Published studies
Printing and printables
-
Study 1
- Digital study complete
- Printer validation not yet run
Printable Ink Efficiency
Measuring Coverage Reduction Across Common Worksheet Rendering Methods
How much estimated printed coverage 36 rendering methods remove from 12 kinds of computer-generated worksheet, and what each one breaks.
Main finding No single ink-saving setting performed best on every page. What a page can save depends on what covers it.
- Removing decoration was the largest lever where decoration exists.
- Page-wide fading saved little on text and line pages.
- Dot patterns never meaningfully beat a flat tint.
-
Study 2
- Digital study complete
- Printer and classroom validation not yet run
Ink Saving by Worksheet Type
Which Print Treatments Keep a Worksheet Working
Study 1's methods applied to 54 real printable test cases, from name tracing sheets to word searches: what each treatment removed from the page, and where lighter printing broke the worksheet.
Main finding Reducing everything equally is the wrong model. The fade that suited outline letters and coloring pages failed every tracing sheet.
- Lightening only the dark example row kept tracing sheets working and about half the saving.
- On practice lists, the word scramble and the crossword, the right result was no ink-saving treatment.
- The kana charts' page layout wastes 1 sheet in 5, more than any ink setting on them could save.
Unicode text tools
-
Study 3
- Benchmark complete: 12 of 13 generators tested
- Pasting into apps not tested
Zalgo Generator Benchmark 2026
Testing Creative Control Across 13 Tools
13 Zalgo and corrupted-text generators run through one procedure: the same ten inputs, fifteen creative tasks and one measurement script.
Main finding No generator led every dimension. Two of the 12 tested let you shape corruption across a sentence, and four could remove Zalgo.
- The three fastest to a copied result have no intensity, placement or shape controls.
- Zalgo.io was the only one with a same-result toggle; PSD-Dude the only one with a separate amount per zone.
- 5 of 12 broke the 👁 emoji when adding marks.
-
Study 5
- Chromium, Firefox and WebKit on Linux complete
- Windows, macOS, iOS and Android not measured
Unicode Rendering Support Study
How Fast Do New Unicode Characters and Emoji Show Up?
Every Unicode character from 8.0 to 18.0 and every emoji sequence, drawn in three browser engines with four font sets, 45 historical emoji fonts and 25 Chrome versions, and the Chromium changes that fixed what fonts could not.
Main finding New characters fail because no installed font has a glyph, not because of the browser. New emoji reach the newest Noto font a median of 55 days after release, new scripts take about two years.
- With Ubuntu 24.04’s own fonts, 0 of 19 Emoji 18.0 sequences drew correctly; with the newest Noto, 19 of 19.
- Two Chrome changes (131 and 142) fixed emoji that no font update could, and we traced each to a Chromium change.
- Our own device check could see 4 of 47,696 missing glyphs; the corrected check sees all of them in Chromium and WebKit, none in Firefox.
-
Study 6
- Comparison complete for Unicode 18.0
- Labels not yet reviewed by people
Emoji Name Drift
How Unicode’s Stable Names Diverge From the Words People Actually Use
The formal Unicode names of 1,400 emoji compared with their CLDR short names and keywords, CLDR’s rename history across 21 releases, and 31 search words from Reddit threads checked against both.
Main finding 68% of the names still match. About 1 in 7 differ in a way that changes how you would describe the emoji, and most of that gap was there from the first CLDR release.
- Keywords keep a dropped Unicode word for 58% of differing names.
- None of the 135 emoji encoded since 2020 differs from its CLDR name.
- The search words neither standard lists are feelings, occasions, picture descriptions and slang.
-
Study 7
- Unicode, Linux font and Markdown tests complete
- Not tested on other systems or in chat apps
Kaomoji Under the Hood
What Kaomoji Are Made Of, and Where They Break in Basic Fonts and Markdown
1,616 faces from our collection pages taken apart by Unicode script and width, checked against three font sets, and run through copy and paste and three Markdown engines.
Main finding Copying never altered a face in our browser tests, but with basic fonts only 61% had a character the fonts could not draw, and standard Markdown changed 4% to 6% of faces.
- 23.6% of faces mix two or more writing systems.
- Every face made only of keyboard characters drew in basic fonts.
- Compatibility normalization (NFKC) would change half the faces and merge 88 into 43.
-
Study 8
- Seven worksheets, three runs, a settings sweep and a re-test after the fix complete
- Not printed on paper or tested on other devices
Name Tracing and Accents
Do Name Tracing Worksheets Keep the Accents? We Tested Seven and Fixed the One That Cut Them
35 names with accents typed into seven name worksheets and eight cursive pages, checking the preview, the saved PDF and the PNG for every cut, lost or moved accent.
Main finding Six worksheets kept every accent. The cursive pages cut the bottom off letters, including the cedilla on Ç and Ş, and 158 of 280 name and page pairs were cut. We fixed the box, and a re-run found none cut.
- 156 of 156 name and worksheet pairs kept every accent on the six print-style worksheets.
- Marks below the line were cut on cursive in 9 of 9 names, and on the print worksheets in 0 of 54.
- Names that needed room now print at 74% to 90% of the old cursive size.
Writing and AI
-
Study 4
- Observational study complete
- Reading experiment not run
AI Writing Footprint Study
What Makes Writing Sound AI Generated? Public Accusations, Writing Patterns and False Alarms
787 public accusations that a text was written by AI, compared with 967 unchallenged texts, with the AI Writing Footprint Checker run on all of it and on writing from before ChatGPT.
Main finding Most accusers named no reason, and no single cue is proof. The patterns we could measure in older writing also appear in comments from before ChatGPT.
- About 57% of accusers named no reason (55% to 66% across three codings); 2.4% named the em dash.
- The em dash character was in 18% of challenged short comments against 2% of unchallenged ones, and made no difference in long articles.
- The checker flagged 52% of unchallenged articles and 21% of books published before 1910.
Planned and not yet run
-
Designed, not run
Physical printer validation for Study 1: a weighing protocol for inkjet, ink-tank, mono laser and color laser printers is designed, not executed.
-
Designed, not run
A reading experiment for Study 4: the same passages with one cue changed, rated by readers. It needs an ethical review route, a consented panel and a licensed set of passages, including writing by non-native English speakers. The protocol is published and has no results.
-
Not yet run
Physical and classroom checks for Study 2: its two holds (the word search and dark dotted tracing) need a real copier, and its tracing result needs teachers, including low-vision readers, to judge a lighter example row.
-
Not yet run
Printer, device and teacher checks for Study 8: a cursive name with a descender or a cedilla printed on a real printer, the cursive worksheet opened on Windows, Mac, iPhone, Android, Firefox and Safari, and a teacher’s view on whether a name at 74% to 90% of the old size still traces well.
Can you help run one of these? Each test needs something we do not have: a printer and a scale, a school or office photocopier, a device we cannot run, or a teacher's eye.
- Offer a printer for the weighing test: inkjet, ink-tank, mono laser or color laser.
- Offer a photocopier for the copy test: the word search and a dark dotted tracing sheet, copied on a real machine.
- Review a tracing sheet as a teacher, including if you teach low-vision readers.
- Print a cursive name sheet and tell us what you see: a name with a descender or a cedilla, such as Gypsy or Çağla, on any printer.
- Open the cursive worksheet on your device and send a screenshot of a name in the preview and in the saved PDF.
Each link opens an email to [email protected] with the test named in the subject. Results you send are published in the next version of that study, credited to you if you want.
Guides built on this research
How these studies are published
- The study page is the canonical record. Its key findings, method and limitations are on the page itself, not only in the PDF.
- The PDF is generated from the study page, so the two carry the same text and version number.
- Every headline finding is tagged with the kind of evidence behind it: measured, simulated, specific to one renderer, not yet tested, or, where a study reaches product questions, an interpretation, a product implication or an unresolved question.
- Data files are published with each study. A corrected study keeps its address, gets a new version number and says what changed.
The evidence labels
- Estimated coverage measured from rendered pages
- Function check measured contrast, visibility or color separation against a stated threshold
- Simulated produced by a model, not a real device
- Chromium-specific a property of one browser build
- Not tested on printers an open question for physical testing
- Physical printer test measured on a real printer; no published study has one yet
- Interpretation the study's reading of its measurements
- Product implication what a result means for a product setting; not a record that anything was built
- Unresolved an open question that needs a printer, a copier or a person