Predictive Text Is Getting Worse with AI

By Malcolm Sutton

891 ABC Adelaide

Topic: AI

Thu 11 Jun 2026 at 6:21am

Some users say predictive text is not as good as it used to be since AI was added.

The next stop. Eating a banana. It's no longer shiny sync.

My phone gave me weird words when I was texting 'bus', 'banana', and 'no longer sensible' the other day.

Many people online say that autocorrect and predictive text have gotten worse in recent years.

There are many complaints on Samsung and Apple forums about their predictive text and autocorrect.

These complaints include autocorrecting a correct word to a wrong word, repeating mistakes, and inserting nonsense words.

Smartphone users have reported predictive text issues in recent years.

Others include reducing two words to contractions that are not relevant and changing words like 'good' to 'food'.

Forum contributors blame operating system updates and say that once useful software now has many errors since AI was introduced in 2023.

A 'demonstrable decline'

PhD candidate Leon Furze studies how AI affects writing instruction and education.

The age of AI is here, but it can be boring and not helpful for some people.

There's been a clear decline in old-fashioned predictive text, and companies have tried to fix it.

In 2023, Apple started using a machine-learning language model to improve its predictive keyboard.

This replaced an old model that had been used since around 2007.

Apple added more AI to its mix in later iOS releases.

Samsung did something similar from 2023 onwards.

Anecdotal evidence

Morteza Namvar is a senior lecturer at the University of Queensland Business School.

He says there's a lot of talk online about predictive text and autocorrect getting worse.

It's hard to say if it's really gotten worse, but users do report more unexpected corrections and suggestions.

Old predictive text systems used simple statistical models and were generally more predictable.

New systems use neural language models that try to understand context and meaning, but can make more noticeable errors.

From a scientific perspective, this isn't necessarily a decline, but a trade-off for more flexibility and understanding.

Modern systems may be more sophisticated, but their errors can be more noticeable.

An unpredictable history

Predictive text was first used by mobile phone companies in the late 1990s with T9 software.

A text message that autocorrected 'local hall' to 'local happy hour'.

T9 associated several letters with each key and predicted words based on the order of key presses.

Push-button QWERTY keyboards came out, but T9 remained popular until smartphones took over in the late 2000s.

Algorithms used statistical n-grams to predict the next word based on previous input.

Predictive text could predict the next word and even phrases.

But when phone companies introduced AI in 2023, things changed.

Phones 'learning' mistakes

Transformer-based AI can learn user patterns and differentiate between words, but it can also learn mistakes.

It's not as accurate for next-word prediction and can 'learn' common typos.

'AI slop' refers to low-quality content created by generative AI with errors.

This happens when large language models don't have enough data and invent responses that are misleading or false.

These errors now appear in predictive keyboards.

Apple is trying to fix this with its iOS 26.4 release, and Samsung and Google are making changes.

Manufacturers are investing in improving predictive text systems with transformer-based language models and on-device processing.

They're also refining personalisation mechanisms to adapt to users' writing patterns.

Updates recommended

Samsung and Google didn't respond to a request for comment.

Apple recommends updating software regularly, with its iOS 26.4 update improving keyboard accuracy.

Autocorrect learns from a user's app and contact names, and typing behaviour, to improve over time.

Issues can be caused by unique vocabulary in a user's personal keyboard dictionary, and if a user makes common spelling errors, it can 'learn' them.

These issues improve if the user corrects the errors, and users can reset their personal dictionary.

Predictive text can write and complete sentences, but can also be turned off.

ABC Adelaide in your inbox

Thu 11 Jun 2026 at 6:21am

Promotion

Praticar Escuta