Progressive disclosure

The pattern rests on one pair of small experiments. Carroll and Carrithers disabled rather than removed the functions new learners rarely need in a commercial word processor — the ones that act as springboards for errors and confusion — and reported better learning efficiency, in one experiment on learning by doing and one on learning from the manual. That is 1984, one lab, a text-only editor, measured as learning rather than as conversion, and the companion paper in Communications of the ACM is paywalled, so its numbers could not be verified in this run. Treat the training-wheels result as a direction rather than an effect size.

The chunking evidence that is genuinely strong is about lessons, not interfaces. Rey and colleagues' meta-analysis of segmenting, pooled in Noetel and colleagues' overview of 29 reviews covering 1,189 studies and 78,177 participants, found chunked material improved learning at g = 0.34 and modestly reduced perceived load at g = 0.23. Two details in the same analysis cut against the way product teams use it. Segmenting cost time — learners took considerably longer with segmented material, g = 0.92 — and segments chosen by the instructor beat segments the learner chose, g = 0.41 against g = 0.20. The gain comes from someone deciding what belongs together, not from handing the user an expander.

In forms, the place teams reach for it most, the controlled evidence runs the other way. Mavletova and Couper randomised 2,000 invitations each to a scrolling and a paging version of the same mobile questionnaire. Paging, one question per screen, took twice as long — 9.1 minutes against 4.5 — and did not reduce dropout: breakoff was 10.2% against 8.0%, not significant at p = 0.067. It moved the dropout instead. First-screen breakoff was 3.3% with paging against 6.0% with scrolling, then 6.9% against 2.1% on the screens after it, and over the same first eight questions the cumulative rates were indistinguishable, 6.0% and 7.5%. Item nonresponse did not differ, and more people reported technical problems with paging, 25.8% against 18.9%. It is one volunteer panel in Russia, but the PC precedent it replicates is large: Peytchev and colleagues ran over 21,000 undergraduates through up to 268 questions, and in both studies respondents rated scrolling more satisfying. Splitting a form chooses where you lose people, not whether.

Hiding the way around a product has a measured cost in discoverability. The best quantitative test is practitioner work, not peer review: Pernice and Budiu ran 179 participants through two tasks on each of six live sites, on desktop and phone, comparing hidden, visible and combo navigation. Hidden navigation was used in 27% of desktop cases against 48% visible and 50% combo, and in 57% of mobile cases against 86% combo; content discoverability dropped by more than 20% on both devices; tasks took at least 39% longer on desktop and 15% longer on mobile; and participants rated the hidden version 21% harder. It is unmoderated remote testing on 2016 sites with no published statistics beyond the differences, so read it as the size of the risk rather than a coefficient.

The deeper layer, meanwhile, is close to unread. Bakos, Marotta-Wurgler and Trossen tracked around 48,000 monthly visitors to the sites of 90 online software firms: one or two shoppers in every thousand opened the licence agreement at all, most of them reading a fraction of it. Obar and Oeldorf-Hirsch had 543 people join a fictitious social network: 74% took the quick-join option and skipped the privacy policy, those who opened it spent 73 seconds on a document needing 29 to 32 minutes at adult reading speed, 97% accepted it, and 98% missed clauses handing their data to the NSA and their employer and their first-born child to the service. Springer and Whittaker add a subtler failure: offered both a coarse summary and a word-by-word explanation of the same sentiment model, their 74 participants split 37 to 37 on which they preferred, and 24% of those choosing the coarse view praised it for weighing their whole text — which was false, since both ran the identical word-level model. Anything behind a disclosure is, for practical purposes, not shown, and a thin layer can leave a flattering wrong mental model rather than none.

The line between deferring detail and hiding terms is drawn in law rather than in taste. Blake and colleagues moved fees from the listing page to checkout for several million StubHub users over ten days, raising revenue about 21% and the transaction rate 14.1%, with at least 28% of the gain coming from people buying better seats than they would have chosen with the fee in view. Deferred cost is profitable, which is why it is now regulated. The FTC's Rule on Unfair or Deceptive Fees, 16 CFR Part 464, effective 12 May 2025, requires any price shown for a live-event ticket or short-term lodging to carry the total price clearly and conspicuously and “more prominently than any other pricing information,” and defines clear and conspicuous in an interactive medium to mean the disclosure “must be unavoidable” — a definition a disclosure behind a tap fails by construction. Article 6(1)(e) of the Consumer Rights Directive requires the total price inclusive of taxes and any delivery charges before the consumer is bound, and under Article 6(6), where those charges were not disclosed, “the consumer shall not bear those charges or costs.” The EDPB endorses layered privacy notices for the concise-and-easily-accessible requirement of Article 12(1) GDPR, then bounds it: the layered approach “may not be misused to hide information in deeper layers or by adding unnecessary layers,” a control buried across many pages is the Privacy Maze pattern in Annex I 4.1.2 and runs contrary to Article 12(2), and for exercising rights “the number of steps should be as low as possible.” Mathur and colleagues' crawl of roughly 11,000 shopping sites puts the measured prevalence low — 1,818 dark patterns across 1,254 sites, of which 5 were hidden costs, 7 sneak-into-basket and 14 hidden subscriptions — but the crawler only visited product, cart and checkout pages, so those are a floor, not a census.

So the honest grade is mixed, and the asymmetry matters more than the average. That hiding the rare and advanced helps rests on one 1984 pair of experiments plus a learning literature about chunked lessons that also measures the time it costs; that what you hide stays hidden has far better support, from one-in-a-thousand licence openings to a fifth less content discovered behind a menu icon. Use progressive disclosure for optional depth — the breakdown, the advanced setting, the explanation someone might want — and keep the complete set of facts a decision actually needs in the first layer, because that is the layer users read.