An enthusiast used AI to extract the iPhone’s predictive-text blocklist: 673 Russian phrases

When someone types something offensive on an iPhone themselves, Apple gets no complaints. But if a system suggestion supplies the rest of the phrase, responsibility becomes less clear-cut. The user might simply have tapped a suggested word by accident, leaving someone to explain how an obscene or extremist statement ended up on the screen.

Why Apple has a blocklist

That is why the iPhone contains a list of things that should never appear as a phrase completion or a swipe-typing result. Alexander Gorny, author of the “Startup of the Day” channel, calls it protection rather than censorship and has no objection to the idea itself. From a product perspective, the logic is understandable: if a feature generates text on its own, whoever released it is responsible for the output. Better to err on the side of caution than have to explain why a phone completed a user’s message with “heil.”

What the enthusiast found

Things got interesting when a vibe coder used AI to extract the blocked phrases from the phone and present them on a landing page called Swipeblocked. Anyone can now open the list and read it.

For Russian, the figures are 673 phrases in 242 groups. Some entries are entirely predictable; others are strange:

• “what Bolotnaya” — a reference to events almost fifteen years ago; it accounts for about half a percent of all restrictions;
• “shoot eight colleagues” — specifically eight. Neither “seven” nor “nine” appears on the list, and neither do friends, acquaintances or university classmates. The author looked for an explanation and found none: there had indeed once been a news story about eight murders, but there had also been one about seven, just in a different year;
• “stupid Americans” — the largest group excluding profanity, with no fewer than 24 spelling variants;
• “disgraced oligarch,” “lazy journalist” and the unexpected “the result is obvious.”

The author estimates that roughly five percent of the sections relate to politics.

What this means in practice

The main lesson extends beyond Apple to any product with suggestions, autocomplete or text generation. As soon as your interface starts suggesting words to users, you become a co-author of their messages, and a protective layer becomes an essential part of the product.

A second observation: lists like these are almost always assembled reactively. An incident happens, a news story appears, and someone in legal or PR asks for a phrase to be added. That explains isolated oddities such as “eight” appearing without “seven”: the rule addresses a specific case rather than a class of problems. Such a list is hard to maintain, generalizes poorly and produces false positives such as “the result is obvious.”

Third, blocklists themselves become a source of reputational risk. All it takes is for someone to extract and publish one, and the company ends up with a collection of curiosities that people use to judge its internal workings. The post’s author concludes that this perfectly illustrates the absurdities produced by a large company’s business processes.

Limitations

Remember that everything we know about the list comes from the post and the website itself. We have not independently verified whether the extraction is complete or current, or precisely how Apple uses these phrases. The Russian-language figures are the author’s own count, and the reasons for including particular phrases remain speculative.

Takeaway for teams

If you build text suggestions or text generation, assign an owner to the blocklist and review it regularly. Record the reason and date for every entry. Group variants together instead of maintaining a scattering of individual phrases. And assume from the outset that, sooner or later, someone will read your list.