Is the established input method market landscape about to be upended by Doubao and Qwen?

The competitive landscape of the input method industry has remained untouched for many years.

Over the past decade or more, players including Sogou, Baidu, and iFlytek have firmly occupied seats in this market, building industry barriers through their thesauruses, personalized skins, and accumulated user habits.

In the public perception, input methods are only basic tools that are sufficient for daily use, and no one thought any new player could bring disruptive innovations to this track.

At the end of 2022, WeChat launched its own input method product. Relying on the narrative of "Tencent Ecosystem + Privacy Protection", it soon loosened the originally solidified market pattern.

In the past six months, Doubao and Qianwen have successively entered the market, pushing the input method industry into the AI Native era.

From "Input" to "Expression"

According to this evolution path, I divide input methods into three generations:

1.0 Tool Era: Represented by Microsoft Pinyin and early versions of Sogou Input Method, its core value is "being able to type", and the competition focuses on thesaurus and encoding technology.

2.0 Traffic/Ecosystem Era: Represented by Sogou, Baidu, and WeChat Input Method, the input method has become an advertising carrier and a super App connector.

3.0 AI Native Era: AI manufacturers represented by Qianwen and Doubao enter the market. They completely skip the product logic of the previous two generations, and regard the input method as the "infrastructure" of their AI strategy, rather than an "independent product".

Among them, there is no essential difference between the input methods of the 1.0 tool era and the 2.0 ecosystem era, their business starting point is still to develop input method products, serve as advertising carriers and super App connectors.

In this regard, many views believe that Doubao and Qianwen enter this market to seize the market share of traditional input methods and try to redefine the input method.

In my opinion, the AI input methods represented by the two will definitely take a part of the market share. But the more core consideration is to secure the user entry point based on their own AI ecosystem.

This also makes their competition with traditional input methods not in the same dimension.

Traditional input methods such as Sogou and iFlytek solve the problem of "input efficiency", and AI is only a post-positioned auxiliary function. They help users input the already formed text in their minds efficiently and accurately. But the real expression state of ordinary people is often completely different: scattered ideas, random spoken language, and difficulty in organizing words.

What AI-native input methods make up for is exactly the short board of "expression efficiency". AI manufacturers build the input method into the "first contact point" and "intention executor" between their large models and users: the input box is a dialog box, typing is creation/execution, and AI capabilities run through the whole process of typing, voice input, continuation writing and polishing.

From "input" to "expression", the difference of only one word reconstructs the paradigm of human-computer interaction.

Take my usage experience in several social and fragmented scenarios as an example: on the way to commute, the environment is noisy with people and vehicles coming and going. The AI-powered voice input can filter background noise, transcribe mixed Chinese and English spoken language in real time, and output words almost without delay while you are speaking. The AI will also rewrite the grammatical errors, interjections, pauses and pet phrases in personal expressions, and maximize the optimization of tone and final text output.

At the same time, in office scenarios, for conference oral minutes and quick reports, where traditional voice input methods often output fragmented spoken sentences, AI will automatically sort out logic, divide into layers and paragraphs, delete redundancies, and directly organize scattered oral expressions into standardized weekly reports, meeting minutes and business scripts.

This is the input experience that traditional input methods can hardly provide. Their main AI functions are mostly attached as auxiliary modules hidden in the secondary menu. Users need to manually click to trigger, and manually edit the voice input results for the second time.

Figure: Screenshot of Qianwen Input Method by Tang Chen

Another significant change lies in the usage scenarios. In the past, the capabilities of large models were locked in exclusive Apps, and users needed to actively open them and manually copy and paste content. Now, the input box of AI input method exists almost imperceptibly.

For example, on my MacBook, I set the right Command as a shortcut key, which can call functions to continue writing, polish and organize my expression needs at any time. AI has changed from "a tool you need to look for" to "a capability you can get easily".

Simply put, traditional manufacturers are only "making input method products", while AI manufacturers are "making AI with input methods".

The essence of subverting the market is replacing three sets of rules

For AI input method players such as Doubao and Qianwen, market share is very important. But more importantly, they are overturning the three sets of rules of the input method industry.

The first is the competition logic, which shifts from internet products to AI native. Traditional input methods compete on the size of thesaurus, the number and personalization of skins, output accuracy, traffic monetization efficiency, etc., with the underlying logic still belonging to the internet industry;

The competition focus of AI input methods has shifted to end-cloud collaboration, voice adaptation in complex scenarios, contextual semantic understanding, and personalized expression assistance, which is built on the foundation of large models.

This generational gap cannot be eliminated by simply iterating functions on the original input method products.

The second is product positioning, which shifts from an input tool to an intention entry point. Traditional input methods serve the text that users have fully conceived; AI-native input methods take over vague ideas, scattered spoken language, and unpolished drafts, completing generation, summarization, sorting and rewriting at the input layer.

One of the competitive focuses of input methods in the future will be who can more lightly and quickly capture the real expression intention of users.

The third is the business model, which shifts from traffic monetization to strategic layout. In the past, input methods were mature traffic monetization tools, and advertising, value-added membership services, etc., all rely on the diversion of the input box to form a commercial closed loop.

The well-known example is the "three-stage rocket" model of Sogou Input Method, in which the input method is only a traffic entry point. The price is that users' input rhythm is often interrupted by commercial redundant designs such as information pop-ups, advertisements, and membership promotions.

As the strategic node of Doubao and Qianwen's AI ecosystem, AI input methods have no historical burden, so they can easily choose the "pure mode", with no advertisements, no pop-ups, and no information push throughout the process.

To a certain extent, this is also a return to the original intention of input method products, they exchange pure experience for a permanent AI entry point. Because the input method is the only always-online component that spans all applications, it can inject AI capabilities into every "input-output" process, and then turn AI from an "independent tool" into an "environmental capability".

More critically, the input method carries the most authentic and unmodified human expression data across the entire network. Compared with public data screened by algorithms and manually modified such as short videos and search results, daily spoken language, instant dialogue, and fragmented expressions are the most original corpora that fit real language habits, which can continuously feed back the large model's ability to understand spoken language and generate life-like expressions.

This is the most precious iteration soil for large models, and it is also the essential difference between large manufacturers' heavy investment in input methods and traditional manufacturers' simple addition of AI functions.

From the perspective of layout, Doubao and Qianwen are taking their own paths. Relying on its self-developed Seed speech large model, Doubao is deeply cultivating daily mobile scenarios; relying on the Qianwen large model and CosyVoice speech capability, Qianwen focuses on PC office scenarios to fill the short board of intelligent input in Alibaba's office ecosystem.

One focuses on mobile social scenarios, the other focuses on office scenarios, with different paths but the same direction.

Figure: Comparison between Doubao Input Method and Qianwen Input Method, produced by Tang Chen

Fundamentally speaking, AI manufacturers developing input methods is a strategic long-distance race of "exchanging entry points for data, exchanging data for models, and exchanging models for ecosystems", and the key to victory also lies in the competitiveness of the AI ecosystem.

The existing market pattern of the input method industry will not be easily overturned by them.

For most ordinary users, the stable and comfortable basic typing experience is still the core standard for choosing an input method. Traditional input methods have been integrated into users' usage habits in terms of thesauruses, cloud memory, and ecological services. If the AI input method cannot guarantee accurate typing, no matter how powerful its functions are, it will not become the first choice for users.

At the same time, as a new thing, AI input methods also have obvious shortcomings. For example, the PC version of Doubao Input Method has relatively simple functions, and is insufficient in in-depth polishing of long texts and structured text generation; Qianwen's real-time voice input has slight delay, and its lightweight daily typing experience is not as smooth as that of traditional input methods.

AI input methods also have a native shortcoming. Half of its capability ceiling depends on model technology, and the other half is restricted by the boundary of privacy compliance.

This is not difficult to understand: AI input methods need full-scenario reading entry points, and the model needs in-depth understanding and polishing, which cannot avoid reading context and even cloud analysis, bringing privacy anxiety and compliance risks at the same time.

At the same time, iFlytek, Baidu, and WeChat are also continuously iterating their AI functions to narrow the experience gap.

So what will be the final outcome of this game?

I have a prediction: In the future, the input method market will not be a zero-sum game of replacing the old with the new, but will form a two-track pattern of hierarchical coexistence.

Traditional input methods rely on stable basic experience and deep user accumulation to hold the mass stock market; AI-native input methods rely on intelligent expression assistance to win high-quality user groups such as office workers, content creators and experience-sensitive users.

The entire industry will be fully AI-enabled, and products that purely rely on traditional input capabilities will gradually fall behind.

The value brought by AI input methods may have gone beyond the industry itself. In the past, there were obvious individual gaps in writing skills, logic sorting, and workplace expression level; now, AI polishing and spoken language to text enable ordinary users with weak expression ability to quickly output appropriate and standardized texts.

This is also a kind of AI inclusive benefit. The input method industry has shifted from competing for input efficiency to competing for expression capability, and from competing for traffic to competing for entry points, its game rules are being rewritten by AI manufacturers such as Doubao and Qianwen.

Of course, they also hope to have greater right to deal cards at the input method market table.

This article is from the WeChat official account "Tang Chen's Official Account", author: Tang Chen, published with authorization from 36Kr.