Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)

Swipeblocked reveals 4513 English phrases the iPhone keyboard won't suggest

A suggestion filter can stop people from writing what they intended. Swipeblocked's creator reconstructed lists of phrases from Apple's language files that the keyboard excludes from suggestions and, in many cases, swipe typing. Sometimes it even changes the previous word to do so. There is a useful lesson here for developers: test not only what the filter removes, but also what happens to the user's text afterward.

AI-processed from TG: @oulenspiegel_channel (Сергей Марков (ML & AI)); edited by Hamidun News
Swipeblocked reveals 4513 English phrases the iPhone keyboard won't suggest
Source: TG: @oulenspiegel_channel (Сергей Марков (ML & AI)). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @oulenspiegel_channel (Сергей Марков (ML & AI)) ↗
Editor-in-Chief
Key metric #1The website publishes 4513 English phrases from the reconstructed list.
Manual typing remains availablethe restrictions apply to keyboard assistance.
Key metric #3A phrase's presence in the table does not explain why Apple added it.

A user slides a finger across the keyboard, expecting a familiar word, and gets a strange replacement. The first thought: recognition has missed the mark again. But the cause may be a filter that checks combinations of words and removes an unwanted continuation. If the system changes text already entered to achieve this, a protective feature becomes a usability problem.

The Swipeblocked project offers a look at these restrictions on the iPhone keyboard. Its creator, Derik Kauffman, reconstructed lists from Apple's language files and published them alongside an explanation of his method. The website makes 4513 English phrases available. For teams building autocomplete, search or AI editors, this is a reason to test their own products: what happens to the text after a filter kicks in?

When suggestions start editing the user's text

The project page shows an example: after the words left wing, someone swipe-types conspiracy. The keyboard replaces the preceding wing with wrong, avoiding a combination on the list. The user gets a change in a place where they may already have finished typing and stopped watching the result.

According to the creator's description, the restrictions affect next-word prediction and swipe typing, with different coverage for each. Typing a phrase letter by letter remains possible. This is an important boundary: the system restricts its own suggestions and recognition candidates while preserving manual input.

Imagine a similar mechanism in an editor for customer correspondence. An employee writes a complaint and quotes a customer's offensive remark. The filter spots an unwanted combination, the editor substitutes another word, and the quotation loses its accuracy. On the surface, the interface keeps working: text appears and there are no errors. But the user's task has already been completed incorrectly. These are precisely the cases that are easy to miss if protection is assessed only by the number of expressions removed.

What could be extracted from the files

According to Kauffman's technical breakdown, the tables contain numerical fingerprints of word sequences. He tried possible combinations and checked for matches. The word “decoded” therefore needs a qualification here: the researcher did not obtain an original document explaining Apple's decisions.

The method has limitations. Fingerprint matches require additional checks, and the reconstruction of some language lists remains incomplete. The creator also tested behavior on devices and in the simulator for selected phrases. It cannot automatically be assumed that every published row will behave identically on every iPhone owner's screen.

The original Russian-language post draws attention to absurd combinations and inconsistencies in the restrictions. But an odd entry does not reveal how it came to be there. It could have resulted from a template, a characteristic of the language data or an individual editorial decision. Without internal company documents, an explanation involving corporate processes remains an interpretation.

For practical purposes, something else matters more: the list shows specific outputs of the mechanism, but does not provide a ready-made assessment of its quality. Assessing quality requires knowing which user tasks it preserves, which it breaks and how often that happens.

What to test in your own product

If your system suggests or rewrites text, start with a small set of scenarios. Take an ordinary complaint, a quotation, a negation, an organization's name and several paraphrases of the same message. Check whether the meaning survives filtering and whether people can get the result they need without fighting the interface.

Test previously confirmed input separately. Changing a preceding word deserves more attention than a disappearing suggestion: the user may have considered that section finished. For a product manager, this is a matter of trust; for a developer, it is a reason to explicitly define the boundaries of automatic editing and make interventions visible.

It also helps to keep measurements separate. How many unwanted suggestions the system hid is one metric. How many legitimate tasks it spoiled is another. An improvement in the first says nothing by itself about the second. You need examples of errors, comparisons between versions and a way to identify which rule triggered a replacement.

Swipeblocked is compelling precisely as material for this kind of analysis. It helps reveal how a compact technical restriction affects everyday typing. The next step for your own team is concrete: run through critical user scenarios with the filter enabled and read the resulting text in full.

Story

Key Questions Answered

When suggestions start editing the user's text?

The project page shows an example: after the words left wing, someone swipe-types conspiracy. The keyboard replaces the preceding wing with wrong, avoiding a combination on the list. The user gets a change in a place where they may already have finished typing and stopped watching the result. According to the creator's description, the restrictions affect next-word prediction and swipe typing, with different coverage for each. Typing a phrase letter by letter remains possible.

What could be extracted from the files?

According to Kauffman's technical breakdown, the tables contain numerical fingerprints of word sequences. He tried possible combinations and checked for matches. The word “decoded” therefore needs a qualification here: the researcher did not obtain an original document explaining Apple's decisions. The method has limitations. Fingerprint matches require additional checks, and the reconstruction of some language lists remains incomplete.

What to test in your own product?

If your system suggests or rewrites text, start with a small set of scenarios. Take an ordinary complaint, a quotation, a negation, an organization's name and several paraphrases of the same message. Check whether the meaning survives filtering and whether people can get the result they need without fighting the interface. Test previously confirmed input separately. Changing a preceding word deserves more attention than a disappearing suggestion: the user may have considered that section finished.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…