select dozens or hundreds of pieces of content from a content pool of hundreds of millions and push it to users. This improvement in precise matching capabilities allows the algorithm to not only accurately "guess what you like", but also more accurately identify risks and support positive and valuable content.
"As algorithms become 'smarter,' platform companies also need to take greater responsibility in meeting user needs and guarding the order of cyberspace." Wu Shenkuo said.
