Tencent, Alibaba and other Chinese tech companies block Doubao’s AI assistant, highlighting a battle over access, security and control of mobile services.
By Lian Ran
ByteDance’s Doubao AI phone assistant has hit resistance from some of China’s biggest technology companies. On Oct. 14, the 30-day public consultation period for its Screen Automation Execution Protocol (SAEP) ended, with 155 third-party apps listed as having explicitly refused to allow AI-driven operations.
The list initially contained 151 apps, but the figure had risen to 155 by the time of publication. It includes 66 apps from Tencent (0700.HK), 23 from Alibaba (9988.HK) (BABA.US), 11 from Meituan (3690.HK) and 14 from Kuaishou (1024.HK). Even rival AI products, including Kimi and DeepSeek, have declined to grant access.
The refusals mark an early confrontation over control of the mobile interface in the age of AI agents. As assistants gain the ability to operate screens and complete tasks across apps, they threaten to disrupt a traffic and business model that has dominated China’s mobile internet for more than a decade.
SAEP brings the rules into the open
Before SAEP, whether a phone AI assistant could operate an app, and to what extent, was decided unilaterally by AI vendors’ technical adaptation. App developers had no formal channel to state their position, mostly intercepting silently through risk-control strategies. Users faced blurred functional boundaries, never knowing why some apps worked with AI and others didn’t.
SAEP aims to return the choice of screen automation to third-party developers. They can declare whether to allow Doubao Mobile Assistant to perform screen recognition, simulate taps and fill in content within their apps, and can specify permissions for particular pages or functions.
During the 30-day consultation period, third-party apps were closed to AI operations by default, with access granted to those that explicitly agreed. After the period ended, apps that had not expressly refused could be opened up progressively according to their risk level. Apps on the refusal list remain off-limits, although developers can change their declaration status at any time.
Doubao’s system-level permissions allow it to read and interact with screens in the background, unlike conventional accessibility tools that require foreground operation. The restrictions apply to AI automation, not ordinary use: users can still open and operate the apps manually.
The protocol formalizes a conflict that had already surfaced during Doubao’s development.
From hidden conflict to public rules
In December 2025, ByteDance and smartphone maker ZTE’s Nubia brand released a technical preview on the Nubia M153 engineering handset. This was Doubao’s screen automation debut, but adaptation conflicts with third-party apps soon surfaced. Users reported login anomalies, frequent CAPTCHA prompts, and crashes in WeChat, Taobao, Xianyu, and Damai. In competitive games, automation triggered risk-control judgments.
Doubao subsequently restricted automation in areas including financial apps, competitive games and applications offering rewards for repeated actions. But the underlying dispute remained unresolved.
The consumer version launched on Sept. 14, alongside SAEP. The Nubia NaviX Ultra, equipped with the assistant, went on sale two days later.
Less than two weeks after that launch, users reported that Honor of Kings, Tencent’s popular mobile game, was flagging the handset for an abnormal device environment. Some accounts were logged out or temporarily suspended for suspected use of third-party tools.
Doubao apologized the following day, saying its investigation found that the assistant had not interacted with the game system or performed prohibited simulated actions. It said it was communicating with Tencent. Subsequent media tests reportedly triggered the same security controls even when GUI automation was disabled.
The episode illustrates how conflicts previously handled through technical restrictions and security systems are now being brought into the open.
The refusal list reveals a divide
Tencent accounts for the largest share of refusals, with 66 apps spanning WeChat, enterprise messaging, QQ, office tools, music and gaming. Alibaba’s 23 include Taobao, Tmall, Alipay, Ele.me, Amap and Xianyu. Meituan and Kuaishou have also refused access across their listed apps.
Other prominent services include Didi, Xiaohongshu, Pinduoduo, Zhihu, Baidu Maps, WPS and several major banks and mobile operators. Kimi and DeepSeek’s presence suggests that resistance is not limited to traditional internet incumbents: competing AI providers also have reasons to restrict another assistant’s access.
By contrast, ByteDance’s own apps and affiliated services are supported by default. JD.com’s apps do not appear on the refusal list, nor do major video platforms such as iQiyi, Mango TV and Bilibili. Official vehicle-control apps from BYD, Nio, Xpeng and Li Auto are also absent, as are most major state-owned banks and leading securities firms.
Absence from the list does not necessarily mean these companies have actively embraced AI automation. But the pattern suggests a possible dividing line: platforms with strong proprietary traffic and self-contained ecosystems have greater incentives to protect their direct relationships with users, while businesses that depend more on external orders may have more reason to remain open.
Security and commercial tensions unresolved
The dispute is not simply a commercial-interest game to protect market share. Screen reading and simulated clicks raise genuine questions about security, privacy and accountability, particularly when AI handles payments, logins, publishing or data deletion.
If an agent makes an unauthorized payment, sends the wrong message or triggers an account restriction, responsibility could fall on the AI provider, the app developer or the user. There is no settled industry-wide framework for allocating that liability. Banks and social platforms may also reasonably treat simulated interactions as potential fraud or unauthorized automation.
SAEP’s significance is therefore broader than its list of permissions. It creates a public mechanism for developers to define their boundaries and gives users a clearer indication of what an AI assistant can do.
The longer-term balance will likely depend on three parties: operating-system providers, which control system-level capabilities; app developers, which set limits on operations within their services; and users, who decide whether to trust a particular assistant.
The refusal by 155 apps is only the beginning. More detailed permissions, clearer responsibility rules and agreements over commercial benefits will be needed before AI assistants can move freely between mobile services.
For now, Doubao’s experience shows that technical capability alone will not determine the success of mobile AI. The willingness of established platforms to open their ecosystems may prove just as important.
Source:
geekparkGO
