The input method industry has long remained stagnant.

For the past decade, giants like Sogou, Baidu, and iFLYTEK have held sway, constructing formidable industry barriers through expansive word banks, personalized skins, and entrenched user habits.

In the eyes of the public, input methods are often seen as mere basic tools, with little anticipation for innovation.

By late 2022, WeChat disrupted the status quo by launching its own input method product, leveraging the narrative of the 'Tencent ecosystem + privacy protection,' which swiftly loosened the previously solidified market landscape.

In the past six months, Doubao and Qianwen have entered the fray, propelling the input method into the AI-Native era.

From 'Input' to 'Expression'

Tracing this evolutionary trajectory, I categorize input methods into three distinct generations:

1.0 Tool Era: Represented by Microsoft Pinyin and early Sogou, the core value lay in 'typing capability,' with competition centered on word banks and encoding techniques.

2.0 Traffic/Ecosystem Era: Led by Sogou, Baidu, and WeChat, input methods evolved into advertising platforms and connectors to super-apps.

3.0 AI-Native Era: Spearheaded by AI vendors like Qianwen and Doubao, they bypass the logic of the first two generations, viewing input methods as 'infrastructure' for their AI strategies rather than standalone products.

Input methods in the 1.0 Tool Era and 2.0 Ecosystem Era share fundamental similarities; their primary focus remains on creating input method products, advertising platforms, and super-app connectors.

Many believe Doubao and Qianwen aim to capture the traditional input method market and redefine the concept of input methods.

In my view, the AI input methods they represent will indeed carve out a portion of the market. However, their primary objective is to secure user entry points within their AI ecosystems.

This places their competition with traditional input methods on a different plane.

Traditional input methods from Sogou and iFLYTEK focus on 'input efficiency,' with AI serving as a supplementary function. They assist users in efficiently and accurately inputting pre-formed text from their minds. However, ordinary users' actual expression states often differ: scattered ideas, casual speech, and difficulty organizing text.

AI-Native input methods bridge the gap in 'expression efficiency.' AI vendors position input methods as the 'first touchpoint' and 'intent executor' for their large models and users: the input box transforms into a dialog box, typing becomes creation/execution, and AI capabilities permeate the entire process of typing, voice input, continuation, and polishing.

The shift from 'input' to 'expression,' though subtle in wording, reconstructs the human-computer interaction paradigm.

Take my experience in various social and fragmented scenarios as an example: During my commute, amidst noise and bustling traffic, AI voice input filters background noise, transcribes mixed Chinese and English speech in real-time, and outputs text almost instantaneously. AI also corrects grammatical errors, interjections, pauses, and catchphrases in personal expressions to optimize tone and final text output.

Meanwhile, in office scenarios like meeting minutes and quick reports, traditional voice input methods often struggle with transcribing fragmented speech. AI automatically organizes logic, segments layers, and removes redundancy, transforming scattered verbal expressions into structured weekly reports, meeting minutes, and business scripts.

This represents an input experience that traditional input methods struggle to provide. Their AI features are mostly external modules hidden in secondary menus, requiring manual triggering and secondary editing of voice input results.

Figure: Qianwen Input Method Screenshot by Tang Chen

Another significant change lies in usage scenarios. Previously, large model capabilities were confined within dedicated apps, requiring users to actively open and manually copy-paste. Now, the input box of AI input methods feels almost invisible.

For instance, on my MacBook, I set the right Command key as a shortcut to invoke continuation, polishing, and organization of my expressions at any time. AI has transformed from a 'tool to seek' to a 'capability at hand.'

In essence, traditional vendors are merely 'making input method products,' while AI vendors are 'leveraging input methods for AI.'

The Essence of 'Revolution' Lies in Replacing Three Sets of Rules

For AI input method players like Doubao and Qianwen, market share is important. More crucially, they are overturning three sets of rules in the input method industry.

First, the competitive logic shifts from internet products to AI-Native. Traditional input methods compete on word bank size, skin quantity and personalization, output accuracy, and traffic monetization efficiency—still rooted in the internet era.

AI input methods now focus on end-to-end cloud collaboration, voice adaptation in complex scenarios, contextual semantic understanding, and personalized expression assistance, with large models as their foundation.

This generational gap cannot be bridged by simply iterating functions on existing input method products.

Second, the product positioning shifts from an input tool to an intent entry point. Traditional input methods serve users with fully conceived text; AI-Native input methods handle vague ideas, scattered speech, and rough drafts, completing generation, summarization, organization, and rewriting at the input layer.

A future competitive focus for input methods will be who can lighter and faster capture users' true expression intents.

Third, the business model shifts from traffic monetization to strategic positioning. Input methods were once mature traffic monetization tools, with advertising and membership value-added services relying on input box traffic for commercial closure.

A well-known example is Sogou Input Method's 'three-stage rocket' model, where the input method merely serves as a traffic entry point. The cost is that users' input rhythms are frequently interrupted by commercial redundancies like news pop-ups, ads, and membership promotions.

AI input methods, as strategic nodes in Doubao and Qianwen's AI ecosystems, carry no historical baggage and can freely choose a 'pure mode' with no ads, pop-ups, or news pushes.

To some extent, this represents a return to the original purpose of input method products, exchanging a pure experience for a permanent AI entry point. As the only component always online across all apps, input methods can inject AI capabilities into every 'input-output' process, transforming AI from a 'standalone tool' into an 'ambient capability.'

More critically, input methods carry the most authentic, unfiltered human expression data across the web. Compared to algorithmically filtered and artificially refined public data from short videos and searches, daily speech, instant dialogues, and fragmented expressions are the most authentic native corpora, continuously feeding large models' spoken language understanding and lifelike expression capabilities.

This is the most precious iterative soil for large models and the fundamental difference between big tech's heavy investment in input methods and traditional vendors' simple AI overlay.

From a strategic perspective, Doubao and Qianwen are playing different cards. Doubao leverages its self-developed Seed voice large model to deepen daily mobile scenarios; Qianwen relies on the Qianwen large model and CosyVoice voice capabilities to focus on PC office scenarios, filling Alibaba's office ecosystem's intelligent input gap.

One targets mobile social, the other office scenarios—different paths, same direction.

Figure: Doubao vs. Qianwen Input Methods Chart by Tang Chen (Click to enlarge)

Fundamentally, AI vendors' foray into input methods is a strategic marathon of 'exchanging entry points for data, data for models, and models for ecosystems,' with the outcome hinging on AI ecosystem competitiveness.

The Input Method Table Won't Be Easily Overturned

For most ordinary users, a stable and smooth basic typing experience remains the core criterion for choosing an input method. Traditional input methods have integrated into user habits with word banks, cloud memory, and ecosystem services. If AI input methods type inaccurately, no matter how powerful their features, they won't become users' first choice.

Meanwhile, AI input methods, as new entrants, have obvious shortcomings. For example, Doubao Input Method's PC version has relatively simple functions, lacking deep polishing and structured writing capabilities for long texts; Qianwen's real-time voice input has slight delays, and lightweight daily typing isn't as smooth as traditional input methods.

AI input methods also have an inherent limitation. Their capability ceiling depends half on model technology and half on privacy compliance boundaries.

This is understandable: AI input methods seek full-scenario read access, requiring deep contextual understanding and polishing, which inevitably involves reading contexts and even cloud analysis, raising privacy concerns and compliance risks.

Meanwhile, iFLYTEK, Baidu, and WeChat continue to iterate AI functions, narrowing the experience gap.

So, what will the endgame of this table be?

I predict: The future input method market won't be a zero-sum game of old replacing new but a dual-track landscape of tiered coexistence.

Traditional input methods will hold the mass market with stable basic experiences and deep user accumulation; AI-Native input methods will capture high-quality user groups like office workers, content creators, and experience-sensitive users with intelligent expression assistance.

The industry as a whole will fully embrace AI, with products relying solely on traditional input capabilities gradually falling behind.

The value brought by AI input methods may already exceed the industry itself. In the past, significant individual gaps existed in writing skills, logical organization, and workplace expression; now, AI polishing and spoken-to-written conversion enable ordinary users with weak expression abilities to quickly output decent, structured text.

This is also a form of AI inclusiveness. As input methods shift from competing on input to expression, from traffic to entry points, their rules are being rewritten by AI vendors like Doubao and Qianwen.

Of course, they also hope to gain greater card-dealing rights at the input method table.