Alt text Word wrote itself

How the text is hidden

Word's image describer writes a short noun phrase into the same descr attribute a human would use — a phrase naming a screenshot, a chart or a diagram — sometimes followed by a marker clause it appends itself. The field is invisible either way; the difference is that nobody chose the words.

Why a model still reads it

It still reaches a model through the alt-text attribute, which is why it is reported at all. It is separated because machine-written boilerplate is not authored concealment, and it never enters summary.concealedRetained.

What we do about it

The parser removes the appended marker clause, but only at the end of the value, then requires three things of what is left: the WHOLE value must match the describer opening list, it must be no longer than the DESCRIBER_MAX_CHARS cap taken from measured describer output, and a sentence break followed by more text disqualifies it. A prefix-matching version was implemented first and watched to excuse a payload appended to a describer phrase, a payload dressed as a chart description, and a long human-authored sentence that merely opened like one. The rule fires at informational, action disclose.

How often it fires

2.7% of open-data portal DOCX — 150 real government publications, written in Word (genre: government, NOT contracts), measured 2026-08-18.

This is an alert-volume number and nothing else. It says how often the alarm sounds on documents as found — not how often it is right, and not whether what it found was harmless. Documents as found may themselves carry concealment. Read it against the population named above rather than as a property of documents in general.