Predictive text in 'demonstrable decline' with introduction of AI-based language models

The next "butks" stop. Eating a "banns bc a". It's "mi longer shiny sync".

The above gobbledegook is what my phone dished up the other day when I was texting the words "bus", "banana", and "no longer sensible", admittedly in a hurry and on an aging operating system.

But if internet chatter is to be believed, the intuitiveness and efficiency of autocorrect and predictive text have fallen off a cliff in recent years.

Online forums moderated by leading mobile phone companies Samsung and Apple show long lists of complaints about the declining intelligence of their predictive text and autocorrect models.

These include autocorrecting a correct word for a wrong word, repeating the mistake several times even if the user deletes and retypes, failing to capitalise letters such as a single "I", or inserting nonsensical words such as "tondel" when a person was trying to write "to" or, as one person said, consistently writing "Theresa" instead of "the". 

Others include reducing two words to contractions that are not relevant to the message, changing words like "good" to "food" and, as was the case referenced at the top of this story, correcting words with the nonsensical arrangement of letters.

Contributors to forums have blamed operating system updates and stated that once useful and fairly accurate software options seemed to be facing a litany of errors with the introduction of AI from 2023.

A 'demonstrable decline'

PhD candidate Leon Furze is studying the implications of generative artificial intelligence on writing instruction and education.

"There's been a demonstrable decline in 'old-fashioned' predictive text, as well as several deliberate patches/updates from companies to fix them," he said.

"Back in 2023, Apple started to use an on-device machine-learning language model, a transformer model like [Chat] GPT, to improve the predictive keyboard.

"This replaced/augmented a deterministic model that had been in use since around 2007."

He said Apple writing tools were introduced in its subsequent iOS releases, "which added even more AI into the mix".

"Samsung followed a similar pattern in the years [from] 2023 onwards," Mr Furze said.

Anecdotal evidence

Morteza Namvar is a University of Queensland Business School senior lecturer and Centre for Enterprise AI affiliate.

He said there was "considerable anecdotal evidence and growing public discussion" by users about a perceived decline in autocorrect and predictive text quality.

"Whether there has been an objective deterioration is difficult to establish, as these systems have become substantially more sophisticated over time," Dr Namvar said.

But he said users did appear to encounter more "unexpected corrections and unusual suggestions" than they did with earlier predictive text technologies and that the increasing integration of AI-based language models was likely to be a contributing factor.

"Earlier predictive text systems relied primarily on dictionary-based approaches and relatively simple statistical language models, which generally produced conservative and predictable outputs," Dr Namvar said.

"Contemporary systems increasingly employ neural language models that attempt to infer context, meaning, and user intent.

"From a scientific perspective, this should not necessarily be interpreted as a decline in capability. Rather, it reflects the trade-off associated with moving from rule-based and statistical approaches towards AI-driven language prediction, where improvements in flexibility and contextual understanding are accompanied by increased variability in system outputs.

"Consequently, while modern systems may exhibit more sophisticated linguistic capabilities, their errors can be more noticeable and occasionally appear less intuitive to users."

An unpredictable history

Predictive text was first adopted by mobile phone companies in the late 1990s with T9 software, or "Text on 9 keys".

It associated several letters with each key and, based on the order in which you pressed them, predicted what word you were trying to write, or gave you several options.

Push-button QWERTY keyboards came into play on some phones, but T9 remained a prominent texting system until the advent of smartphones and touchscreen technology took the market by storm in the late 2000s.

Through the use of statistical n-grams, or sequences of words or characters in a text, algorithms determined the highest probability for the next word based on previous input into a phone's QWERTY keyboard.

It meant predictive text was no longer just offering up words based on the order a keypad was pushed, but was predicting the next word, and even phrases.

But from about 2023, when phone companies started introducing AI into the mix, things started taking a different turn.

Phones 'learning' mistakes

Mr Furze said transformer-based AI could help in that it could "learn" user patterns and differentiate between words such as "effect" and "affect", for example, and was better at complex grammar. But, he said, it fell short in other areas.

"It isn't as accurate for next-word prediction, and it can even 'learn' mistakes and common typos," Mr Furze said.

This is the word used to describe when large language models (LLM) do not have enough underlying data or solid information to generate an accurate response, so they invent a response — often misleading or completely false — and confidently present it as fact.

"These now come into predictive keyboards," Mr Furze said.

"So Apple is trying to fix all this up with the March 2026 iOS 26.4 release, and both Samsung and Google are making their own changes."

Dr Namvar said manufacturers like Apple and Samsung were investing heavily to improve the performance of predictive text systems, which, as well as transformer-based language models, also included increased on-device processing to reduce latency and enhance privacy and expanded support for multiple languages and dialects.

"These companies are also refining personalisation mechanisms, enabling models to adapt to users' writing patterns while providing greater user control over corrections," he said.

Updates recommended

Samsung and Google did not respond to a request for comment.

Apple has recommended customers update their software regularly, with its iOS 26.4 update, released in March, designed to improve keyboard accuracy when typing quickly.

The company said autocorrect learnt from a user's app and contact names, and typing behaviour, to better improve the feature over time.

It said some issues could be caused by unique vocabulary that had become part of the user's personal keyboard dictionary and, if a user consistently made common spelling errors, it could "learn" them.

These would improve over time if the user corrected the errors, and users could also reset their personal dictionary, which would return the keyboard dictionary to its default state and reset all custom words and shortcuts.

Apple said predictive text could write and complete entire sentences with a few taps, but could also be turned off altogether.