Lexicon

Why this lexicon exists

Most disagreements about media technology are disagreements about words. “AI” covers a dictation tool and a text generator. “Personalisation” covers a newsletter topic checkbox and an inferred behavioural profile. “Verification” covers a phone call and a probability score from a detector. When a discussion uses one word for several mechanisms, the participants can agree completely and still mean different things.

The entries below define each term as narrowly as the subject allows, and where a term is genuinely contested the entry says so rather than picking a side. Definitions are written to be quotable: each is self-contained, avoids referring to other entries by number, and states the distinction that makes the term worth having. Terms are listed alphabetically and tagged by the area of practice they belong to.

Defined terms

Assisted writingProduction

Any use of a machine tool inside the drafting or editing of an item, from spelling checks to generated prose. The term is only useful when qualified by the artefact produced, because the verification burden differs completely between a transcript and a generated paragraph.

Attention timeMetrics

The period during which an item was active and in view, as measured by an analytics tool applying its own idle threshold. Better evidence of reading than a pageview, but still unable to distinguish absorption from confusion, and unfair to formats that succeed by being brief.

Automated newsProduction

Items generated by filling human-written templates from a structured data feed, according to rules that decide what is newsworthy. Distinct from generative drafting: an automated item cannot contain a sentence that nobody wrote, but a faulty rule affects every item in a run.

Batch identifierProduction

A label attached to every item produced in one run of an automated pipeline. It allows a single correction to be applied consistently across the whole run and lets a desk withdraw an entire batch in one action rather than editing items individually.

Candidate generationDistribution

The first stage of a ranking system, which reduces a large catalogue to a shortlist using cheap criteria such as recency, section or topic similarity. Items excluded here never reach the scoring stage, so the stage shapes exposure more than its simplicity suggests.

Chain of custodyVerification

The documented record of how a file passed from capture to the desk: who supplied it, when, in what form, and what was asked and answered. The record travels with the asset so that qualifications attached at one desk are not lost when it moves to another.

ChronolocationVerification

Establishing when material was captured, using shadow direction and length, weather records, visible clocks, seasonal vegetation and event markers. Rarely precise, but frequently sufficient to eliminate a claimed time, which is usually what a verification desk needs.

Cold startDistribution

The condition of a newly published item, a new section or a new reader for which a behavioural ranking system has no history. Without a deliberate exploration budget or an editorial slot, cold-start items are systematically under-exposed.

Consent ModeRules

A signalling arrangement in which a page declares default states for storage categories such as advertising and analytics, and may update them after a user chooses. On this site all advertising signals are set to denied permanently and no advertising or measurement tag is installed.

Content credentialVerification

A signed provenance record attached to a file by supporting capture devices or editing software, describing origin and subsequent edits. Strong evidence where present and intact; neutral where absent, because most genuine material carries no credential at all.

Correction noteStandards

The visible statement accompanying an amended item, recording what was wrong, what it now says, when the change was made and, where it matters, the superseded wording. A silent edit is not a correction, and deletion removes the record without removing the error.

DeepfakeVerification

Synthetic audio, image or video depicting a real person saying or doing something that did not occur. A subset of synthetic media, distinguished by the depiction of identifiable people, which is why it attracts specific transparency duties under European rules.

Diffusion modelProduction

A class of generative system that produces an image by repeatedly removing noise from a random field under the guidance of a learned representation, usually conditioned on a text description. Understanding the mechanism explains where its characteristic errors appear.

Disclosure thresholdRules

The published line a publisher draws between machine assistance that is disclosed to readers and assistance that is not. Thresholds usually turn on whether machine-produced material reaches the reader substantially as generated. Publishing the threshold matters more than the exact position of the line.

Editorial floorDistribution

A fixed set of items shown identically to every reader before any adaptive module appears. It preserves the civic function of a shared front page while allowing relevance below it, and its size and position can be published as a commitment.

Exposure diversityMetrics

The variety of material a reader actually encounters, measured by counts such as distinct sections seen per week, the share of readers reached by public-interest coverage, and the overlap between what two random readers were shown. Measurable without collecting additional personal data.

Fabricated outputProduction

Confident, plausible and unsupported material produced by a generative system: an invented statistic, a citation to a document that does not exist, a quotation attributed to someone who never said it. Dangerous precisely because it reads like competent journalism.

Filter bubbleAudience

The claim that algorithmic selection seals readers into homogeneous information environments. The strong version is weakly supported by research; the narrower effects — amplified existing preference and reduced incidental exposure — are better evidenced and can be measured on a publisher’s own site.

GeolocationVerification

Establishing where material was captured by matching landmarks, signage, road markings, vegetation and skyline against mapping and street-level imagery. Slower than running a detector and considerably more conclusive, provided the frame contains identifiable features.

Hard wallSustainability

An access arrangement in which almost all material is closed to non-supporters. It concentrates editorial effort on depth for a defined audience and raises a public-value question that many publishers answer by keeping safety, election and health coverage permanently open.

Impression concentrationMetrics

The share of total impressions absorbed by the most-shown items. If the top five items take most of the impressions, the effective front page is five items long regardless of how the site is designed. Read alongside catalogue reach.

Large language modelProduction

A statistical system trained to predict continuations of text, which produces fluent prose without any representation of whether a statement is true. Fluency and accuracy are independent properties of its output, which is the central editorial fact about it.

Liar’s dividendVerification

The advantage that the existence of convincing fakes gives to anyone wishing to deny authentic material. Reduced by documenting provenance for your own output as a matter of routine, so that a chain of custody can be produced when authenticity is challenged.

Live blogCraft

A continuously published, continuously amended account of an unfolding event. Fast, and structurally hazardous, because early entries remain visible and quotable after being overtaken. Requires timestamps, marked supersessions, a current-state summary and a settled article at the end.

Metered accessSustainability

An arrangement allowing a set number of items in a period before access closes. It creates a specific editorial pressure towards volume, because the meter must be filled, and it can be implemented using essential storage without any advertising identifier.

ProvenanceVerification

The documented origin of a file and the record of what happened to it since capture. The most reliable basis for verification, and the first thing destroyed when material is re-encoded and stripped of metadata during ordinary sharing.

PromptProduction

The instruction given to a generative system. Prompts encode framing, tone and default assumptions, which makes writing one an editorial act; desks that take assistance seriously store prompts alongside drafts in the way they store interview notes.

Recommender systemDistribution

Software that selects and orders items for a reader according to a measurable objective. An editorial instrument with no editor in the room: the choice of objective is the whole editorial argument compressed into one line of configuration.

Re-rankingDistribution

The final stage of a ranking system, applying explicit rules on top of learned scores: deduplication, section caps, reserved editorial slots, legal holds. The layer where almost every realistic editorial commitment about ranking can actually be implemented.

Registration wallSustainability

An arrangement requiring an account rather than a payment. It builds a direct channel and moves the publisher from handling largely anonymous traffic to holding identifiable records tied to reading behaviour, with the data-protection duties that follow.

Returning readersMetrics

The count of people who come back within a defined period. Harder to inflate than pageviews and slower-moving, which makes it a better indicator of the health of an audience than any monthly total, though it under-counts readers who clear their storage.

Scroll depthMetrics

How far down a page a reader travelled. Useful for diagnosing layout problems and misleading as a measure of value, because a long page can score well while being abandoned in the middle and a short one cannot score highly at all.

Synthetic mediaVerification

Material generated or substantially altered by a model. The editorially useful test is not whether a file was touched by software but whether the alteration changes what a reasonable viewer would understand it to record.

Vertical videoCraft

Short, portrait-orientation video designed for handheld viewing in a feed, usually watched once and frequently without sound. Its hard editorial limit is qualification: a clip short enough to hold attention may not be able to carry the condition a claim requires.

How these terms are used on this site

Every page uses these definitions consistently. Where a page needs a narrower sense of a term — for instance, “assistance” meaning specifically generated prose rather than the whole spectrum — it says so in the sentence rather than relying on context. Where a term appears in a legal instrument with a defined meaning that differs from ordinary usage, the page uses the instrument’s meaning and notes that it is doing so.

Three words are deliberately avoided across the site. “Hallucination” is not used as a technical term, because it implies a perceptual failure in a system that has no perception; the entry above uses “fabricated output” instead. “Algorithm” is avoided as a standalone noun, because it collapses candidate generation, scoring and re-ranking into a single opaque object. And “engagement” is used only with a stated definition attached, since it otherwise means whichever metric the speaker prefers.

Suggestions for missing terms are welcome through the contact page. The lexicon is intended to stay small enough to be maintained: a definition that is never reviewed becomes wrong quietly, and a wrong definition is worse than a missing one.

Questions about the lexicon