I don’t know if you have noticed that these days, even the input method is becoming more and more difficult to use. It's not that it's completely difficult to use, but it's secretly and intentionally difficult to use. It's like it seems to know exactly what you want to type, but it goes against you and deliberately targets you. The most common thing is that commonly used words are mysteriously missing. For example, the word "是" is often pushed to the first place by "thing"; you obviously want to type "difficult to use", but the first choice is "for men".
Not only are commonly used words not given priority, word parsing is often incredibly wrong.
If you want to type "right to left", it will pop up "onion knife holder" for you; if the programmer wants to enter "Python", the result will be "red butt hole";
If you want to make a "perseverance", your first choice is "Yili", which also makes many people think that this is a smuggled product.
The ability to associate idioms with ancient poems has also dropped off a cliff. By typing "ypzyflbylh", it used to be able to perfectly identify "it's a pleasure to have friends come from afar", but now, it can give you "spicy clothes".
I tried to type "I came at a bad time", but it was recognized as "I lied to my teacher about not having plump lips"? ? ?
Sometimes, you type a word and subconsciously press confirm. When you look at it, you select the wrong word. The word you want is in the second one.
So you typed it again, and this time you learned the lesson and pressed 2 directly. As a result, the input method also predicted your prediction and pushed the word you mistakenly selected last time to the second position, while the correct word returned to the first position.
Do you think this is a problem unique to Android third-party input methods? The Apple Party would be happy too early.
After the iOS 26 upgrade, many users reported that Chinese pinyin input had been significantly degraded, and the association function had greatly regressed. In the past, they could associate a complete sentence, but now they can only enter two or three characters separately.
For example, if you enter woyebuzhidao, you think the first choice should be "I don't know either", right? Sorry, what it gives you is "Play with it and don't knit it too much"; enter nidaodizaiganshenme, and the first choice is "If you are big enough, pass it on to me."
Play Oh, don’t knit the big apples, hand them over to the big liver! ! !
We took a look and found out that similar complaints about the input method to reduce intelligence can be said to be a "ghost story" that has been rumored for a long time in the world.
It's just that nowadays, when it comes to this kind of "reverse intelligence", everyone has begun to habitually blame AI.
But in fact, AI cannot bear all the blame.
In January this year, Sogou specifically explained to AI that the core foundation of the input method is the understanding of pinyin strings, words and uses. AI will not pollute these links. Officials even said that our input method has not become stupid at all.
According to background data, typing accuracy is slowly improving year by year, which is different from what everyone feels.
In fact, the two have different definitions of accuracy. On the input method engineer's side, as long as the target word appears in the top five candidate boxes, the input is considered successful; but for the user, the first word that comes out is not the desired word, and this input method is mentally retarded.
So if AI is not the fault, then why?
The first one is that the explosion of Internet corpus has caused serious pollution of data.
How does the input method know what words you want to type?
Simply put, there is a huge statistical language model behind it, based on probability theory, using the previous text to predict the next most frequently occurring word.
This system worked very well in the past because there were few Internet users in the early days and the vocabulary was relatively concentrated. A set of general cloud vocabulary libraries (the corpus mainly comes from news, publications, and standard web text, and the word frequency distribution is very stable) can cover the input needs of most people.
But things are different now. Markets such as e-sports, 2D, fandom, short videos, etc. are creating and discarding massive vocabulary at an exponential rate every day.
In order to meet the needs of these people, the input method's vocabulary must continue to expand its corpus sources.
In order to cover these new words, manufacturers have to introduce large-scale spoken language materials such as e-commerce search logs and short video comment areas.
The amount of data has increased, but the quality has declined.
Of course, the most serious source of pollution may be the users themselves. Many people are not so rigorous when typing. They type a large number of typos in Pinyin, but the message is still sent without error. For example, the modal particle "ya" at the end of the sentence is often typed as "pressure" or "ya". Who hasn't been confused by their parents' typos?
This may be easier for us humans to handle. We can figure it out by making a guess based on the context and the person we’re chatting with.
But for the input method, after these texts with typos enter the vocabulary, it can't figure out who is right and who is wrong.
There are also some people who, because of the strict review of the platform, deliberately use typos to avoid suspicion (such as words such as broadcast room), which is also further contaminating the input method's vocabulary.
Ironically, these situations occur not because the input methods have become stupid, but because they have become smarter.
In many posts complaining about input methods, there are always people who miss the era of feature phones. At that time, you could carry a PHS in your pocket to reply to messages, so many people think that is smart.
But to be honest, everyone found it easy to use at that time precisely because the input method was too stupid. One key corresponded to three or four letters, the vocabulary was rigidly fixed, and the association function was almost non-existent.
In this way, the order of words is always fixed, and when you reach a certain stage, it is certain which word appears in which position.
After using it for a long time, your fingers will form muscle memory. Pressing the confirmation button a few times and scrolling down a few grids are all automated operations. Its ease of use is entirely due to people adapting to the tool.
Today's input methods have become smarter. They have to actively learn from you, predict you, and adapt to you. As a result, the frequency of words is constantly changing, and the positions of candidate words are also dynamic. Yesterday, the first word was "yes", but today it becomes "thing". Users have no time to form muscle memory, thus further deepening the impression that it is not easy to use.
Moreover, there is another reason that cannot be ignored. In the past, if I changed my mobile phone or something, I would log in to my account and the vocabulary library would follow. The input habits I have used for many years will not be lost at all, and part of this relies on radical data collection and cloud synchronization.
However, in the past few years, the "Personal Information Protection Law" and the "Data Security Law" have been implemented one after another, and input method manufacturers have been forced to comprehensively reduce data collection.
Cloud vocabulary synchronization is not necessarily enabled by default. Users must actively bind accounts and actively authorize. Many people have to retrain the input method from scratch when changing devices, which naturally makes it difficult to use.
Does that mean that all manufacturers are out of business? That's not necessarily the case.
Because Sogou, iFlytek, and Baidu account for 96% of the market share of third-party mobile phone input methods, the entire market is like a pool of stagnant water.
If you look at Sogou, iFlytek, and Baidu, the long-term update announcements are basically "fixing known bugs." The last time the QQ input method had a substantial update on the PC was several years ago.
Then their focus has changed to skins, ads, and adding messy new features. Who bothers to worry about input accuracy?
Moreover, after the mobile Internet wave, PC-side input methods have been left out of favor to the point that manufacturers themselves have to admit it. Sogou has said that the update speed of the computer side is much slower than that of mobile APPs.
But the real demand for input methods is on the PC side. Input method errors encountered at these times will be amplified into the overall impression that the entire input method brand has become dumb.
Having mentioned so many crimes against the input method, are you feeling a little desperate? But don't worry, things have turned around.
As mentioned earlier, AI cannot be blamed for input methods becoming difficult to use. In fact, AI cannot be blamed. If input methods want to make a qualitative leap, they must ultimately rely on AI..
The AI mentioned here is not those fancy functions such as "helping you write", "helping you reply with high emotional intelligence", "one-click association search", etc., but actually using the ability of large models to improve the level of input methods.
Remember what we said earlier, what an input method should do most is to have human-like thinking and guess what you want to say and what you want to express.
Coincidentally, isn’t this the working principle of large AI models?
So you can also see that in the past two years, Tencent’s WeChat input method and ByteDance’s Doubao input method have entered the market successively, and many of our editorial departments have become voluntary promoters of both.
On the one hand, the WeChat input method has successfully captured a piece of the market by relying on its own ecosystem and its unique selling point of transferring files across devices.
Subsequently, the Doubao input method did not have the hard scrolling function at the keyboard typing level and the old input method. They directly used Byte's own Seed-ASR speech recognition model to make a dimensionality reduction attack.
The error rate of this model on the public test set can be reduced by up to about 40% compared to similar domestic models.
What concept? It's the kind of thing where after you finish speaking, the text on the back has been recognized, punctuation has been added, and capitalization has been standardized. Sometimes I even say the wrong word, and it can help me correct it based on the context.
Moreover, this guy really has the ability to associate context.
Our editorial department now often discusses the AI model Claude. When inputting voice, it will initially recognize it as Cloud - after all, the pronunciation is similar. But when the keywords "large model" and "AI" are mentioned later, it will automatically correct the previously recognized Cloud to Claude, and the whole process is very fast, almost real-time correction, and quite smooth.
In the past two decades, the input method has changed from a pure typing tool to an advertising container, a hodgepodge of functions, and a data island bound by privacy regulations. The explosion of Internet corpus has polluted its vocabulary, excessive commercialization has hollowed out its core, and the old technical architecture has come to an end.
But as the new input method players open their books, we also see a possibility: the future of input methods is not to help you write love letters, nor to help you compile text for your circle of friends.
Instead, stay quietly at the bottom of the screen and return your muscle memory to you.
The day when it truly eliminates the need to look down at candidate words, that useful input method will truly come back.