Visual Compression

It is Monday morning and a security engineer faces forty browser tabs: weekend vulnerability disclosures, three vendor advisories, a thread dissecting one of them, two papers someone said she must read. She has an hour before stand-up. She will open perhaps six tabs and skim two. The other thirty-four will be judged — kept or closed, read or abandoned — on the thinnest of evidence: a title, a favicon, a remembered recommendation, the first three lines glimpsed on hover. Every judgement is a bet. Most bets are cheap when wrong. A few are expensive: the advisory she dismisses because its title reads like routine patch noise is the one describing the library her team ships.

This chapter is about that bet. Not about understanding a page — that is Chapter 5’s question — but about the narrower, prior decision: is this worth my attention at all? The proposed intervention is visual compression: a small set of generated visuals that previews a page’s structure before the reader pays the cost of reading it. The chapter earns a deliberately narrow verdict on that intervention, and it builds the error framework Part II will reuse at much larger scale.

The operational definition first, because everything else hangs on it:

Visual compression succeeds when a smaller, faster representation preserves enough task-relevant information for a reader to make an accurate selection decision.

Not fewer words. Not fewer pixels. Not prettier cards. A preview that halves reading time while hiding the decisive caveat is a failure. A preview that takes the same ten seconds as a title but prevents one catastrophic skip is a success. The unit of analysis is the selection decision, and the chapter never lets the discussion drift from it.

Selection has its own machinery: scent

Readers deciding what to open are foragers, and foraging theory gives this chapter its mechanism. Pirolli and Card’s information-foraging theory (1999, with a programme of Web studies through the 2000s: the SNIF-ACT model, scent-based navigation, the “law of surfing”) models seekers who judge distal content — the page at the other end of the link — from proximal cues: titles, snippets, surrounding text, images, layout. That judgement has a name: information scent, the imperfect, subjective perception of a source’s value and cost obtained from its cues. Two behaviours follow. Scent-following: pursue the trail whose cues promise most. Patch-leaving: abandon the area when scent runs thin. Anyone who has watched themselves skim search results has watched both policies execute.

A visual preview is, in this vocabulary, an engineered scent cue. It stands where the snippet stands and makes the same promise: this is what lies down this trail, and here is what it will cost you. That framing does three useful things. First, it locates the preview’s job precisely — it does not need to convey the content, only to predict the value of engaging with it. Second, it predicts the failure modes exactly: scent can be wrong in both directions, and foraging theory already names them. Third, it connects this chapter to Part II, where the same scent logic governs attention allocation across hundreds of items rather than dozens of tabs.

The two failures, stated asymmetrically because they are asymmetric:

FALSE OPEN
preview looks relevant
→ reader opens, wastes minutes, returns unharmed

FALSE SKIP
preview looks irrelevant
→ reader never sees the source; the loss is unbounded

A false open costs time. A false skip can cost the one item that mattered — the advisory, the contraindication, the contradictory result. Any evaluation that averages the two into “accuracy” buries the chapter’s most important number. EXP-04 below reports them separately, and the design treats a preview regime that trades fewer false opens for more false skips as regressing, however attractive its cards.

What the evidence actually says about visual previews

The closest existing literature is not about AI-generated previews at all — it is about visual and graphical abstracts in science communication, and it reads like a cautionary tale this chapter takes seriously.

Visuals attract; attraction is not selection quality. Chapman and colleagues’ randomised trial of surgical-research dissemination (British Journal of Surgery, 2019: 41 manuscripts across plain-English abstracts, visual abstracts, and standard tweets) found visual abstracts drew significantly more engagement overall — but the engagement came overwhelmingly from healthcare professionals, public engagement stayed near zero across all arms, and standard tweets produced the most click-throughs to the full text. Engagement with the cue is not the same as correct routing to the source. A preview can win attention while losing the decision.

Knowledge gains match plain language, not exceed it. Buljan and colleagues (Journal of Clinical Epidemiology, 2018: three randomised trials comparing infographic, plain-language summary, and scientific abstract of Cochrane reviews, across students, patients, and doctors) found no difference in knowledge acquired from infographic versus plain-language summary — both beat the scientific abstract — while participants preferred the graphical form. Preference and performance separate exactly as Chapter 2 warned. For this chapter the implication is sharp: a preview readers love is not yet a preview that routes them correctly.

Compressed visuals inherit — and can amplify — reporting distortion. A 2024 cross-sectional study in BMJ Evidence-Based Medicine examined 119 randomised trials with non-significant primary outcomes and found spin in 33% of infographics against 26% of abstracts and 26% of full texts — a difference that did not reach statistical significance, but with infographics two to four times more likely to carry spin in the results section specifically, and with the standard repertoire (spotlighting secondary outcomes, presenting non-significance as equivalence, highlighting benefit despite the primary result). Spin, in other words, survives compression and sometimes concentrates in it. A generated preview that drops the “no significant difference on the primary outcome” qualifier while illustrating the encouraging subgroup is not a compression artefact. It is a misrepresentation with good layout.

Readers substitute; substitution raises the stakes. The RIVA-C reporting-guideline project (BMJ Evidence-Based Medicine, 2024) documents the behaviour that makes all of the above dangerous: large fractions of readers use infographics as a substitute for the full text, believing them detailed enough to stand alone — while content analyses find most infographics omit the elements needed for correct interpretation (population, comparator, harms, effect size, uncertainty, limitations). And the BMJ’s own pragmatic evaluation of its visual abstracts noted a novelty halo: strong early social-media performance that faded, suggesting some of the apparent win was unfamiliarity, not utility. A preview designed for selection will be used for understanding no matter what the designer intended. Chapter 9’s preservation instrument exists partly because of this predictable misuse; this chapter notes it as a constraint on preview design (every card links to source; nothing in the card set may contradict the source’s headline result).

Taken together, the evidence licenses a sober, useful posture: visual previews are plausible scent cues with real attractiveness, no demonstrated selection-accuracy advantage yet, demonstrated distortion risks, and a user population inclined to over-trust them. That is precisely the posture from which to run the experiment rather than to declare the victory.

Three effects that must not be confused

If a visual preview ever wins in EXP-04, the next question is why — because three entirely different victories hide inside one accuracy number:

1. It carries more useful semantic information. The cards surface the page’s actual structure — its claims, comparisons, exceptions — in less space than the excerpt. This is the victory the book wants, and it predicts accuracy gains that survive even ugly, unstyled cards.
2. It communicates structure faster. Same information as the text summary, decoded sooner — parallel visual parsing, spatial layout doing the work Chapter 2 described. This victory shows up in decision speed at equal accuracy, and it is genuinely valuable under attention scarcity.
3. It is merely more attractive. Colour, polish, novelty, and fluency pull the eye and raise confidence without improving the underlying judgement. This victory shows up in preference and possibly speed, with accuracy flat or worse — the most dangerous outcome, because it feels like success while manufacturing false skips and false opens with higher confidence.

The experiment therefore measures three outcomes separately — decision quality (correct selections), decision speed (time to a committed keep-or-close verdict), and preference (which preview readers say they want) — and the chapter predeclares its most valuable finding: a condition that wins preference while losing selection accuracy. That result would do more for the book than a clean win. It would prove the evaluation discipline is measuring the right things, and it would warn every later chapter that attractiveness is a confound to be controlled, not a goal to be maximised.

When the right preview is no preview

Transformation is an intervention, and interventions must earn their use. Even in a chapter devoted to visual compression, a substantial class of pages should receive UNCHANGED — no generated preview at all, just title and source. Four cases, each with its mechanism:

• Qualification-heavy argument. A page whose point is its hedges — a methods critique, a legal analysis, a careful negative result — compresses into apparent certainty. Any card set either reproduces the full hedge structure (no compression achieved) or drops it (the point destroyed). The honest preview is the title plus the abstract.
• Primarily numerical distinction. Two treatments differing by a hazard ratio of 0.87 versus 0.91, or two plans differing by 3% of budget: visual encoding invites magnitude illusions (bar lengths, area comparisons) that the numbers stated plainly would not cause. Text carries this; pictures distort it.
• Short piece already summarised by its title. “CVE-2026-XXXX: unauthenticated RCE in Acme Router firmware < 4.2” needs no cards. Generating a preview spends compute, attention, and distortion risk to restate what twelve words already decided. NOTHING — or rather, title-only — is the correct compression level, a preview of Chapter 17’s ladder.
• Concept inviting a false visual metaphor. Some ideas come with stock imagery that smuggles wrong structure: the brain as circuit board, evolution as a ladder, a network protocol as plumbing, risk as a traffic light. A generated image that reaches for the obvious metaphor teaches the wrong mental model faster than prose would. Where the available visual vocabulary misleads, text is the safer channel.

These are not edge cases to be patched later. They are the chapter’s positive contribution to the book’s recurring principle: transformation is an intervention that must earn its use. The router of Chapter 5, the ladder of Chapter 17, the attention policy of Chapter 19, and the delegation contracts of Part IV all inherit this default-deny posture from here. The system that cannot say UNCHANGED cannot be trusted with anything stronger.

A selection representation is not a summary

One distinction remains, because Chapters 4 and 13 will otherwise bleed into each other across Part I and Part II. This chapter’s object:

source
  ↓
compression
  ↓
selection representation

is built for one purpose: help me decide whether to spend attention on the source. A summary, in the ordinary sense, implies something broader: help me understand the source without reading it. Those are different contracts with different fidelity requirements. A selection representation may legitimately omit everything except what discriminates this source from its neighbours — topic, stance, key result, decisive caveat — because its job ends the moment the reader decides. A summary that omits the same material has failed. Conversely, a selection representation that tries to teach the content has overbuilt: slower to scan, more to get wrong, duplicating Chapter 5’s job. And a final scope fence for both chapters: previews and enhancements are scored on selection accuracy and task-bounded understanding respectively — neither chapter measures retention, transfer durability, or learning, and neither claims them.

The book-level ownership, fixed here and referenced henceforth:

• Chapter 4 — compress for selection. Route attention; measured by decision quality and false-skip rate.
• Chapter 5 — enhance for understanding. Improve the encounter with the page itself; measured by comprehension and transfer.
• Chapter 13 — compress a long source into its important information units. Source-level condensation; measured by coverage and fidelity.
• Chapters 14–15 — compress those units for this person’s purpose and knowledge. Task relevance, then personal novelty; measured against declared goals and known-concept ledgers.

Each stage narrows the previous one’s output under a new criterion, and each keeps the path back open: selection expands to source, enhancement exposes source, units link to timestamps, personal compression preserves the unfiltered set. The ladder of Chapter 17 will formalise what this chapter begins.

What this chapter earned

Visual compression is legitimate only as decision support: a preview earns its place by improving selection accuracy — with false skips counted separately and weighted heavier — without hiding decisive information. Information scent is the mechanism; attractiveness is the confound; UNCHANGED is a standing candidate; and the selection representation is fenced off from the summary. The method for verifying all of this is the experiment below. Whether generation can improve understanding is explicitly not decided here — it is the next chapter’s question.

The next step keeps the page intact and asks when a generated representation genuinely improves it.

References

• Pirolli, P. & Card, S. (1999). Information foraging. Psychological Review, 106(4), 643–675. Foundational theory: proximal cues, scent-following, patch-leaving. Used for the selection-decision mechanism.
• Pirolli, P. (2005). Rational analyses of information foraging on the Web. Cognitive Science. Scent as utility assessment under uncertainty. Used for the false-open/false-skip framing.
• Chi, E.H., Pirolli, P., Chen, K. & Pitkow, J. (2001). Using information scent to model user information needs and actions on the Web. Proc. CHI 2001. Proximal-cue → distal-content prediction. Used for the preview-as-scent-cue mapping.
• 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. 41 manuscripts; visual abstracts ↑ engagement (chiefly professionals); public engagement ~zero; standard tweets most full-text click-throughs. Used for engagement ≠ routing.
• 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. J. Clinical Epidemiology, 97, 86–94. Infographic ≈ PLS > scientific abstract on knowledge; graphical form preferred. Used for preference ≠ performance.
• BMJ Evidence-Based Medicine (2024). Do infographics ‘spin’ the findings of health and medical research? 119 RCTs with non-significant primaries: spin in 33% infographics vs 26% abstracts vs 26% full texts (n.s.); results-section spin 2–4× higher in infographics. Used for compression-concentrates-distortion risk.
• RIVA-C checklist and guide (BMJ Evidence-Based Medicine, 2024). Substitution behaviour; missing harms/effects/uncertainty in most infographics. Used for the substitution constraint.
• Pragmatic evaluation of The BMJ’s visual abstracts (Information Design Journal, 2020), DOI 10.1075/idj.25.1.08sta. Novelty halo effect. Used as a caution on early engagement metrics.

Proposed experiment EXP-04: preview regimes for triage

Status: PROPOSED. Hypothesis: a card preview improves selection decisions for suitable pages and defers (UNCHANGED) for unsuitable ones. Corpus: frozen pages spanning structured technical topics plus pre-registered UNCHANGED-appropriate items (qualification-heavy argument; numerical-distinction page; title-sufficient short item; false-metaphor-risk concept), each paired with a declared reader goal. Conditions: title only / title+excerpt / text summary / 3–5 generated visual cards / cards+minimal text. A semantic-extraction step precedes image generation and is logged, so card content is auditable. Measurements, reported separately, never averaged: decision quality (correct keep/close against a human-adjudicated relevance key; false skips and false opens as separate rates, skips weighted heavier); decision speed (time to committed verdict); preference (which regime readers say they want); confidence (calibration of confidence against correctness). Failure criteria: cards win preference while losing on false skips (valuable negative result — attractiveness confound confirmed); cards underperform title+excerpt on quality (no evidence for the intervention); UNCHANGED-appropriate items show systematic false skips under cards (over-application demonstrated). Artifacts expected: page/goal packs, relevance keys, extraction logs, per-outcome results. What a positive result would not justify: any claim about comprehension, recall, or understanding — selection only, per the Ch 4/5 ownership split.