Recent Posts
Before online knowledge could be searched, shared or understood by machines, it first had to acquire a language for describing itself
Every object on the internet has a second existence.There is the thing we encounter: an essay, a photograph, a map, a recording, a dataset. Then there is the usually invisible description that accompanies it. Somewhere behind the screen may be a title, a creator’s name, a publication date, a language, a subject, a file type and an identifier. These details are not the content itself. Yet they often determine whether the content can be found at all.
This descriptive layer is called metadata. The term can sound bureaucratic, as if it belonged mainly to database administrators and archivists. In practice, metadata shapes almost every journey through digital information. It helps a library catalogue distinguish two books with the same title. It tells a podcast application which image and episode description to display. It allows a search engine to decide whether a page is relevant to a question. It gives an archive a way to preserve not only a file, but also its origin and meaning.
Without metadata, the web would still contain information, but much of that information would resemble books dumped into a warehouse without covers, shelf marks or a catalogue. The objects would exist. Their relationships would not.
The need for a common descriptive language became urgent during the web’s early expansion. The answer that emerged was neither an advanced search engine nor an elaborate piece of software. It was a short vocabulary created by people willing to argue about ordinary words.
That vocabulary became Dublin Core.
The crisis was abundance
The early web is often remembered as a sparse digital frontier: grey backgrounds, blue hyperlinks and pages assembled by hand. Yet to the people trying to organise it, the web already looked dangerously large.
Once graphical browsers made navigating online spaces easier, universities, publishers, researchers and individuals began placing documents on servers at an accelerating rate. A person could publish material without passing through a library, newspaper, broadcaster or commercial distributor. This freedom was revolutionary, but it separated publication from cataloguing.
A printed book normally entered an established descriptive system. Its title page named its author and publisher. Its physical form revealed something about its type. A library might assign it subject headings, a classification number and a carefully structured catalogue record.
An online file might provide none of these signals. Its filename could be cryptic. Its location might change. A text could be copied, revised or separated from its original context. Early discovery services were capable of locating servers and filenames, but they could not reliably answer conceptual questions about authorship, subject or relationships between works.
The problem was not simply that searching technology was immature. The material itself often had no standard way to introduce itself.
A page needed the digital equivalent of a label on a museum object or a catalogue card in a library. It needed to say, in a form both people and computers could interpret: this is what I am; this is what I am about; this is who made me; this is when and where I belong.
But whose descriptive system should the web adopt?
Libraries already possessed highly developed cataloguing traditions. These systems could express subtle distinctions between editions, contributors, formats and publication histories. Their precision had been earned over generations. They also required specialist knowledge and considerable labour.
The web presented a different environment. Its descriptive system would have to be usable by a professor uploading a paper, a government department releasing a report, a photographer publishing an image and an amateur building a personal page. If the standard demanded professional cataloguing expertise, most creators would ignore it. If it were too vague, machines would gain little from it.
The task was to discover the smallest useful vocabulary between those extremes.
A meeting about nouns
In March 1995, 52 people gathered at the headquarters of the Online Computer Library Center in Dublin, Ohio. The workshop was jointly hosted by OCLC and the National Center for Supercomputing Applications.
The participants represented fields that did not ordinarily share a professional language. There were librarians, computer scientists, publishers, indexing specialists, museum and archive professionals, and experts in imaging and geographic information. They agreed that networked resources needed better description. Agreement became more difficult when they tried to decide what that description should contain.
Words that seem simple in everyday speech become unstable when turned into technical standards.
Consider “author”. It works naturally for a novel or an essay. But who is the author of a map, a photograph, a collaborative database or a digitised performance? Should the organisation that published a resource be treated differently from the person who created its intellectual content? What about an editor, translator, illustrator or transcriber?
A technical system cannot settle these questions through intuition. Each term must work across different media and professional communities. It must remain understandable when a document moves between institutions, software systems and countries.
The workshop’s participants were therefore designing more than database fields. They were negotiating a small philosophy of information.
The first meeting produced 13 descriptive elements. They covered ideas such as title, subject, author, publisher, date, language, format, source, relation, identifier and geographic or temporal coverage. Further workshops revised the vocabulary, eventually producing the familiar set of 15 Dublin Core elements.
What mattered was not only which words survived. The system was built around a distinctive set of principles. Its elements were optional, because not every resource had every property. They could be repeated, because a work might have several creators or subjects. They were intended to remain independent of a single technical syntax. They could also be refined when a community required greater precision.
This was a modest design, but its modesty was strategic.
The intelligence of incompleteness
Standards often fail because their designers try to anticipate every possible situation. Each unusual case produces a new category; every category produces qualifications and exceptions. The resulting system may be intellectually impressive while becoming too difficult for ordinary use.
Dublin Core accepted incompleteness.
Its creators did not attempt to replace professional library cataloguing. Instead, they produced a common floor: enough description to support discovery and exchange, but not enough to capture every distinction that a specialist might want.
A basic record could answer a handful of practical questions:
What is this resource called?
Who created it?
What is it about?
When was it produced?
In what language and format does it exist?
How is it related to other resources?
Who controls its publication or rights?
These questions apply to a surprising range of objects. A poem and a geological map are profoundly different, but both can have a title, a creator, a date, a subject and an identifier. By concentrating on properties shared across media, Dublin Core enabled independently developed systems to exchange at least a minimal amount of meaning.
This is the paradox of a successful standard: its power can come from what it refuses to specify.
A highly detailed vocabulary may describe one community’s objects beautifully while remaining unintelligible elsewhere. A smaller vocabulary loses detail, but it can travel. It provides a bridge between richer local systems.
The designers also recognised that description should not be tied permanently to the technology of 1995. The meaning of “creator” or “subject” should survive even if the method used to encode it changed. Dublin Core could appear in HTML metadata, database records, XML documents or later in Resource Description Framework statements.
Separating meaning from syntax helped the vocabulary outlive the particular tools that surrounded its birth.
Every label contains a judgment
Metadata is sometimes defined as “data about data”, but the phrase can conceal the human choices involved. Describing an object is never merely mechanical.
Someone must decide which name is authoritative, which date matters and which subjects deserve emphasis. A photograph of a public demonstration could be classified under the place where it occurred, the political issue involved, the people represented, the photographer’s career or the history of visual journalism. Each description makes some paths to the object easier and others more difficult.
The distinction between creator and contributor can assign prestige. A language label can help one audience find a resource while making multilingual complexity disappear. Geographic coverage can centre modern political borders even when the material concerns communities that understand territory differently. A rights field can preserve vital legal information, but it can also expose how frequently the conditions governing cultural material are uncertain.
Metadata does not simply report a world that has already been organised. It participates in organising that world.
This makes the diversity of the Dublin workshop particularly significant. Librarians brought experience in durable description and public access. Computer scientists understood what automated systems could process. Publishers knew how works moved through production and distribution. Specialists in maps, images, museums and archives brought objects that resisted assumptions based on conventional books.
Their disagreements revealed the boundaries of their respective worlds. Consensus did not eliminate those differences; it created a vocabulary that could function despite them.
The history of Dublin Core therefore challenges a familiar mythology of technological progress. Important infrastructures are not always invented by a lone engineer having a brilliant insight. Sometimes they emerge from meetings, draft documents and compromises over terminology.
The dramatic achievement is not a machine. It is the agreement that allows many machines to cooperate.
The description becomes infrastructure
Dublin Core’s original elements were later formalised and standardised. The present core includes contributor, coverage, creator, date, description, format, identifier, language, publisher, relation, rights, source, subject, title and type.
None of these terms seems especially futuristic. That is partly why they have endured.
The larger family of Dublin Core terms has evolved beyond the original list, supporting more precise relationships and linked-data applications. A resource can be connected to another resource of which it is a version, part, replacement or source. Instead of treating metadata as a flat label attached to an isolated object, systems can use it to describe a network of entities and relationships.
This development reflects an important change in how machines handle meaning.
A simple metadata record resembles a card describing an object. Linked data resembles a set of statements: this person created this work; this work belongs to this collection; this digital file represents this physical object; this edition replaces an earlier edition.
Such relationships help build catalogues, repositories and knowledge graphs. They also allow information created in one institution to be understood by another without requiring both institutions to use identical internal databases.
The contemporary web uses many additional descriptive systems. Social platforms rely on metadata to generate preview cards. Search services interpret structured data about articles, products, events and organisations. Digital publishing formats store information about titles, languages, identifiers and rights. The specific vocabularies vary, but the underlying ambition remains familiar: give machines a structured account of what a resource is.
Metadata has become so ordinary that users rarely notice it—until it fails.
A missing preview image, an incorrect author attribution or a meaningless search result exposes the hidden descriptive machinery. The visible page may be perfectly intact, yet its digital identity has broken.
From common descriptions to private predictions
The current information environment differs sharply from the web for which Dublin Core was designed.
The early model assumed that publishers would describe resources and that different services could use those descriptions. The structure was imperfect, but it encouraged interoperability: a shared vocabulary allowed information to move between organisations.
Large platforms increasingly organise content through proprietary systems. Rather than relying only on explicit metadata, they infer categories from behaviour, text, images and networks of users. A platform may know how long someone paused over a video, which posts they ignored and what purchases followed a search. These behavioural descriptions are often more valuable to the platform than traditional fields such as subject or creator.
This is metadata too, but it operates differently.
Open descriptive metadata tells multiple systems something about a resource. Proprietary behavioural metadata tells one company something about a user. The first helps objects travel. The second helps platforms predict attention.
Generative artificial intelligence introduces another change. An AI interface can produce an answer assembled from many documents while hiding the boundaries between them. The user receives fluent prose but may not see a stable title, creator, date or source for each contributing work. Information becomes easy to consume while its provenance becomes difficult to inspect.
That trade can appear convenient. It is also dangerous.
Knowledge depends on more than a plausible statement. Readers need to know who made a claim, in what context, using which evidence and at what time. These are, fundamentally, metadata questions. When an interface removes them, it does not make information neutral. It makes the process of selection harder to examine.
The coming challenge is therefore not merely to produce more powerful models. It is to build systems capable of carrying context forward.
The next vocabulary
The web now needs descriptive agreements for problems that were marginal or nonexistent in 1995.
How should a synthetic image identify the model and prompts involved in its creation? How can a piece of writing distinguish human authorship, machine assistance and automated generation? How should systems represent contested provenance? What metadata should accompany a dataset used to train a model? How can a summary preserve links to the works from which it was derived?
Technical solutions are emerging, but the history of Dublin Core suggests that syntax alone will not be enough. The difficult part will be establishing shared meanings and incentives.
A company may be able to create a proprietary label quickly. A durable public standard requires negotiation among people who want different things: technologists, artists, publishers, researchers, archivists, educators, regulators and users. The process is slow because its purpose is not merely efficiency. Its purpose is legitimacy.
The Dublin Core workshop succeeded by treating description as common infrastructure. No participant could know precisely what the web would become, so the group created something broad enough to adapt. Their vocabulary was neither complete nor perfect. It was usable, extensible and public.
Those qualities remain instructive.
We often imagine the internet as a universe made from documents and links. Beneath them lies another structure built from names, categories and relationships. It is this second structure that enables a file to become a source, a source to become part of a collection, and a collection to become available to strangers.
Content may be what we go online to encounter. Description is what allows the encounter to happen.
The future of digital knowledge will depend on whether that descriptive layer remains a shared language—or becomes a collection of private judgments made by machines we cannot question.
