Previously, we looked at the overall category of AI writing assistants and explored how each of the three broad classes work and what their different use cases are. In this post we focus specifically on the text prediction class of AI writing assistants.
Text prediction software helps users write faster and more accurately by providing real-time suggestions of letters, words and even entire sentences, as the user types. By analysing what has already been written, the AI can make accurate predictions of what the user is likely to want to write next. The user can then quickly select the suggested word or phrase, (normally by using the ‘tab’ key) which saves keystrokes and therefore time.
We are starting to see these type of AI writing assistants integrated into various text input software such as email composers, word processers, customer support tooling and CRMs. With browser-based predictive text tooling, the scope of their use is very wide as they can be applied to almost any browser-based situation.
The abilities of predictive text AI is rapidly increasing in sophistication as the field of NLP and the associated language models become ever more advanced. It’s expected that use of such tooling will increase substantially over the next 5 years, with such functionality eventually becoming ubiquitous for almost any text input scenario (in the same way that spellcheck is no longer seen as a feature, but rather an expected standard in most text entry situations).
Although the category of AI writing assistants is large with dozens of products sitting under the umbrella term, the number of true text prediction tools remains quite small. This is in part because of the complexity of predictive text modelling. Let’s look at the main players who offer some form of predictive text capabilities.
Users of Gmail will likely be already familiar with their ‘Smart Compose’ predictive text system. In the big G’s own words, ‘smart compose helps you write emails faster’. It does this a few ways; firstly, the autocomplete functionality which helps you quickly find the person you want to send your message to (if you have already sent them at least one email previously).
Then it’s smart enough to add the first name of the recipient after you type ‘Hi’, ‘Hello’ or another standard greeting. It will then suggest a number of common opening lines after you type the first word or two.
For example, if you type:
‘How’ it will suggest the words 'are you?’
It also suggests common closing phrases, such as
‘All the best’
There is a feedback function which allows you to give direct feedback to Google on any specific suggestion.
Finally, Gmail has a standard spellcheck built in and will pick up on typos and some grammatical errors, providing you with suggestions.
Limitations: Smart Compose saves a bit of time when writing an email, but currently is quite limited in its functionality. Don’t expect it to be composing entire emails for you quite yet. The system is also limited to Google products only, so you are not able to use its predictive text capabilities anywhere else outside of Google’s ecosystem.
The functionality of Outlook’s predictive text system is very similar to that of Gmail’s Smart Compose. The level of functionality is on par, with the same sorts of word suggestions being offered, after one or two words have been typed. It also has a spellchecking functionality.
Microsoft’s prediction system seems slightly more sophisticated than Google’s offering, with more suggestions being offered. When we compared them side by side, typing the same words into each, in several scenarios Outlook suggested relevant phrases, where Google suggested nothing.
However, as text prediction AI tends to learn the more you use it, it’s possible that this was due to the amount of ‘training’ rather than the sophistication of the AI itself.
Limitations: As with Smart Compose being locked for use in Google products only, so is the case with Text Suggestions only being available within the Microsoft ecosystem. It’s also not advanced enough yet to offer many word suggestions, but it’s likely it will get better and better.
Apple iPhone keyboard has an option to activate predictions where suggestions for next word or emojis are offered, based on recent typing activity, and information from use of other apps.
Apple’s predictive text functionality has mixed reviews; there are numerous articles reporting that the quality of prediction and autocorrection has decreased and even articles explaining how to turn off text prediction.
However, in our own assessment of Apple keyboard's capabilities, Apple performs better than Google in prediction accuracy and autocorrection abilities.
Limitations: Currently Apple only offers its predictive text functionality via the iPhone keyboard, with no availability when using a laptop or desktop device. So even though the quality of Apple’s predictive text is not bad, its scope is limited to largely individual consumer use.
Typewise (that’s us) is a Swiss deep tech company that builds advanced text prediction and autocorrect technology, that surpassess that of Google and Apple. We initially used this AI to power our smartphone keyboard app which has been downloaded more than 2 million times and has won the CES Innovation Award two years in succession.
We now also use our text prediction engine to power our AI writing assistant tool for customer service teams, which significantly increases the productivity, quality and consistency of writing and replying to emails and support chats.
In addition to its off-the-shelf text prediction capabilities, Typewise uses a custom AI model, that is able to learn specifically from the company’s previous communications, and therefore provide word and sentence predictions that are tailored for the business in terms of style, tone, quality and common phrasing. Typewise works on the client’s own servers, meaning the system offers 100% privacy.
Limitations: As a B2B product, Typewise AI writing assistant is aimed at customer support teams, rather than individual consumers, so only businesses can currently use their advanced technology.