Apple’s Hidden Keyboard ‘Grey List’ Exposed: 100,000 Phrases iOS Won’t Suggest

Apple’s iPhone keyboard has long been praised for its predictive text, but an independent researcher has now revealed that the system quietly withholds suggestions for tens of thousands of words and phrases. The discovery, first reported by Reuters and detailed on the website swipeblocked.com, exposes a hidden layer of content moderation operating silently on hundreds of millions of devices.

• Hidden catalogue: Researcher Derik Kauffman identified roughly 100,000 restricted entries across nearly 30 languages, including about 4,513 English phrases.
• Not a ban: Users can still type any phrase manually; the list only suppresses autocomplete and swipe-typing suggestions.
• Political and puzzling: Entries range from ‘dictator Trump’ and ‘Tibetan independence’ to ‘Santa isn’t real’ and ‘iMessage isn’t working’.

How the grey list was uncovered

Kauffman began investigating after noticing that swipe-typing would stall on certain phrases, such as ‘bad boy’ or ‘white man’, while similar terms like ‘bad girl’ and ‘black man’ worked fine. Using OpenAI’s Codex and Anthropic’s Claude Code, he dug into iOS language assets and found a list of cryptographic hashes rather than readable text. By brute-forcing combinations of two to four words against those hashes, he recovered the English list in about ten minutes on an ordinary computer. Security researcher Matt Suiche, an engineer at cybersecurity firm Tolmo, independently reconstructed the same list and cross-referenced it with Apple’s internal data, yielding a near-perfect match.

The scale is striking: the English catalogue has grown from roughly 1,700 phrases in iOS 14 to about 4,513 today, while the total across 28 languages ranges from over 100,000 to around 200,000 entries. The system operates as a ‘grey list’ rather than a hard block—users can still type each letter manually, but the keyboard will never proactively suggest the phrase.

What’s on the list—and what isn’t

Most entries are vulgar, offensive, or related to drugs, sex, and violence. But the list also includes politically charged expressions from across the spectrum: ‘dictator Trump’, ‘sleepy Joe’, ‘right wing conspiracy’, ‘left wing conspiracy’, ‘Tibetan independence’, ‘United Ireland’, and ‘Taiwan is country’. Regional variations appear in other languages: the Hindi list includes ‘Modi braggart’, the Hebrew list includes ‘occupied territories’, and the French list includes ‘immigration too high’. The Russian-language list contains 637 entries, including military-related phrases such as ‘send in troops’ and a reference to ‘Putin drove out’.

The inconsistencies are harder to explain. Kauffman found that ‘bad boy’ is allowed while ‘bad girl’ is not; ‘black man’ autocompletes but ‘white man’ does not; ‘read smut’ can be swipe-typed while ‘read gay smut’ cannot. Seemingly benign phrases like ‘Santa isn’t real’, ‘muffin man’, and ‘iMessage isn’t working’ also appear, suggesting blunt phrase-matching rules rather than a coherent harm-prevention policy. Reuters tested a small subset and confirmed that phrases like ‘Tibetan independence’ and ‘United Ireland’ were difficult to enter via swipe-typing, often resolving to ‘Tibetan independences’ or ‘United Iceland’.

It’s a very light-touch form of censorship. But at the scale that Apple is operating, it only takes a light touch to have a big impact.

Apple declined to comment on Kauffman’s findings or explain why specific political phrases are restricted. The company told Reuters that its keyboard is designed to avoid ‘potentially inappropriate or offensive suggestions’ and that users can always type what they want and override any suggestion. Both points are accurate, but neither accounts for the presence of regional independence terms, product complaints, or gendered asymmetries.

The transparency problem

David Greene, senior counsel at the Electronic Frontier Foundation, said that while part of him found the list ‘creepy’, he could see an argument for curbing unwanted suggestions. ‘But if they are going to do it, this is something users should be really well aware of and have some ability to customise,’ he said. ‘This is something that should not be secret.’

The absence of published criteria or a review process matters because predictive keyboards shape how people communicate. When a phrase requires extra keystrokes, some users will rephrase, abbreviate, or drop the thought entirely. That is a form of influence, whether intended or not. The fact that it took a combination of two AI coding assistants to reverse-engineer the list says much about how opaque Apple’s curation has become. Publishing the filtering criteria, clarifying regional variations, and creating a transparent review process for contested entries would bring the keyboard in line with the scrutiny its scale demands.