A filled PDF form field holds two things: the typed answer, and a small drawing of it, which is what the box shows. When the drawing fails, the answer is usually still in the file. On 20 November 2026 we filled a test form in English, Malay, simplified and traditional Chinese, Thai and Vietnamese, saved it with pdf-lib's built-in Helvetica and a flag asking viewers to draw the boxes, and opened the file in four viewers on one Mac. pdf.js drew every answer. Chrome's viewer dropped two Chinese characters and four Vietnamese letters. poppler left the Chinese and Thai boxes empty; macOS Quick Look left the Vietnamese one empty too. Yet all seven stored answers read back exactly as typed.
Every form field carries a typed answer and a separate drawing of it
A field's answer sits in its value entry, /V. The box on the page is a widget with its own appearance stream, /AP: instructions for painting the glyphs. Software that changes the value but cannot draw it must still leave the box something, and that decides what a reader sees.
Our form was one 420 by 400 point page with six text fields and a dropdown, every text box already holding "OLD VALUE" as if an earlier answer were being corrected. The new answers, all fictional, were Tan Ah Kow, Nur Aisyah binti Abdullah, 陈大文,新加坡, 陳大文,臺北市, สมชาย น้ำใจ, Nguyễn Thị Hương and, in the dropdown, 新加坡. The saved file was 8,948 bytes, and pdf-lib 1.17.1 and pdf.js 4.10.38 read all seven values back exactly.
Four viewers drew the same saved file differently
| Field | pdf.js 4.10.38 | Chrome 149 viewer | poppler 26.10.0 | Quick Look |
|---|---|---|---|---|
| Simplified Chinese | Drawn | Drawn | Empty | Empty |
| Traditional Chinese | Drawn | 陳, 臺 missing | Empty | Empty |
| Thai | Drawn | Drawn | Empty | Empty |
| Vietnamese | Drawn | "Nguyn Th Hng" | "Nguyn Th Hng" | Empty |
| Dropdown | Drawn | Drawn | Empty | Empty |
pdf.js ran in a canvas inside Chrome for Testing 149.0.7827.55; Firefox's viewer is built on it, but we did not test Firefox. poppler means pdftoppm; Quick Look is the macOS 26.7.2 thumbnailer, a different program from Preview. Quick Look's boxes were pixel-identical to the blank form's, border included. Font fallback depends on installed fonts, so these are our Mac's results.
An old answer can haunt the box after the value underneath has changed
Worse than an empty box is a plausible, false one. A save that sets the form's NeedAppearances flag, skips redrawing and leaves the previous appearance stream in place hands any viewer that trusts /AP the old picture. Our harness did this with plain pdf-lib calls, and Chrome's viewer and Quick Look showed "OLD VALUE" over a field we had cleared.
pdf.js pull request 15587, merged in November 2022, says that when a character is missing on save it sets NeedAppearances "and remove[s] the appearance stream". The PDF Association's Arlington model of PDF 2.0 (ISO 32000-2:2020) lists that flag from PDF 1.2 and as deprecated in 2.0, where it is still defined.
Until 20 November 2026 our PDF form filler did exactly this: on this test form, Quick Look showed "OLD VALUE" in 4 of 4 boxes changed to Chinese, Thai or Vietnamese. The revision of that date gives each such box an empty drawing and keeps its border; only a list box loses its drawing outright, as pdf.js does.
Helvetica's WinAnsi set leaves out Vietnamese tone letters and keeps Rumi Malay
The pdf-lib README says its standard fonts use WinAnsi encoding, which "only supports 218 characters in the Latin alphabet". We passed every code point of Unicode's Basic Multilingual Plane through pdf-lib 1.17.1's Helvetica encoder and got 218: printable ASCII, all of U+00A0 to U+00FF, and 27 others such as € and curly quotes.
Script is a poor guide to what fails. The Rumi Malay name lost none of its 25 code points; the Vietnamese name lost 4 of 16: ễ, ị, ư and ơ. A Malay string in Jawi lost 8 of 9; every Chinese and Thai letter fell outside the set. Our guide to languages that come out as empty boxes traced Vietnamese failing in a web font; in a PDF the limit is WinAnsi, and tone-marked vowels fail beside the hooked ones.
Flattening has to draw every answer, so a Chinese one stops it
Flattening paints each answer into the page and deletes the fields, so no viewer is left to redraw anything. Helvetica cannot paint Chinese. Plain pdf-lib's save() throws WinAnsi cannot encode "陈" (0x9648). The filler we tested refuses before writing and names the five fields that stopped it.
Subsetting a Chinese font quietly lost most of the answer
Embedding fonts fixes the drawing. With our Noto Sans SC 2.004, cut to 9,783 code points for this site, and Noto Sans Thai 2.002 embedded whole, all four viewers drew the Chinese, Thai and dropdown answers, in a file of 2,030,597 bytes. Subset, the file came to 14,542 bytes, and only 陈 and the full-width comma survived. By ink in poppler, the simplified answer kept 20% and the traditional one 1% of what the full font drew, and every viewer showed the same loss.
The cause was reported before us. In pdf-lib issue 1232, brucx wrote on 23 July 2026 that the subsetter saves the glyph location table in short form, which, per Microsoft's 'loca' specification, stores each offset divided by 2, so an odd offset cannot survive. They pointed to the upstream fix, fontkit pull request 191, released in fontkit 1.8.0 in 2019; pdf-lib's fork, @pdf-lib/fontkit 1.1.1, predates it. brucx measured 29% of the full font's ink on a static Noto Sans SC instance; our subset, with 4,911 of 9,778 glyph records at odd lengths, is a cross-check. The Thai font had none and subset cleanly, to 6,239 bytes on our 3,968-byte blank form.
Thai tone marks lose their lift when the glyph offsets are dropped
fontkit lays out สมชาย น้ำใจ as 12 glyphs for 11 characters, because ำ is drawn as two, and raises the tone mark in น้ำ by 309 font units. pdf-lib 1.17.1 writes the glyph numbers and ignores the offsets, so at 400 dpi in poppler the tone mark sits on the nikhahit, the small circle that begins ำ.
Copying the text back out gains a letter. pdftotext and pdf.js both returned 12 code points with an extra า, because pdf-lib's character map labels the nikhahit glyph as the whole ำ and the second glyph as า. Our guide on CJK and Thai text that copies blank or garbled found the reverse in shaped page text, where ำ went missing; both follow from one character drawn as two glyphs.
The stored value is what a machine reads, whatever the box shows
Form-reading software reads /V, which in every file we saved held the exact string typed: 陳大文,臺北市 with both characters Chrome's viewer dropped, and Nguyễn Thị Hương with all four letters poppler lost. A person reads the drawing. The two are written separately, and one can be empty, stale or short of letters while the other stays exact. Nothing in the file records which program will open it.