Ducking annoying: why has iPhone’s autocorrect function gone haywire?
Since Apple released iOS 26, many iPhone users report bizarre autocorrect swaps such as "come" to "coke". Experts link the glitches to Apple’s new on‑device transformer model, a shift from older n‑gram methods, making errors harder to diagnose. The article explains the technology, its history, and…
Since the September launch of iOS 26, a wave of iPhone users has taken to social media complaining that the built‑in autocorrect is producing absurd suggestions – turning "come" into "coke", "winter" into "w Inter", or even mangling "thumb" into "thjmb". A video of the latter mistake has amassed more than nine million views, fueling speculation that Apple’s latest software update has broken the keyboard.
What changed in iOS 26?
Apple markets iOS 26 as the first version to embed a full‑on‑device machine‑learning language model for text input. In earlier releases, autocorrect relied on statistical n‑gram tables – essentially large lists of word sequences derived from public text corpora. Those tables could predict the next word based on the two or three words that preceded it, and they were simple enough to be inspected and tweaked by engineers.
The new model, described by Apple as a "transformer language model," is a miniature version of the architectures that power ChatGPT, Gemini and other large language models (LLMs). It learns from a user’s typing habits, adapts to personal vocabularies, and runs entirely on the phone without sending data to the cloud.
Why the glitches appear now
Transformers are powerful but opaque. Unlike n‑grams, which can be traced back to specific probability tables, a transformer’s internal weights are not human‑readable. Jan Pedersen, a statistician who contributed early autocorrect research at Microsoft, explains that this lack of transparency makes it difficult to pinpoint why a particular substitution occurs.
Kenneth Church, a computational linguist who helped create the first statistical spell‑checkers in the 1990s, adds that Apple’s “deep, dark secret” approach to its language model means the company does not publicly disclose the training data or the exact architecture. Consequently, when the model makes a mistake, developers have limited tools to debug it.
Apple’s own statement to the press clarified that the viral video showing "thumb" becoming "thjmb" was unrelated to autocorrect; the issue stemmed from a separate keyboard rendering bug. Nevertheless, the timing of the video – a month after iOS 26’s release – reinforced the perception that the new AI‑driven system is unstable.
From spellcheck to AI‑driven autocorrect
Spellchecking dates back to the 1970s, when Unix utilities simply compared each word in a document to a static dictionary and flagged unknown entries. Early autocorrect added a layer of statistical decision‑making: if a user typed "graff", the system would weigh the likelihood of "giraffe" versus "graph" based on edit distance and frequency.
For decades, the dominant technique was n‑gram modeling. It worked well enough that most users never thought about it, except when the system clumsily replaced uncommon names or sanitized profanity in ways that felt "ducking annoying". The shift to transformer‑based models promises better context awareness – the system can now consider an entire sentence rather than just the preceding two words.
However, the trade‑off is interpretability. Pedersen notes that Apple’s on‑device model must be small enough to run on a phone’s processor and memory, yet sophisticated enough to handle nuanced language. When such a compact model misfires, the error can be more glaring than with older methods, because the system is confident in its prediction even when it is wrong.
What users can do now
Apple advises users that the autocorrect engine is continuously updated via background patches, and that the specific video incident was not caused by the language model. In practice, users can mitigate odd corrections by adding frequently mistyped words to the keyboard’s personal dictionary, disabling autocorrect for specific apps, or temporarily reverting to an earlier iOS version if possible.
Developers and researchers are watching the situation closely. The community hopes Apple will eventually release more diagnostic tools or at least a way to opt out of on‑device learning, similar to the privacy toggles offered for Siri and location services.
What may happen next
Given the rapid pace of AI integration across Apple’s software stack, it is likely that future iOS updates will refine the transformer model, improve on‑device training efficiency, and perhaps introduce explainability features for developers. Until then, the current wave of complaints serves as a reminder that even well‑funded tech giants can stumble when replacing decades‑old, well‑understood algorithms with cutting‑edge AI.
For now, iPhone owners should stay alert for official patches, report persistent errors through Apple’s feedback channels, and keep an eye on community forums where workarounds are shared.
Why it matters
Understanding the shift to AI‑driven autocorrect reveals why everyday typing errors have become more noticeable and highlights the challenges of deploying opaque machine‑learning models on consumer devices.
Key points
- iOS 26 introduced an on‑device transformer model for autocorrect, replacing older n‑gram methods
- Transformers improve context awareness but are harder to debug, leading to visible glitches
- Experts warn that Apple’s secrecy around the model makes troubleshooting difficult
- Users can add words to the personal dictionary or disable autocorrect while Apple issues fixes
- Future iOS updates are expected to refine the model and possibly add explainability tools
Frequently asked questions
Is the "thumb" to "thjmb" video actually caused by autocorrect?
Apple says the video captured a separate keyboard rendering bug, not a failure of the autocorrect engine.
Can I turn off the new AI autocorrect on my iPhone?
You can disable autocorrect entirely in Settings > General > Keyboard, but you cannot selectively disable the on‑device language model while keeping basic spellcheck.
Why does the new model make more bizarre mistakes than older versions?
The transformer model makes predictions with higher confidence; when it misinterprets context, the error is more pronounced and harder to trace.
Will Apple release more information about the language model?
Apple has not announced plans to disclose model details, but community pressure may lead to more developer‑focused documentation in future updates.
