Written to Be Agreed To, Not Read

Everyone taps “I Agree.” Almost no one reads what they are agreeing to, and that is not a failure of diligence — it is the design. In 2013, a draft paper (Reck and Reck), “Mobile Application User Agreements: A look at variation and complexity,” set out to measure this exact problem across a corpus of mobile-app End User License Agreements. The draft made the argument well but stalled before the measurements were done; its key sentences still read, in the surviving file, “Of 100 EULAs sampled only XX contained a version number, and XX contained a date.” The XX was never filled in.

This finishes the paper. The original corpus survived — 92 readable mobile-app agreements collected in 2012–13, from banks, games, social apps, shopping, and mapping. Run through a bibliometric analysis, they replace every XX with a number, and the numbers are worse than the draft guessed.

Unreadable by design

Take the agreements purely as text and measure them the way you would measure any corpus. The median agreement runs 2,875 words — about eleven minutes of reading — with sentences averaging 27.7 words, well past the ~20-word ceiling plain-writing guidance recommends. Run the standard readability formulas and they agree with each other: a median Flesch Reading Ease of 29, which sits below the “very difficult” line, and a consensus grade level of 20 — postgraduate. The cited figure of ~41 apps on the average phone turns that into roughly seven and a half hours of postgraduate reading to know what a single phone has agreed to.

The difficulty is not incidental; it is lexical. Across 330,000 words, the register is relentlessly legal: shall appears 3.1 times per thousand words, “including without limitation” 286 times, “sole discretion” and “indemnify” throughout. This is not language chosen to be understood. It is language chosen to be enforceable.

What you actually agree to

Difficulty would matter less if the terms were benign. They are not. Classifying every agreement by the substantive clauses it contains produces a clear ledger — and it tilts one way. The most common clause, in 86% of agreements, disclaims the company’s liability (“provided AS IS”). The next two, at 75% and 73%, are the ones that should give any reader pause: the company may change the terms at any time, and you are bound to the new terms simply by continuing to use the app. Well over half also make you indemnify the company (59%), let it terminate you at its discretion (57%), and pin any dispute to its choice of law and venue (61%).

Read the colors, not the bars: nearly every high-prevalence clause is a right the user gives up or a shield the company keeps. There is no symmetric column of user protections, because these documents are not negotiated. They are issued.

The consent trap

Here is where the draft’s instinct was exactly right, and the data makes it undeniable. Put two facts together. Three-quarters of these agreements reserve the right to rewrite themselves at any time, and bind you to whatever they become by your “continued use.” Now the numbers the draft left blank: only 3% carry a version number, and only 37% carry a date.

That combination is a trap. You are bound, automatically and continuously, to a document that can change without notice — and that carries no version stamp and usually no date, so you cannot tell whether it has changed, or which text you supposedly accepted. Informed consent is not merely unlikely here; it is structurally impossible. You cannot consent to a moving target you are not permitted to identify. The clauses say so plainly:

Naugatuck (bank): “Continued use of the Mobile Banking Service and any related Software constitutes your agreement with and acceptance of these Terms and Conditions, as well as any future changes.”

Conduit: “Conduit reserves the right to modify… this EULA at any time and without prior notice… By continuing to use the Software… you agree to be bound by the modified EULA.”

The fine print, verbatim

Aggregate percentages can feel abstract. The individual clauses do not. These are pulled directly from the corpus — the agreements users tapped “Agree” on.

An app for recipes reserves the right to sell your personal information to “carefully selected companies.” A travel app and a music app each make you waive your right to a jury trial and to join a class action. A payment app makes you agree to defend and indemnify the company. A state government’s own app can cut off your access without notice, for any reason it deems harmful. None of these is hidden in a rogue outlier; each is representative of a whole column in the ledger above.

What would actually fix it

The draft gestured at the remedies, and thirteen years later they still stand — because they were never really adopted. The Plain Writing Act of 2010 showed that plain-language mandates are possible; they simply were never extended to consumer contracts. Two structural fixes would cost vendors almost nothing and change everything for users: version and date every agreement, so a user can tell what they accepted and whether it changed; and converge on a small number of standard, plain-language licenses the way open-source software did with the GNU licenses, instead of 92 bespoke documents that vary on every axis. A change log at the top of each agreement — the ordinary courtesy of any other living document — would do more for informed consent than another thousand words of “including without limitation.”

Until then, the finding of the finished paper is simple. These agreements are written to be agreed to, not read. The bibliometrics, at last with the XX filled in, say so precisely: postgraduate prose, eleven minutes each, three-quarters silently rewritable, three percent versioned. The reader never had a chance — and that was the point.

Notes and references

1. Reck, R. and Reck, R. (2013, draft). Mobile Application User Agreements: A look at variation and complexity. The unfinished draft and its 2012–13 corpus of 92 mobile-app EULAs — the source material completed here.
2. Corpus bibliometrics computed over 330,391 tokens: length, type-token ratio, mean words/sentence, and a readability suite (Flesch Reading Ease, Flesch–Kincaid, Gunning Fog, SMOG, Coleman–Liau, ARI).
3. Plain Language Action and Information Network. Plain Writing Act of 2010. plainlanguage.gov.
4. Verbatim clauses quoted from the corpus agreements of Epicurious, Travelocity, Spotify, Pinterest, Electronic Arts, PayPal (Here), the State of Michigan, Conduit, Naugatuck Savings Bank, and BancorpSouth.