A researcher has discovered an undisclosed multilingual list containing tens of thousands of words and phrases that Apple’s iOS keyboard avoids suggesting through autocomplete and swipe-typing, Reuters reported.Derek Kauffman, the researcher who uncovered the list, described it as a “very light-touch form of censorship.” However, he cautioned that given the scale of Apple’s operations, “it only takes a light touch to have a big impact.”The list does not constitute a formal ban. iPhone users can still manually type any words or phrases in their emails and messages. The restriction simply means that, in most cases, the keyboard will not suggest these terms through predictive text or swipe-to-type features.
Which words and phrases are on the list
Most of the flagged words and phrases are vulgar, offensive or related to drugs, sex and violence. However, the list also includes politically sensitive terms such as “dictator Trump” and “sleepy Joe,” referring to US President Donald Trump and his predecessor, Joe Biden. Trump has repeatedly mocked Biden as “sleepy Joe.”Other phrases excluded from iOS’s predictive text features include “right wing conspiracy,” “Tibetan independence” and “United Ireland.” The list also has an India connection.Different languages have their own set of flagged words and phrases. In Hindi, for instance, the keyboard does not allow users to swipe-type references to Prime Minister Narendra Modi and Rahul Gandhi, the Lok Sabha leader of the opposition.Similarly, French speakers cannot swipe-type the phrase “immigration too high,” while Hebrew speakers face restrictions on the phrase “occupied territories.”“Every major language has some version of this,” Kauffman told Reuters. While acknowledging that Apple may be trying to prevent users from accidentally sending obscene messages or inflammatory slogans, he highlighted what he described as “inconsistencies” in how the company applies the feature.As an example, he pointed to the word “smut,” which means a sexually explicit material. He said users could swipe-type “read smut” but not “read gay smut.”
How the list was discovered
Kauffman discovered the list in September after noticing that swipe-typing — gliding a finger across the keyboard to enter words — would stutter over certain phrases, such as “bad boy” and “white man,” but not “bad girl” or “black man.”Curious about the pattern, he used OpenAI’s Codex and Anthropic’s Claude Code to investigate, eventually uncovering a list stored in a location he described as “deep in the bowels” of Apple’s operating system. While the count varies depending on how word combinations and their permutations are calculated, Kauffman estimated that there are around 100,000 entries, excluding Chinese characters.Reuters independently corroborated his findings, with phrases such as “United Ireland” and “Sleepy Joe” failing swipe-typing tests.Separately, Matt Suiche, an engineer at agentic AI cybersecurity company Tolmo, recovered the same list and cross-checked Kauffman’s findings against Apple’s internal data, obtaining a near-perfect match.Suiche noted that the list had been hidden in plain sight for years, offering an insight into how “human communication has been potentially influenced and filtered for years inside messaging apps.”
Why the list raises transparency concerns
Users have questioned iOS’s predictive text restrictions for years, with several taking to social media to highlight phrases they could not swipe-type. Earlier this week, a Norwegian systems administrator independently uncovered a list of flagged words, including “al-Qaeda,” “Auschwitz” and “Hooters.”However, the existence of a much broader list of sensitive phrases is a first.Apple is not the only tech company to face scrutiny over predictive text. Alphabet-owned Google has repeatedly faced criticism, and even lawsuits, over controversial autocomplete suggestions.Apple, too, drew attention last year when users reported that dictating the word “racist” on their iPhones briefly caused the word “Trump” to appear on screen. The company attributed the issue to an error and said it had been fixed.
