Why is the AI ​​mobile phone that can do things in one sentence said "no" by major applications?

Recently, my country's first AI smart phone that has completed the complete compliance process from large model registration to terminal network access has been put on sale. But just after it was launched, the intelligent operation was rejected by many App manufacturers. What's going on?

Recently, the "Doubao Mobile Assistant", which focuses on the concept of an intelligent agent that can do things in just one sentence, has been collectively rejected by more than a hundred applications. It does not mean that these apps cannot be used on Doubao phones, but that these applications prohibit Doubao Mobile Assistant from helping users to control applications. Why do major apps join forces to say "no" to such a convenient function?

In fact, this is an industry collision that will inevitably occur when AI moves towards real industrial implementation.

From the perspective of Doubao or all smart phones, the original intention is good, and it is more intelligent to meet user needs. However, this demand will inevitably break the existing balance between various apps and break through each "information island". However, because the data of each company is not interoperable, Doubao Mobile Assistant can only use "screen automation" technology at this stage, which means that AI can operate the phone and complete various tasks by understanding the screen and simulating human clicks and slides.

From the standpoint of major apps, there are also very practical considerations. The first and biggest concern is safety and who will be responsible. If the person who clicks on the screen is not the owner, but an authorized agent, once the agent is implanted with a virus or its privacy is leaked, the security responsibilities have not yet been clearly defined. Second is the business ecosystem. App manufacturers are required to hand over all or part of their hard-earned user relationships, core data and traffic portals to third-party AI for scheduling. This is somewhat idealistic within the existing business logic before a specific cooperation model is reached.

At present, the global technology circle is still groping on the road to making AI "collaborate across applications". In addition to "screen reading" like Doubao, there are also cooperation routes such as "directional API interface", which means that the platforms agree on three chapters in advance to delineate permissions and responsibilities. Objectively speaking, the former is highly versatile, but is prone to running into ecological and security defense lines; the latter has clear rights and responsibilities, but lengthy business negotiations make it difficult to cover all applications in a short period of time.

In fact, this kind of collision between technological breakthroughs and business order is not uncommon in the history of the Internet. In the early PC Internet era, search engines crawled all over the Internet to crawl data, which also triggered collective panic and boycott of websites. It was not until later that the industry proposed the "Robots Protocol", which allows websites to declare through code independently what is allowed to be crawled and what is refused to be crawled. This initiative eventually evolved into a common standard adhered to by the entire industry, laying the foundation for the development of search engines.

Today, if an AI agent wants to penetrate the high wall of the App again with "one sentence", it also has to go through a long ice-breaking process. It is worth noting that in order to help itself out this time, the Doubao team also proposed a solution, the Screen Automation Statement of Operations Protocol (SAEP), and simultaneously launched a 30-day rule announcement. According to this agreement, third-party applications can independently declare operating boundaries and explicitly allow or deny AI assistants to perform screen automation operations within this application. Doubao Mobile Assistant stated that it will fully respect the wishes of developers and will not perform automated operations on third-party applications that explicitly refuse relevant operations.

Zheng Senwen, a senior engineer at the Institute of Software, Chinese Academy of Sciences: The SAEP protocol mainly solves these problems: First, the "one size fits all" platform risk control, the platform often confuses user-authorized agent behavior and malicious plug-in behavior. Second, conflicts in the past broke out after the fact. Now there is a 30-day public notice period to turn conflicts after the fact into negotiations beforehand. Third, there are millions of applications on mobile phones, and the standardization statement reduces negotiation costs. For enterprise initiatives to become an industry consensus, it also needs to rely on the efforts of all parties to respond and game. The competent authorities also need to provide guidance and standardization, and finally settle into industry standards.

The development of technology is inseparable from standards, specifications and multi-party cooperation. It requires everyone to cross the river by feeling the stones together. How to break down barriers while maintaining a safe bottom line? This is not only the only way for technology to move forward, but also a difficult problem faced by the industry and managers before entering the era of intelligent agents.