In one paragraph
Study 1 looked at 39,867 quiz answers from 2,590 learners. People missed the last question of a course about as often as the first one, so putting the important material first does not help. Learners also got about 1.5% faster with each lesson without making more mistakes. The study was preregistered on the Open Science Framework and is published as a preprint at edarxiv.org/9e5jy.
Study 1: Does attention fade during a short course?
Course designers are often told to put the most important material first, because attention supposedly drops as a course goes on. It is a sensible-sounding rule. We could not find anyone who had tested it on a real learning app, so we did.
A NerdSip course is 3 to 10 short lessons on one topic, taken in order, often spread over several days. Each lesson is about 130 words, shown a screen at a time, and ends with one easy three-option question. You have to answer it correctly to move on. We asked a simple question: is the question at the end of a course missed more often than the question at the start?
We wrote down our hypotheses and our analysis plan and registered them publicly on the Open Science Framework. Thirty-three minutes later we pulled the data: 39,867 first answers from 2,590 learners, across 994 courses and 5,169 questions, frozen on 20 July 2026.
What we found
| Question | Answer | How sure |
|---|---|---|
| Are later lessons missed more often? | No. The last lesson of a course is missed about as often as the first. Odds ratio 0.978 per position, and the interval rules out any effect big enough to justify front-loading. | Preregistered. Held in every check we ran. |
| Do people slow down or speed up? | They speed up, by about 1.5% per lesson, and they do not make more mistakes while doing it. | Preregistered. Some of the speed-up is people tapping through, and we say so. |
| Do errors rise inside one long sitting? | A little, but only in very long sittings, only for a few dozen learners, and smaller than the smallest effect we had registered as meaningful. | Preregistered. Detectable, below our own bar. |
What we thought we found, and did not
For a while it looked as if longer courses had more errors. The gap was there from the very first lesson, though, so nothing was building up. And it turned out to track who had generated the course, not how long it was. We report that openly in the paper. It is not a finding.
What this does not tell you
It is one app, with deliberately easy questions, and it describes learners who keep going. People who quit after a wrong answer are not in the later lessons. We did not measure attention itself, only whether a question was missed. Everything is an association, not a cause. The paper spells all of this out.
Read the paper
- Preprint on EdArXiv, open access, CC-BY 4.0
- PDF, 8 pages
- Preregistration, registered 20 July 2026 before the data were extracted
- Code, result tables and figure data on the Open Science Framework
How to cite
Sergelius, P., & Hänze, M. (2026). Quiz errors do not rise with lesson position in mobile micro-learning: a preregistered analysis of 39,867 outcomes. EdArXiv preprint. https://edarxiv.org/9e5jy
The preprint has been submitted to a peer-reviewed journal. This page will link the published version when it exists.
Study 2: What people pick from the feed
The second study looks at 253,000 swipe decisions in the NerdSip feed and asks whether taste in topics differs by country, or mostly by person. It uses a split sample: we explored one half, wrote down what we expected to see, registered it at osf.io/7xp9q, and only then opened the other half. The write-up is in progress.
How we do research
- We register first. Hypotheses, sample rules and the exact analysis are written down in public before we touch the data. Anything we add afterwards is labelled as such.
- We publish everything. Results that support our hunches and results that do not. Every departure from the plan is listed, numbered, with the reason.
- We share the code, not the people. Analysis code, result tables and figure data are public. Row-level learner data is not: it is pseudonymous personal data under the GDPR. Every published number rests on at least 10 different learners.
- We say who we are. The authors founded NerdSip. That is a conflict of interest, and the preregistration, the frozen dataset and the open deviation log are our answer to it.
- We say what AI did. Analysis code, figures and drafting were produced with the help of Claude (Anthropic). The authors checked every analytic decision and are responsible for the content.
For journalists
Three sentences you can use, all backed by the paper:
- In 39,867 quiz answers from a learning app, the last lesson of a course was missed about as often as the first. The idea that attention fades across a short course did not hold up.
- Learners got about 1.5% faster with every lesson and did not make more mistakes while doing it.
- The study was registered publicly before the data were touched, and the code is open.
Questions, interviews, or the underlying tables in another form: [email protected]. Dr. Philip Sergelius and Dr. Max Hänze are the authors and the founders of NerdSip, Reinbek near Hamburg, Germany.
Related reading on NerdSip
- How to build a longer attention span
- How to design a micro-learning lesson
- How NerdSip fact-checks its lessons
- How to learn faster, the science-backed method
- The human brain, region by region
Questions people ask
Is the study peer reviewed?
Not yet. It is a preprint, which means it is public and citable but has not been through journal review. It has been submitted to a journal. This page will link the reviewed version when it is out.
Can I quote or reuse the findings?
Yes. The preprint is published under a CC-BY 4.0 licence, so you can copy, quote and share it as long as you name the authors.
Did you publish user data?
No. Learners were replaced by random codes on the server before anything was extracted, and only aggregate numbers are published. Every published figure rests on at least 10 different learners. The row-level data and the key to the codes are not shared.
Why publish a result that says nothing happened?
Because the belief we tested shapes how courses are built, and a careful, registered test that finds no effect is exactly the evidence needed to stop following a rule that has no support.
Did AI write the paper?
AI helped. Analysis code, figures and drafts were produced with Claude by Anthropic. The authors reviewed and approved every analytic decision and take full responsibility, and the paper says so.