How dub.mom works, and what it’s built on

Everything in the program comes from published reading research or from our own measurements. This page shows which is which. Last updated: 11 September 2026.

We wrote the main text for readers whose English is a second language. Every section opens into sources and details: numbers, sample sizes, what each study got wrong. Every claim on the landing page links to the part of this page that backs it.

The one idea

Reading faster does not come from moving your eyes faster. Eye movement is a small part of reading time. Most of the time goes into recognizing words and putting them together into meaning. [1][2]

So there are only a few ways to read faster that also keep understanding:

  • Make the language easier to process: know the words, read in phrases, know the shape of the text.
  • Decide what not to read.
  • Practice at the edge of your speed, with a check that you understood.
  • Slow down for the parts that will hurt you if you get them wrong.

That is the whole program. Every part below is one of those four levers.

Sources and details

Rayner and colleagues' 2016 review is the standard summary. It finds a real trade-off between speed and understanding, and concludes that the way to gain speed without losing comprehension is practice and better language skill, especially vocabulary [2]. Removing the ability to re-read makes comprehension worse [1].

What AI English is, and why it’s harder for you

Text from AI tools is not ordinary English. Measured against human writing, it has longer sentences with more clauses, rarer and more formal words, and more padding: it restates the question, adds headings, and ends with caveats. One study put ChatGPT text at a reading grade of about 16.6, against 12.1 for essays by non-native students [21]. And there is a lot of it: in a study of 82,845 conversations between developers and an AI tool, the AI’s answers were about fourteen times longer than the questions, counted in tokens [24].

Two more things make it harder than it looks. AI text sounds confident whether it is right or wrong, so mistakes don’t look like mistakes. And people rate AI-written summaries as clear even when they understand them less well than human ones [25]. Feeling fluent is not the same as understanding.

Sources and details

[21] compared 50 student essays with 50 ChatGPT texts on readability and syntactic complexity; it studied essays, not workplace documents, so we treat the exact grade levels as indicative. [24] is cited at v1 (September 2025), which analysed 82,845 developer–AI conversations and reports the ratio as a median of token lengths. The revised v2 covers 587,568 conversations but no longer contains any response-versus-prompt length finding, so the larger sample cannot be attached to this number. [25] tested 150 people on plain-language summaries. On the “AI overuses these words” point, a 2025 study of 15 million abstracts found 379 words whose frequency jumped once AI writing tools arrived [22]; we use that list, but whether learning it speeds reading is untested — see the third column of the evidence table.

How the program works

Ten minutes a day, one text at a time.

  1. New textDay 1
  2. Read againDay 2, audio
  3. New textDay 3
  4. Read againDay 4, audio
  5. CheckpointDay 5, no help

A point on your chart. Read-again days never put one there.

New-text days
You read one new text of the kind you meet at work (a ticket summary, a call recap, a report), then answer ten questions with the text hidden. Seven right is a pass; if you pass, your speed goes on your chart. If not, we slow down. This is the protocol that raised second-language reading speed by about half in nine weeks, in studies from Korea, Taiwan and Vietnam [3–7]. Most of the gain came in the first ten texts [3].
Read-again days
You read yesterday’s text twice: once along with audio a little faster than your own pace, once on your own. This is repeated reading with audio, the only method in this group that also improved comprehension against a control group [8], and one whose gains were still there three months later [9].
Checkpoints
Every fifth session, no help at all: plain text, no glosses, no shading. This is how we measure. The texts are matched for difficulty so a checkpoint is comparable with the last one and with your benchmark [7].
Help that fades
While you need it, the text is broken into phrases, words can be tapped for a translation into your language, and padding is lightened. Each kind of help goes away when your own numbers say you don’t need it, and comes back if they say you do. Phrase-broken layout has improved comprehension in classrooms for more than fifteen years [12–14]. Translations in your first language work better than explanations in English [17]. Lightening the padding helped readers skim faster with better understanding in two small studies [15, 16]. And people learn more with a medium amount of help than with a lot, even though they prefer a lot [18], which is why the fading is not optional. That study measured note-taking on lecture video, not reading; carrying it across to reading help is our inference, not its finding.
Reading on your own
Easy texts with no questions, and your own AI texts from work, with the same tools under your control. Volume of easy reading has a solid record on its own [10].
Catching mistakes
From Day 5, some texts contain a planted error. You mark what you think is wrong; we tell you afterwards. In one small study, people caught a wrong AI conclusion only a quarter to a half of the time; they did best when the answer gave them enough reasoning to check the conclusion against [23]. We measure your catch rate from the start and add practice once we have a baseline for you.
Sources and details

The five L2 speed-reading studies [3–7] are the core. Their weaknesses: the earliest had no control group and did not report comprehension [3]; samples were 29 to 84 students; and part of the measured gain fades after the course ends [4], which is why dub.mom has checkpoints on matched texts, a four-week check, and a maintenance mode. The “first ten texts” finding is from [3] and should be read as “in that study.”

What we measure

Effective reading speed
How fast you read a new text when you also understood it: at least seven of ten questions right. Readings below that do not count. Speed and understanding are never averaged together; understanding is the floor, not a factor.
Catch rate
Planted errors you noticed, out of the ones you have seen, with false alarms counted next to it.
With help and without
Your regular speed and your checkpoint speed, and the gap between them. The gap closing is the program working.
Transfer
At checkpoints 10, 20 and 30, a different kind of text (another genre, a human-written memo, a long text) to see whether the skill leaves the training texts. Macalister’s 2010 study is the closest precedent: a comparison group, and transfer to authentic texts [33].
Four weeks later
One more checkpoint after the program, to see how much stayed.

What we deliberately keep apart: speed on a text you read yesterday (inflated by repetition) and speed on your own reading with no questions (unchecked). Both are shown, both are labelled, neither goes on your main chart.

Sources and details

Our 7-of-10 gate is the lower half of a two-sided rule. Nation sets a band, not a floor [6]: seven or eight out of ten is the target, and a higher score means the reader is going too slowly and taking too much from the text, a signal to speed up, not a better pass. Below seven he has the reader continue at the same speed until comprehension catches up, where we ease the next text instead. Both departures are ours, not his, and both are deliberate. The 250 wpm mark is his, given as a reasonable goal for second-language readers on material with no unknown vocabulary or grammar [6]. Speeds are normalised by characters so texts of different genres compare [26]. The pre/post matching of checkpoint texts follows the correction Gorsuch and Taguchi made to their own earlier work [8].

What’s tested, what’s mixed, what we’re testing

Well tested

Replicated, or meta-analysed

Speed comes from language skill and practice, not eye movement1, 2
Timed reading with a comprehension check: about +50% in nine weeks3–7
Repeated reading with audio; gains held at three months8, 9
Volume of easy reading; adults benefit most10, 11
First-language glosses beat English glosses17

Mixed or small

Classroom, small N, or one study

Phrase-broken layout (one study found plain spacing did as well)12–14
Lightening the padding (N=18 and N=24)15, 16
Medium help beats maximal help, measured on note-taking, not reading18
Individual font choice and speed20
Priming before reading (N=44, preliminary)19

We’re testing

Nobody has studied this yet

That any of this transfers to AI-generated textour own study
Catching planted errors as a trainable skill23
Drilling the words and bundles AI overuses22
Fading help on a learner’s own numbers—
Four-week retention after ten minutes a day4

The right-hand column is the honest part. Nobody has studied training people to read AI-generated text. Everything in the program is borrowed from research on ordinary reading, and the borrowing is the bet. Our own study says how we’re testing it.

What we don’t do, and why

No flashing one word at a time.
Rapid serial presentation keeps comprehension only at ordinary speeds; above about 350 words per minute, understanding drops, and it forbids re-reading, which good readers do [1, 27, 28]. For a reader in a second language it is worse, not better.
No bold half-words.
“Bionic” formatting made 2,074 readers slightly slower in one company’s experiment and made no difference in two controlled studies [29–31].
No streaks, no badges.
They push sessions, not skill, and the number that matters is on your chart.
No self-report as a result.
We ask how a text felt, and we keep the answer separate from what you scored, because the two disagree [25].
No “read 1,000 words a minute.”
The research says that speed is not reading [2].

Our own study

The gap in the evidence is transfer: does training on AI-generated text make you faster and safer on the AI text in your own inbox? We are running a pre-registered study to find out: professionals at B1–B2 English, several first languages, four weeks at ten minutes a day, against a group who read the same amount without the program. We measure effective speed on new AI texts, catch rate, transfer to human-written text, and a four-week retest. The plan is registered before the data exists, and the results will be published on this page whether they are good or bad.

Until then, the landing page makes no claim about results. It says what the program is built on, and it says “we’re testing.” We have not published any findings yet; the research notes page says what we will publish and at what threshold.

Stop dubbing in your head.

Every non-native reader knows the moment. The English sentence is too long, so you dub it into your mother tongue in your head, understand it, and switch back. It works. It is also slow, and by the afternoon it is exhausting.

We named the product after that moment, because that is the moment we want to remove. dub.mom teaches you to read AI English at the speed of your mother tongue, without the dubbing.

Sources

Numbered as cited above. Sample sizes and the main weakness of each are given so you can weigh them yourself.

  1. Schotter, Tran & Rayner (2014). Psychological Science. Eye-tracking; preventing re-reading harms comprehension. doi.org/10.1177/0956797614531148
  2. Rayner, Schotter, Masson, Potter & Treiman (2016). “So Much to Read, So Little Time.” Psychological Science in the Public Interest. Review. doi.org/10.1177/1529100615623267
  3. Chung & Nation (2006). English Teaching 61(4). N=49 Korean university students; ~73→132 wpm over 23 texts. No control group; comprehension not reported. doi.org/10.26686/wgtn.12552209
  4. Macalister (2008). The TESOLANZ Journal 16, 23–33. N=29 pre-university students in New Zealand on 400-word timed texts; 25 of the 29 gained between 5 and 143 wpm, and on a delayed post-test 15 of 29 showed no change or a decrease. Deliberately unlinked: the journal is not indexed and we found no stable URL.
  5. Tran Thi Ngoc Yen (2012). RELC Journal. Vietnamese EFL; treatment vs control; >50 wpm gain, transferred to other text types. doi.org/10.1177/0033688212439996
  6. Nation (2009). “Reading faster.” International Journal of English Studies 9(2), 131–144. Open access. The comprehension band is on p. 142, the 250 wpm goal for second-language readers on known vocabulary on p. 143. revistas.um.es/ijes/article/view/90791
  7. Chang (2010). Reading in a Foreign Language 22(2). N=84; +29 wpm vs +7 control. doi.org/10.64152/10125/66837
  8. Gorsuch & Taguchi (2008). System 36(2). N=50; assisted repeated reading; comprehension gain vs control. doi.org/10.1016/j.system.2007.09.009
  9. Chang & Millett (2015). System 52. N=64; audio-assisted reading; gains held at 3 months. doi.org/10.1016/j.system.2015.05.003
  10. Nakanishi (2015). TESOL Quarterly 49(1). Meta-analysis, 34 studies, N=3,942; d=0.46 (group), 0.71 (pre–post). Most included studies were small. doi.org/10.1002/tesq.157
  11. Jeon & Day (2016). Reading in a Foreign Language. Meta-analysis; adults benefit most. doi.org/10.64152/10125/66901
  12. Walker et al. (2005). Visual-syntactic text formatting; 75.6% vs 62.5% quiz scores over seven months. Developer-led study.
  13. Park et al. (2019). British Journal of Educational Technology. N=1,085 fourth and sixth graders; one year. One related study found plain extra spacing did as well as the phrase layout. doi.org/10.1111/bjet.12689
  14. Dempsey, Christianson & Van Dyke (2024). Reading and Writing. Linguistically-driven formatting improved comprehension for English learners and native speakers. doi.org/10.1007/s11145-024-10548-1
  15. Gu, Arawjo, Li, Kummerfeld & Glassman (2024). CHI. Grammar-preserving text saliency; N=18. doi.org/10.1145/3613904.3642699
  16. Gu, Raymond, Al Madi & Glassman (2024). CHI Extended Abstracts. Eye-tracking follow-up; N=24. doi.org/10.1145/3613905.3650956
  17. Yanagisawa, Webb & Uchihara (2020). Studies in Second Language Acquisition 42(2). Meta-analysis, 42 studies, N=3,802; L1 glosses > L2 glosses. doi.org/10.1017/s0272263119000688
  18. Chen, Ruan, Ju, Yap & Wang (2025). “More AI Assistance Reduces Cognitive Engagement: Examining the AI Assistance Dilemma in AI-Supported Note-Taking.” Proceedings of the ACM on Human-Computer Interaction (CSCW). Peer-reviewed. Within-subject, N=30, three conditions on lecture video: intermediate AI produced the highest post-test scores and automated AI the lowest, while participants preferred the automated setup. Measured on note-taking, not reading. doi.org/10.1145/3757632
  19. Srivastava, Healey & Jain (2025). “Priming at Scale.” CHI Extended Abstracts. N=44; AI-generated primes sped reading by about 7%. Preliminary. doi.org/10.1145/3706599.3720153
  20. Wallace, Bylinskii, Dobres et al. (2022). “Towards Individuated Reading Experiences.” ACM Transactions on Computer-Human Interaction. Individual font choice changed reading speed materially, without comprehension loss. One study. doi.org/10.1145/3502222
  21. Fredrick & Craven (2025). Frontiers in Education 10:1616935. ChatGPT vs L2 student essays; FKGL 16.61 vs 12.12. Essays, not work documents. doi.org/10.3389/feduc.2025.1616935
  22. Kobak et al. (2025). Science Advances 11(27). 379 excess “style words” in 2024; 900-word annotated list. doi.org/10.1126/sciadv.adt3813
  23. Friedman et al. (2026). IUI workshop; arXiv 2603.06878v1. N=24; error detection on incorrect AI outputs 25% (short), 54% (medium), 31% (long). Small; workshop paper. arxiv.org/abs/2603.06878v1
  24. Zhong, Zou & Adams (2025). arXiv 2509.10402v1. 82,845 developer–LLM conversations; median response-to-prompt token ratio about 14:1. Preprint. v1 is cited deliberately: the revised v2 reports a larger corpus (587,568 conversations) but drops the response-length finding entirely, so the ratio and that sample size cannot be quoted together. arxiv.org/abs/2509.10402v1
  25. Guo et al. (2026). Journal of Biomedical Informatics 179. N=150; LLM summaries rated clearer, comprehended worse. doi.org/10.1016/j.jbi.2026.105038
  26. Kramer & McLean (2019). Reading in a Foreign Language. Character-based speed measures for cross-text comparison. doi.org/10.64152/10125/66930
  27. Benedetto et al. (2015). Computers in Human Behavior. RSVP (Spritz) and comprehension. A formal reply was published in the same journal [34]; nothing here draws on it yet. doi.org/10.1016/j.chb.2014.12.043
  28. Acklin & Papesh (2017). American Journal of Psychology 130(2). RSVP; inferential comprehension degrades with speed. doi.org/10.5406/amerjpsyc.130.2.0183
  29. Readwise (2022). Company experiment, N=2,074; Bionic Reading 2.6 wpm slower on average. Not peer-reviewed.
  30. Snell (2024). Acta Psychologica 247. No significant difference in reading times with Bionic formatting. doi.org/10.1016/j.actpsy.2024.104304
  31. Beelders (2025). Journal of Eye Movement Research 18(5). Eye-tracking; no effect of Bionic formatting. doi.org/10.3390/jemr18050049
  32. Lee et al. (2025). CHI. N=319 knowledge workers; confidence in AI associated with less critical thinking. doi.org/10.1145/3706598.3713778
  33. Macalister (2010). Reading in a Foreign Language. A different design from the 2008 study: a comparison group, and transfer to authentic texts. The closer precedent for our transfer measure than for the fade. doi.org/10.64152/10125/66649
  34. Ricciardi & Di Nocera (2017). “Not so fast: A reply to Benedetto et al. (2015).” Computers in Human Behavior. A formal comment on [27] in the same journal. We have not read it, so no caveat on this page is drawn from it. doi.org/10.1016/j.chb.2016.12.047

For researchers and L&D

The instrument
The benchmark and checkpoints use 280–500-word passages at a fixed reference difficulty per block, rendered as plain text, with ten four-option questions (seven literal, three inferential), a comprehension floor of 70%, and one planted error detectable from the passage alone. Speeds are reported in raw and character-normalised words per minute.
Norms
Cohort comparisons are shown only for first-language × field × starting-band cohorts of 200 or more consented results, never as percentiles.
Data
Aggregate, consented results are available for research on request; individual data are not. Team dashboards receive checkpoint aggregates only.
Contact
[email protected]

Change log

  • 11 September 2026. Page published.
  • Entries added whenever a claim, a number, or a source changes, with what changed and why.