v1.71: the spam filter learns in every language

FormFlow v1.71.0 shipped on 2026-10-19 with two changes that share one theme: text that is not written in English letters and numbers used to be second-class in two places, the spam filter and the submission search. Release v1.71.0 removes that limit in both. This post restates what the Changelog entries say, ties them to the parts of FormFlow you already use, and lists what the entries do not claim.

What the Release changes

There are two entries in v1.71.0, both marked improved.

• The spam filter learns in every language. When you mark submissions as spam or not spam, your form learns from them in any language: Arabic, Russian, Greek, Chinese, Japanese, Korean, Thai and the rest, not just English. Spam that swaps in look-alike letters to dodge the filter is caught too.

• Search finds submissions written in any language. Searching your submissions now finds words in Arabic, Cyrillic, Chinese, Japanese, Korean, Thai and every other script, not only English letters and numbers.

Nothing in either entry asks you to change a form, a setting or a habit. They describe behaviour that now works in more scripts, not a new control.

How the filter learns, and why marks matter

In FormFlow's vocabulary, marking a submission as spam or not spam is a Correction: a human's explicit judgment, made from the inbox list or from the submission's own page. A Correction is the only input allowed to train the Spam model.

The Spam model is per form. Each form builds its own picture of what spam looks like for that form, trained exclusively from the Corrections you make on it. When the model is confident, its verdict surfaces as a learned Spam signal, and signals combine into a submission's spam score. For words the form has not seen yet, the model backs off to a Team prior, so a new form can flag from its first submission.

The practical reading of the v1.71.0 entry follows from that. A form whose inbox is mostly Arabic, Russian or Chinese gets the same benefit from its Corrections as an English-language form does: what you mark is now learned from, whatever script it is written in. Because training is human-driven, the filter is only as good as the marks you give it. A Team that has made no Corrections has nothing to learn from, in any language.

Look-alike letters

The entry also says that spam which swaps in look-alike letters to dodge the filter is caught. That describes the common trick of writing a word with a character from another alphabet that looks the same on screen. The entry states the outcome, not the mechanism or the list of substitutions covered. Read it as a real improvement against that trick, not as a guarantee against every disguise.

Search in any script

Submission search is the other half. Searching your submissions now finds words in Arabic, Cyrillic, Chinese, Japanese, Korean, Thai and every other script. Before this Release, the entry's own wording makes the limit plain: search covered English letters and numbers only. If you collect leads in more than one language, one search box now works across all of them.

The entry says search finds words. It does not describe ranking, fuzzy matching, spelling variants or how a query in one script behaves against text in another, so do not assume any of those.

Where it is blunt

The entries are short, and so is what can honestly be said about them. Specifically:

• No accuracy figures. The entry does not say how many Corrections a form needs before the learned signal becomes confident, in any language.

• The language list is not exhaustive. "And the rest" and "every other script" are the entry's words; the named scripts are examples.

• The filter still depends on you. It learns from your marks. It does not arrive pre-trained for your language, and a Correction is still the only training input.

• This post does not say what happens to marks you made before v1.71.0, or whether earlier submissions are re-scored. The entries do not say.

What this does not decide

v1.71.0 does not change who decides what is spam. A learned verdict is one signal that feeds a spam score, and your Corrections remain the source of truth. It does not add a setting, and it does not change how submissions are counted or delivered. Later patch Releases in the v1.71 line are separate Releases with their own Changelog entries; this post covers v1.71.0 only. The Changelog entry stays the canonical record of the Release. This post is commentary on it.