Rauh, C. (2023). Clear messages to the European public? The language of European Commission press releases 1985–2020. Journal of European Integration 45(4): 683–701. doi:10.1080/07036337.2022.2134860. Open access. Available: the PDF only (20 pages, figures included). The replication corpus (Harvard Dataverse, doi:10.7910/DVN/UGGXUF) was not used; every number in the game is read off Figure 2 at 400 dpi and marked as such.
Primary: estimates. Mean values of three clarity indicators across 44,978 Commission press releases (1985–2020) and five benchmark corpora: UK and Irish government press releases, political-science abstracts, broadsheet and tabloid politics reporting (Fig. 2, upper panel); the annual Commission series (Fig. 2, lower panel); topic-matched Commission-vs-national differences (Fig. 3). Secondary: a measurement. The three indicators themselves, defined exactly: Flesch reading ease (sentence length and word length), average Google Books frequency of the words used, and the share of verbs among verbs and nouns (note 5).
A counterintuitive result, plus a mechanism made intelligible. The EU’s executive gained power and became publicly contested over 35 years, and repeatedly said it would communicate better. Its press releases did not get easier to read: they sit near academic abstracts, below national governments and far below newspapers, on every topic, and the line is flat from 1985 to 2020. The second half — what makes a text read like that — is the measurement: long sentences and long, noun-heavy words. The paper insists clarity is not “dumbing down” (pp. 686–7); the game lets the player check that claim by hand.
The European Union has an office in Brussels that writes news announcements meant for ordinary people. Over thirty-five years that office got more powerful, people started arguing about it, and it kept promising to explain itself better. A researcher measured how hard its announcements are to read, and compared them with governments, newspapers and professors. They read like the professors — and they never got easier. So: here is one of those announcements. Can you make it as easy to read as a government’s, without leaving anything out?
Mixed, labelled by part.
Aim (start screen, play screen, How to play): “Make a Commission press release as easy to read as a UK government one — with just two edits.” Verdict lines: (you) “Your release scores 45 — easier than the average UK government release.” / “…still harder than a UK government release.” (paper) “The Commission’s real releases average 27: closer to academic abstracts than to any government, on every topic, and flat from 1985 to 2020.” (means) “Every version you could pick said the same thing. Plain was possible; the paper asks why the Commission doesn’t write it.”
0:00 Start screen: a press release headline, “Say It Plainly”, the aim in a box. I press Start. 0:15 A ruler 0–80, “reading ease: higher is easier”. Five labelled dots: academics 14, Irish government 33, UK government 40, broadsheets 48, tabloids 65. “Where do the EU Commission’s releases sit? Tap the ruler.” I think: they’re government press releases, so near the UK… or maybe worse, it’s Brussels. I tap 35. 0:40 The blue dot drops at 27. “You put it at 35 — too clear. The Commission sits between the academics and the governments.” Two small rulers below: same picture for familiar words (Commission lowest of all) and verbs. I grin: worse than professors on vocabulary. 1:10 “Maastricht, the Santer resignation, the euro crisis, Brexit: Europe got contested. Did the Commission’s releases get easier to read?” I guess Easier — surely they learnt. The line draws itself year by year past each event: flat. 27.6 → 25.6. I frown; that’s the paper. 1:40 “Your turn.” A five-paragraph release, gauge at 21, two edit tokens. I read: the quote is awful (“a significant contribution to the empowerment of consumers…”). I tap it; two rewrites appear. “Shorter sentences” barely changes words; “Say who does what” is plain. I pick it: the marker jumps 21 → 35. One left. 2:30 Which next? The first paragraph is the longest sentence; I split it: 35 → 40.5? No — 39.5. Missed by half a point. “Still harder than a UK release.” I see the list: what each edit did, and “best possible with two edits: 47”. 3:00 Play again: quote with verbs plus the last paragraph in everyday words: 47. Easier than broadsheet? Not quite, but past the UK. I understand: long words cost as much as long sentences. 3:30 End screen: verdict; the paper’s line; why it might be (three reasons, as open questions).
The Midday Briefing: the bets put the paper’s finding in the first ninety seconds, and the editor gives a real decision with a budget — and the thing learned (what the score punishes) is the paper’s own measurement. Runner-up: Who wrote this?, switch if the author supplies real snippets from the replication corpus — it would then carry the “unchanged over time” finding as an experience rather than a chart.
Screens: start → bet 1 (place the dot) → bet 2 (the trend) → editor → end; How to play and About reachable throughout. One round each; the editor is replayable (“Try the edit again” skips the bets). Score: none for bets (graded in words); editor yardstick is the reading-ease score itself, in the aim’s units, against the UK line. Presets: none; the benchmarks are the paper’s corpora. Order: the paper’s result first (it’s the surprise), mechanism second (it explains the surprise).
Controls: tap a ruler (bet 1); three buttons (bet 2); tap a paragraph, tap one of its two rewrites (editor). Mechanism: mechanism.js — Flesch over the whole draft. Feedback: the drop (bet 1), the trajectory (bet 2), the counter-on-a-ruler with a delta per edit (editor). Moment: the line drawn flat past Brexit; then the quote rewrite jumping the marker by 14 points.
Fig. 1 (number of releases per month) except as a line in About; Fig. 3 (topic-by-topic) except as one sentence (“on 56 of 60 topic comparisons”); the topic-matching robustness check; the familiar-words and verb measures in the editor (we cannot compute the Google Books measure, and the paper’s verb share depends on spaCy tagging conventions we cannot reproduce on its scale — both appear only in the bet reveal, as the paper’s values).
From the paper: Fig. 2’s dot-and-ruler plot — the Commission blue dot among yellow (governments) and green (academics and newspapers). Rules: one horizontal ruler is the only chart type; Commission blue, governments yellow, others green, as in the paper; the release is set like a press release (headline, dateline, serif text); the player’s marker is a big ink triangle. Legibility: on the editor the player looks at one ruler (full width, 5 labelled ticks) and five paragraph cards; nothing else moves.
Fidelity result (node fidelity.js, full output in fidelity-result.txt): ALL PASS.
PASS platypus sentence = 24.44 (published 24.4, Wikipedia 'Flesch–Kincaid readability tests')
PASS "The cat sat on the mat." = 116.15 (published 116.1)
PASS every release word has a hand count (169 words)
PASS syllables() matches the hand count for all of them
PASS Commission below both governments on all three indicators (p. 692)
PASS Commission below broadsheets and tabloids on all three (p. 692)
PASS Commission closer to political-science abstracts than to broadsheets on reading ease and verbs (p. 692)
PASS Commission below political scientists on familiar words (p. 692: "performs even worse")
PASS Commission in Flesch's graduate band (<30), UK above it (p. 692: university graduates vs high school) — 26.7 vs 40
PASS annual reading ease flat: range 3.0 points, first 27.6 last 25.6 (p. 693: "hardly changed")
PASS in no year does the Commission reach the Irish government average (Fig. 2 lower panel)
PASS original release scores 21.1: inside the Commission spread (Fig. 2 lower panel, ~10–45) and below the Irish average
PASS original within 7 points of the Commission average 26.7
PASS with 2 edits, 9 of 40 plans reach 40 (best 46.7: -1,-1,-1,1,0)
PASS a random plan succeeds 23% of the time (< 50%: the choice matters)
PASS no two-edit plan reaches the broadsheet line 48.1 (the aim is the UK line, not beyond)
PASS a player who always picks the single biggest improvement reaches 46.7 (sensible play succeeds)
PASS plain syllable heuristic: plan ranking correlates with the game's (Spearman 0.983)
PASS plain syllable heuristic: its best plan (-1,-1,-1,1,0) also reaches the UK line in the game (46.7)
INFO Flesch–Kincaid grade formula: plan ranking correlation with the game's Spearman 0.584 — this formula weighs sentence length more, so the order of middling plans shifts; the claim the game makes is checked below
PASS Flesch–Kincaid grade formula: its best plan (-1,0,-1,1,-1) also reaches the UK line in the game (42.9)
PASS hand-checked counter: mean one-edit gain — shorter sentences 5.4, everyday words 9.1, who-does-what 8.9
PASS plain heuristic: mean one-edit gain — shorter sentences 5.4, everyday words 9.2, who-does-what 8.8
PASS every rewrite improves the score under both formulas (no edit is a trap)
PASS single best edit under both formulas: p4 verbs (Flesch +14.9; FK grade −2.9) / FK best p4 verbs
INFO one-edit gains (Flesch points / FK grades): p1-split 4.8/2.36 p1-verbs 4.1/0.91 p2-split 8.4/2.82 p2-words 9.2/1.86 p3-words 7.6/0.94 p3-verbs 8.6/1.44 p4-split 2.9/1.28 p4-verbs 14.9/2.88 p5-words 10.5/1.76 p5-verbs 8.0/1.37
PASS split rewrite of p1 adds sentences (1→3) and keeps word length (1.70→1.81 syll/word)
PASS split rewrite of p2 adds sentences (1→3) and keeps word length (1.65→1.58 syll/word)
PASS split rewrite of p4 adds sentences (1→2) and keeps word length (2.32→2.33 syll/word)
ALL PASS
Robustness: under a plain syllable heuristic the ranking of all 40 two-edit plans is nearly identical (Spearman 0.98). Under the Flesch–Kincaid grade formula, which weighs sentence length more, the ranking shifts (Spearman 0.58). There, splitting sentences is worth about as much as the other rewrites, but the single best edit (the quote, with verbs) is the same, and that formula’s best plan also reaches the UK line in the game. So the end screen’s lesson is worded “on this score”, and About states the caveat.
Build log.
session-test.js, log in session-log.txt): bet, trend, two edits with help opened and returned mid-edit, end screen, retry on the best plan, About and back, restart. No errors, no horizontal scroll at 380 px.