Case study · Product design · 2026
Making research readable for the people it affects.
Plain Paper turns a dense research abstract into a plain summary, decoded jargon and the questions worth asking. This is how we found the problem, what we tried, what we cut, and what changed.
All figures are illustrative sample data for a design prototype.
Outcome at a glance
Readers went from understanding a third of an abstract to explaining it back correctly.
- Comprehension score
- 34% → 81%
- Time to understand
- 11.2 → 3.4 min
- Confidence to discuss
- 2.1 → 4.3 / 5
- Task success (beta)
- 89%
Role
Lead product designer
Team
1 PM, 2 engineers, 1 researcher
Timeline
10 weeks
Platform
Responsive web
01 · The problem
Research is public. Understanding it isn't.
Abstracts are written for peer reviewers, not for the patients, parents, students and journalists who increasingly read them.
People were finding papers that mattered to them through news, social media and their doctors, then hitting a wall of jargon, statistics and hedged language. Most gave up, or worse, walked away with the wrong conclusion.
68%
of non-expert readers stopped reading an abstract before the end
1 in 3
misread the main finding, most often confusing correlation with causation
74%
turned to secondary news coverage, which overstated results in 41% of cases
Source: baseline study, n = 212 non-expert readers
02 · Who was affected
Four groups, one shared wall.
We recruited across groups who read research out of need, not profession.
38% of participants
Patients and caregivers
Reading studies about a diagnosis or treatment, often under stress, and needing to bring questions to a clinician.
27% of participants
Students
Undergraduates outside the field citing papers they only partly understood for essays and projects.
19% of participants
Journalists and writers
On deadline, needing to report accurately without overstating what a study found.
16% of participants
Policy and NGO staff
Briefing decision makers and needing the bottom line plus its limits, fast.
03 · Research
What we did to understand it.
Five methods over four weeks, mixing behavioural evidence with what people told us.
- Baseline comprehension test
- 212 readers summarised one of 6 abstracts in their own words. Two raters scored accuracy against an expert summary (inter-rater agreement κ = 0.82).
- Contextual interviews
- 18 sessions of 60 minutes watching people read a paper they cared about. We logged where they stalled, what they searched and when they quit.
- Diary study
- 9 participants logged every research encounter for two weeks: 143 entries, 61% triggered by a news story or a social post.
- Competitive audit
- Reviewed 11 tools: AI chat assistants, paper search engines and science news sites. None combined a summary with the original wording and its limits.
- Expert review panel
- 4 researchers (medicine, ecology, economics, physics) rated our generated explanations for accuracy and flagged overstatement.
Key insights
- 01Jargon was not the biggest barrier. Not knowing which parts mattered was. Readers spent 58% of their time on methods detail they did not need.
- 02People didn’t trust a summary they couldn’t check. 14 of 18 interviewees wanted to see the original words next to the explanation.
- 03The caveat is the finding. Misreadings almost always came from missing a limitation in the final sentence.
- 04Readers wanted to act, not just know: what to ask a doctor, what to check before sharing.
04 · Possible solutions
Six directions we explored.
We sketched broadly, then tested low-fidelity versions with 24 participants.
| Direction | Comprehension | Trust | Effort to build | Verdict |
|---|---|---|---|---|
| A. One-line TL;DR | +12% | Low | Low | Merged |
| B. Plain-language summary | +31% | Medium | Low | Chosen |
| C. Inline jargon highlights | +18% | High | Medium | Chosen |
| D. Open-ended chat with the paper | +22% | Low | High | Scrapped |
| E. Full-paper PDF upload | +26% | Medium | Very high | Deferred |
| F. Guided reviewer questions | +9% | High | Low | Chosen |
Comprehension lift vs. reading the raw abstract, n = 24
05 · What we chose
Summary, context, then judgement.
We combined three directions that each solved a different insight, in the order people need them.
- Three-part summary
- What they asked, what they found, what it means. A fixed structure beat free-form summaries by 14 points because readers always knew where the caveat lived (insight 3).
- At a glance band
- Bottom line, study type, who was studied and a "read with care" note. It absorbed the TL;DR idea (A) and puts the limitation next to the headline instead of after it.
- Jargon decoded in place
- Terms are highlighted inside the original abstract, not in a separate glossary. Readers could verify every claim against the source, which lifted trust ratings from 2.8 to 4.2 out of 5 (insight 2).
- Questions worth asking
- Three or four reviewer-style questions to take to a doctor, editor or teacher. Low effort, and the most-copied element in beta (insight 4).
- Reading levels
- Simple, Everyday and Informed. Students and policy staff wanted the numbers kept; caregivers wanted them gone. One toggle served both.
06 · Trade-offs
What shipped, and what we scrapped.
Ten weeks forced hard choices. Each cut protected accuracy or trust.
Shipped
- Abstract-only input. Abstracts are short, public and structured, which kept accuracy at 96% in expert review.
- Streaming summary so the first sentence appears in under 2 seconds, even when the full result takes 8.
- A permanent AI disclaimer in the footer, despite design debate about tone.
- Nothing stored. No accounts, no history. Privacy over convenience.
Scrapped or deferred
- Chat with the paper. Engaging, but answers drifted beyond the abstract in 23% of test sessions. Too risky for medical readers.
- Full PDF upload. Accuracy dropped to 81% on long papers and parsing tables doubled engineering scope. Deferred to v2.
- Credibility score per paper. Experts disagreed on scoring 40% of the time; a single number implied false certainty.
- Saved library and accounts. Requested by 22% of beta users, but not core to understanding. Cut to ship on time.
07 · Impact
How it changed reading.
We reran the baseline comprehension test with 96 new participants using the prototype, plus a 3-week open beta.
Correctly explained the main finding
Identified the study's main limitation
Finished reading to the end
Minutes to understand an abstract
4,860
abstracts explained in a 3-week beta
62%
of beta users returned within 7 days
+58
Net Promoter Score among caregivers
“I finally had the right questions for my son’s neurologist. I didn’t feel stupid walking in.”
08 · Next
What we’d do next.
- Full-paper support
- Revisit PDF upload with section-aware parsing, targeting 92% accuracy before release.
- Source linking
- Paste a DOI and pull the abstract automatically, removing the copy-paste step that 31% of users struggled with on mobile.
- Clinician handoff
- A printable one-page brief combining the summary and questions, requested by 12 of 18 caregivers.