# Visual Compression
# Book: Language
# Chapter URL: https://programmer.ie/books/language/05-chapter/
# Prose fingerprint (sha256, normalized): 17a4f9260afded41db911fd0790cefcfc10878860f900f7a47506d643630e4da
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

It is Monday morning and a security engineer faces forty browser tabs: weekend vulnerability disclosures, three vendor advisories, a thread discussing one of them, and two papers someone said she must read. She has an hour before stand-up. She will open perhaps six tabs and skim two. The others will be judged (kept or closed, read or abandoned) on the thinnest of evidence: a title, a favicon, a remembered recommendation, or the first three lines glimpsed on hover. Every judgement is a bet, and the cost of losing the bet depends on what she misses. An advisory she dismisses because its title reads like routine patch noise might describe the library her team ships. Opening every tab is not a solution either. An hour spent on marginal material can crowd out the one disclosure that needed investigation.
This chapter is about that bet, and about the first thing a flexible representation can do to help with it. The question is not how someone understands a page; that comes later. The question here is the narrower decision that comes first: is this worth my attention at all? Previews have always existed to answer it. Titles, snippets, thumbnails and abstracts are all previews. The new possibility is to derive a preview from a source and a declared current task, choosing a form that makes the reasons for selection quick to inspect. A compact text card may do this better than an image. So the real question is not whether a visual can be generated. It is whether a generated cue improves the decision over the best ordinary cue that already exists.
The chapter follows one path. The page stays fixed. A smaller representation is derived from it. Making something small is compression, and compression has effects: something becomes easier to see, and something may disappear. The question the whole book keeps returning to follows from that: what had to survive?
A smaller representation, and what it is for The purpose comes first, because everything else depends on it:
A preview succeeds for a stated task and attention budget when it helps the reader select worthwhile sources, without concealing the distinctions that make that selection correct.
This is a criterion to be tested, not a property that comes automatically from making something shorter. Testing selection quality requires a task-specific reference judgement: someone has to decide in advance which sources matter for the task. The costs of a missed source and a wasted opening must also be declared, not assumed to be equal.
So a good preview is not defined by having fewer words, fewer pixels or prettier cards. A preview that halves inspection time while hiding the decisive caveat can fail badly. A preview that takes the same ten seconds as a title but prevents an important skip may be valuable. The unit of analysis is the selection decision made under a limited budget. The useful questions are: what was opened, what was missed, what was delayed, and at what cost? The rest of the chapter stays close to those observations.
Readers deciding what to open behave in part like information foragers. Pirolli and Card’s information-foraging theory models how people use visible cues about material they have not yet reached: titles, snippets, surrounding text, images and layout. These cues provide information scent, which is a reader’s estimate of how valuable, and how costly, following a link is likely to be. Scent is not the actual value of the hidden page, and it is not necessarily accurate. It depends on the reader’s current goal and prior knowledge. The theory helps explain why a cue attracts a click or sends someone elsewhere. It does not, by itself, show that the resulting choice was correct.
In this vocabulary, a generated preview proposes a new scent cue. It takes the place of the ordinary snippet and suggests what lies down the trail and why opening it might matter for this task. A generated preview cannot know whether a source is worth engaging with just because it sounds confident. Its claims need to remain traceable to the source, and its value has to be judged against actual selection decisions. Two kinds of error matter here, and their costs depend on the work and the budget:
FALSE OPEN the preview looks relevant → the reader opens it and spends time that may displace another source FALSE SKIP the preview looks irrelevant → the reader does not open it in this pass and may miss task-critical evidence A false open costs attention, and when the budget is fixed, it can indirectly cause a false skip somewhere else. A false skip can lose the one item that mattered: the advisory, the contraindication, or the contradictory result. But the actual cost of a false skip depends on the task, on the consequences of delay, and on whether the source can be recovered later. Overall selection accuracy hides this asymmetry. So record false opens and false skips separately, distinguish high-consequence misses from routine ones, and evaluate the set of sources selected under the budget, not each card in isolation. A more attractive preview that increases consequential misses has not earned its place.
Conceptual failure comparison. The two errors are asymmetric and must be reported separately; a shorter card may omit background, but it may not change the claim. Open full-size diagram. Compression is not neutral shortening It is easy to describe compression as neutral shortening: the same thing, with less of it. It is not. Shortening involves choosing which relations, qualifications, entities, numbers and sources stay visible. Sometimes this choice makes the topic or conclusion easier to spot. Sometimes it removes the population, condition or exception that decides whether a source deserves attention for this reader’s task. A preview does not need to repeat the full content, but it must not turn a conditional claim into an unconditional one. Four effects are easily confused with one another:
attention gained ≠ understanding gained ≠ correct inference ≠ safe action A preview can draw the eye without helping anyone judge relevance. It can lead to a correct decision to open a page without conveying what the page says. A reader may also understand a short preview correctly and still lack the evidence needed to act safely. These are different outcomes: selection, comprehension, inference and action. A study of selection should measure selection. It should not borrow the language of understanding or safety until those outcomes have been tested separately.
The evidence discussed here mostly concerns visual and graphical abstracts used to communicate scientific research. It does not concern AI-generated previews used to sort a mixed browser queue. That makes these studies analogies, not validation of the proposed system. They show how attention, preference and reporting accuracy can come apart. None of them directly measures whether a task-conditioned preview helps the engineer choose the six pages she should open.
Visuals can attract engagement without showing whether readers chose the right source. Chapman and colleagues (2019) randomised 41 accepted surgical manuscripts to be shared as plain-English abstracts, visual abstracts or standard tweets. Visual abstracts produced more overall social-media engagement than the plain-English condition, and most of that engagement came from healthcare professionals rather than the public. The study measured interactions with posts. It did not measure whether readers understood the papers or made better decisions about which sources to read.
Liking a format is not the same as learning from it. Across three randomised trials involving students, consumers and doctors, Buljan and colleagues (2018) found no difference in measured knowledge between infographics and plain-language summaries of Cochrane reviews. Participants nonetheless preferred the infographic and rated its reading experience and user-friendliness more highly. This is evidence of a gap between preference and knowledge for those materials and populations. It is not a result about triaging sources.
Compressed scientific communication can also carry distortion. An observational study by Muller and colleagues, published in 2025, compared 119 infographics with the matching abstracts and full texts of randomised trials whose primary findings were non-significant. They found spin, meaning a framing that makes findings look more favourable than the data support, in 33% of infographics and 26% of abstracts and full texts. The overall difference between these was not statistically significant, although some specific comparisons of spin in results sections did differ significantly. The study therefore establishes that spin appeared in every form examined and that some reporting patterns varied by form. It does not establish that compression caused the spin, or that all infographics are worse. For a generated preview, the practical risk is concrete: highlighting a positive subgroup while leaving out the non-significant primary finding changes the cue the reader uses to choose.
Designers also cannot assume that people will stop at the intended boundary. The RIVA-C reporting-guideline project (Zadro and colleagues, 2024) was motivated in part by evidence that people sometimes use research infographics in place of the full article, even though infographics may leave out information needed for interpretation. The project’s output is a reporting checklist. It is not a trial showing that checklists or previews improve behaviour. The warning carries over to this design problem. A card labelled preview may be treated as the whole answer, so the interface must show what the card cannot establish and make the unchanged source easy to open.
These studies do not show that AI-generated previews cause bad selections, or that well-designed ones prevent them. What they show is why engagement, preference, source fidelity and task performance have to be measured separately. The proposed system would need its own experiment before anyone could say that compression improved a reader’s triage.
Three ways a preview could win If a visual preview does beat a title or an excerpt, it matters why it wins. Three different kinds of advantage can hide inside one result.
It carries more task-relevant evidence. A card can make a specific affected version, comparison, exception or negative result visible when a generic excerpt buries it. To avoid crediting the visual form for better information selection, compare it with a text card that carries the same facts. It conveys the same evidence faster. A spatial arrangement can help a reader find a relation sooner than prose does. To test this, compare matched factual content in text and visual layouts, and measure speed at comparable error rates. It is merely more attractive. Colour, polish and fluent labels can attract attention or raise confidence without improving selection. To test this, compare styled and plain versions with the same content. Preference alone is not evidence that the reader chose better. A fair test therefore measures what was selected, what was missed, how long decisions took, and how readers rated the format, without merging these into a single score. It must also separate two different advantages: choosing better evidence to show, and laying out the same evidence better. Without that separation, the experiment cannot tell whether generation, content selection or visual design produced any apparent gain.
A proposed selection experiment Start by freezing a corpus of short documents. Include routine advisories, relevant vulnerabilities, false alarms, qualified results, and sources whose importance is visible only in a body paragraph. Give each reader a declared task and a fixed time or opening budget. Before any previews are generated, independent reviewers should inspect the unchanged documents, record their task relevance and the consequence of missing them, and resolve any disagreements. A document may legitimately be marked uncertain rather than forced into a relevant or irrelevant label.
Compare at least four conditions on the same source–task pairs:
title and source only title with an ordinary extracted snippet a text-only task-conditioned preview a visual task-conditioned preview Where feasible, add an authored concise baseline, which strengthens the comparison. Hold source identity, access to the original and the selection budget constant. Where the hypothesis concerns layout rather than content, build a matched text and visual pair that carries exactly the same adjudicated facts. Counterbalance the order of conditions and the assignment of readers, so that one condition does not simply inherit what readers learned from another.
Record consequential false skips, ordinary false skips, false opens, sources selected within the budget, time to decision, source-opening behaviour, subjective preference and unwarranted confidence. Report denominators and uncertainty, and inspect errors by task and by document class. Before the run, declare the primary selection measure and any cost weighting. A headline improvement in clicks or satisfaction does not replace an improvement in correct selection under the budget.
This is a proposed experiment, not a result from this book. It could reject the visual approach altogether. A good excerpt or a well-written text card may turn out to be faster, safer and cheaper than generated imagery.
When the right preview is no preview Compression is an intervention, and interventions have to justify themselves. The default alternatives include the original title, an extracted passage, a supplied abstract, and no generated preview at all. Some sources should receive no additional visual rendering, and some should be marked insufficient evidence for a safe preview. The cases below are reasons to test abstaining. They do not claim that every instance of their class is impossible to compress.
An argument that is its hedges. In a methods critique, a legal analysis or a careful negative result, the qualifications may carry the conclusion. A tiny visual card that drops them is misleading. An excerpt that preserves the conditions, or the author’s own abstract, may be the better preview. If even that cannot preserve the point that matters for selection, leave the source uncompressed.
A distinction that is mostly numbers. Close hazard ratios, confidence intervals, absolute risks and budget percentages can be distorted by poorly scaled marks and areas. A compact table of exact values, with units and uncertainty intact, may work better than a decorative chart or an isolated number. Visual form is not ruled out. The risk is inexact visual encoding.
A short piece that its title already summarises. A title such as “Unauthenticated remote code execution in Acme Router firmware before 4.2” needs no cards. Generating a preview spends compute, attention and distortion risk to restate what a dozen words already say.
A concept that invites a false picture. Familiar visual metaphors can import structure the source never asserted. Evolution drawn as a ladder, a protocol drawn as plumbing, or risk drawn as an unqualified traffic light are examples. A generated image can make the wrong relation memorable. If the metaphor cannot be linked to supported claims and its limits cannot be shown, an exact excerpt is the better choice.
These cases suggest a selection policy: try the least transformative view that fulfils the task, and generate a richer one only if it promises a measurable advantage and can pass the relevant preservation checks. Record when the system abstains and why. A system that cannot leave a page alone cannot be trusted with anything stronger.
For a security advisory, the minimum fields a preview must allow a reader to check might be: the named component, the affected version range, exploit preconditions, severity or impact as stated by the source, important uncertainties, the publication or update date, and a link to the exact source passage. A shorter card may leave out most of the background. It must not silently change not affected into affected, turn under investigation into confirmed, or separate a version range from the component it qualifies. A relevance label generated by a model is a proposal. Checking it against the user’s actual dependency inventory is a separate operation.
A preview is not a summary One last distinction keeps this chapter from spilling into the next. A preview is defined by its intended job: to help someone decide whether to open a source. A summary is designed to convey more of a source’s substance. The two may look equally short, and users may treat either as a substitute for the original. Their obligations to preserve the source therefore differ according to intended use, not visual appearance or word count. A preview may omit background while keeping the cues needed for selection: topic, stance, the task-relevant result, the decisive qualification, and provenance. It must not invent an unsupported implication in order to make the source look worth opening. The interface should clearly state the preview’s limited purpose and provide a direct route to the unchanged source.
The claim here also stops at selection. Nothing in this chapter says that a preview helps anyone understand, remember or apply what the page says.
Where this leaves us A compressed representation is a candidate for decision support. It is not an automatically legitimate substitute for reading. It earns its place when, for a declared task and budget, it improves selection while keeping consequential false skips visible and avoiding unsupported additions. Compression is not neutral shortening. Attention gained is not understanding gained, correct inference or safe action. Information scent explains why a cue is followed, not whether it should have been. Attractive cards need an independent test, and leaving the source alone remains a serious baseline. The experimental question is whether a generated cue improves the chosen set of sources without creating a more persuasive route to the wrong ones.
The preview keeps the reader outside the page. The next question is what can be done once the reader is inside it: when does adding a diagram, a table or a timeline to the page itself help someone understand it?
References Pirolli, P. & Card, S. (1999). Information foraging. Psychological Review, 106(4), 643–675. Pirolli, P. (2005). Rational analyses of information foraging on the Web. Cognitive Science, 29(3), 343–373. Chi, E.H., Pirolli, P., Chen, K. & Pitkow, J. (2001). Using information scent to model user information needs and actions on the Web. Proceedings of CHI 2001. Chapman, S.J. et al. (2019). Randomized controlled trial of plain English and visual abstracts for disseminating surgical research via social media. British Journal of Surgery, 106(12), 1611–1616. Buljan, I. et al. (2018). No difference in knowledge obtained from infographic or plain language summary of a Cochrane systematic review: three randomized controlled trials. Journal of Clinical Epidemiology, 97, 86–94. Muller, R., Ferreira, G., Bejarano, G., Gamble, A.R., Kirk, J., Sindone, J. & Zadro, J.R. (2025). Do infographics “spin” the findings of health and medical research? BMJ Evidence-Based Medicine, 30(2), 84–90. DOI 10.1136/bmjebm-2024-113033. Zadro, J.R. et al. (2024). Development of the Reporting Infographics and Visual Abstracts of Comparative studies (RIVA-C) checklist and guide. BMJ Evidence-Based Medicine, 29(5), 342–345. DOI 10.1136/bmjebm-2023-112784.