Rest assured, you’re not experiencing madness.
Your iPhone’s AutoCorrect has been unusually erratic, unexpectedly altering words like: “Come” to “Cola” and “Winter” to “W Inter”. If you’ve encountered this, you’re not on your own.
Many online comments reveal that others share your frustrations. Hundreds of internet users express concern, fearing this issue might persist indefinitely.
Following the release of its latest operating system, iOS 26, in September, conspiracy theories began to emerge. A video showing a user’s iPhone keyboard changing “thumb” to “thjmb” has gained over 9 million views.
“Autocorrect manifests in various forms,” states Jan Pedersen, a statistician known for his pioneering work on autocorrect at Microsoft. “It’s somewhat challenging to identify the technology behind user predictions, as it operates beneath the surface.”
An early pioneer of autocorrect suggested that those seeking explanations may remain in the dark concerning this recent change, primarily due to Apple’s approach.
Kenneth Church, a computational linguist who developed foundational autocorrect techniques in the 1990s, remarked, “Apple’s operations have consistently been shrouded in secrecy. The company excels at maintaining confidentiality over most others.”
For several years, the internet has been buzzing about autocorrect issues, long before the launch of iOS 26. Notably, there is at least one significant distinction between today’s autocorrect and its earlier versions: the incorporation of artificial intelligence, or what Apple refers to as such. The introduction of iOS 17 brought “on-device machine learning language models” designed to adapt based on user input. However, this can encompass a variety of interpretations.
In response to inquiries from The Guardian, Apple stated that it has continually updated AutoCorrect using the latest technological advancements, asserting that the keyboard complications showcased in the video aren’t linked to autocorrect.
Autocorrect has evolved from earlier spell-checking technology, which originated in the 1970s. This initial spell-checking featured a primitive command in Unix that identified all misspelled words within a text file. It simply compared each word against a dictionary and alerted users to any discrepancies.
“One of my initial tasks at Bell Labs was to obtain the rights to the British dictionary,” Church recalls. He utilized these dictionaries during his early research into autocorrect and speech synthesis programs.
The task of autocorrecting words—such as suggesting “them” instead of “they” in real time—is far more complex. It involves mathematical calculations, wherein the computer statistically evaluates whether “graph” is more likely to refer to a giraffe (just a couple of letters apart) or a homophone like “graph.”
In more complex scenarios, autocorrect must discern if the actual words used align with the context. For instance, figuring out whether your teenage son excels at “math” rather than “meth.”
Until recently, cutting-edge technology relied on N-grams. This system was sufficiently effective that most users took it for granted. If I suspected that a unique name might not be recognized, I would replace any expletives with bland alternatives (a tactic that could be frustrating), or introduce randomness. This led to amusing text changes, like altering “I delivered the baby in a taxi” to “I devoured the baby in a taxi.”
After newsletter promotion
In simple terms, an N-gram serves as a rudimentary version of contemporary LLMs like ChatGPT. They statistically anticipate what you’re inclined to say based on your previous words and common sentence completions. Various engineering approaches affect the data utilized by N-gram autocorrect, according to Church.
Yet, they are no longer at the forefront of technology. We have entered the AI era.
Apple’s innovative Transformer Language Model signifies a more advanced technology than conventional autocorrect, as Pedersen explained. Transformers represent significant progress behind models such as ChatGPT and Gemini, making these models more adept at handling human inquiries.
The implications for the new autocorrect remain ambiguous. Pedersen notes that whatever Apple incorporates will likely be significantly smaller than widely-recognized AI models, making it feasible for mobile devices.
However, grasping what is malfunctioning with the new autocorrect may prove more challenging than with prior models due to the inherent difficulties of interpreting AI.
“A vast domain of explainability and interpretability exists, and people desire clarity regarding how mechanisms operate,” Church stated. “Old methods can still yield insights into actual operations. The latest innovations appear somewhat magical—they outperform older systems, yet when they don’t function correctly, the results can be dire.”
Source: www.theguardian.com
