Recent Posts
Imagine two students beginning the same research project. The first opens a search engine, scans a page of results, and starts making choices. One link appears to be a university archive; another is a magazine essay; a third comes from an advocacy group whose commitments are clear from its name. The student opens several tabs, notices that the sources disagree, follows a footnote, abandons a weak lead, and gradually learns what kind of question is actually being asked.
The second student types the same query into an AI assistant. A clean paragraph appears. It contains the major points, a short list of qualifications, and perhaps several citations placed at the bottom. The answer is readable and probably useful. It may even be more accurate than the first few pages the other student encounters. Yet the two students have not performed equivalent intellectual tasks. One has explored a field of claims. The other has received a finished surface.
This difference is easy to dismiss as a matter of convenience. Why insist on a longer route if a system can deliver the result? But the route through information is not mere overhead. It reveals who has spoken, which ideas depend on which evidence, where interpretations diverge, and how a conclusion was assembled. When an interface removes that route, it changes more than the time required to find an answer. It changes the meaning of finding.
AI summaries are becoming a standard entrance to online knowledge. Their usefulness is real. They can clarify technical language, provide an initial orientation, and help a reader formulate better questions. The danger begins when orientation becomes substitution: when the summary ceases to be a doorway and becomes the entire room. A culture that receives knowledge mainly as completed answers risks losing the visible architecture of thought.
The shortcut becomes the road
Search engines and AI assistants respond to the same practical desire. People want to move from uncertainty to understanding without wasting time. Traditional web search addresses that desire indirectly. It produces a ranked set of possible destinations and asks the user to continue the work. An answer engine addresses it directly. It reads or models many sources, composes a response, and presents the result as a single piece of prose.
That directness is the attraction. Many web searches are chores. A person trying to reset a device, understand a tax form, compare train schedules, or identify a plant may not want an education in the history of the subject. A concise answer can eliminate repetitive browsing, search engine optimization clutter, and pages designed primarily to display advertising. There is no virtue in friction for its own sake.
The problem is that the same interface is now used for questions that do not have the structure of a simple lookup. Political concepts, historical causes, medical evidence, literary interpretation, and scientific disputes contain unresolved judgments. A smooth paragraph can make such questions resemble settled facts. The format supplies a sense of closure before the reader has learned where the uncertainty lies.
Even citations do not fully solve this problem. A row of links beneath an answer is different from an argument in which each claim visibly emerges from a particular source. Readers need to know whether a link supports one sentence, supplies background for a whole section, contradicts another source, or merely resembles the topic. Provenance is not a decorative badge attached after synthesis. It is part of the meaning of the synthesis.
The distinction becomes clearer when an AI answer makes a mistake. A conventional webpage has an author, publication date, editorial context, and address. Those features do not guarantee truth, but they give criticism somewhere to land. A generated response may be recreated differently for the next user or altered after a model update. Its sentences can inherit information from many places without preserving the path by which any particular claim arrived. The answer has a voice, but its responsibility is dispersed.
When this mode of reading becomes habitual, users begin to expect knowledge to arrive without visible construction. The shortcut becomes the normal road, and the longer route starts to look like a defect. At that point, the issue is no longer whether a summary saves time. It is whether the surrounding information system still teaches people how conclusions are made.
Knowledge has a shape
The web was designed around a simple but demanding idea: information becomes more useful when documents can point beyond themselves. Tim Berners-Lee developed his 1989 proposal at CERN because a changing scientific organization could not be represented adequately by one fixed hierarchy. People, experiments, software, equipment, and documents formed overlapping relationships. A linked system could preserve those relationships without forcing them into a single tree.
The first page of Tim Berners-Lee’s 1989 proposal for a distributed hypertext system. Courtesy CERN.
The link was therefore more than a navigation button. It expressed a relation. One document might refer to another, depend on it, revise it, provide evidence for it, or place it in a new context. A reader could move along that relation and inspect both ends. The resulting network did not eliminate interpretation; it made interpretation visible and traversable.
The title page of Ephraim Chambers’s Cyclopaedia (1728), an early experiment in cross-referenced knowledge. Courtesy Wikipedia.
Earlier visions of hypertext placed even greater emphasis on this property. In 1945, Vannevar Bush imagined the memex, a personal system in which a researcher could connect items into associative trails and share those trails with others. Two decades later, Ted Nelson argued for forms of electronic writing that would preserve complex interconnections among texts. These proposals differed in important ways, but both treated the movement between documents as an intellectual act. A trail recorded how one thought prompted another.
Ted Nelson’s 1965 paper introducing hypertext as a form for interconnected writing. Courtesy the Internet Archive and Ted Nelson.
The commercial web implemented only part of that ambition. Most links point in one direction. Pages disappear, ownership is often obscure, and quotation rarely maintains a durable connection to its source. Still, the basic unit survived: one document openly indicates another. The reader can leave. That small freedom gives the web much of its generative power.
Google’s early search engine succeeded partly because it treated this network as information. PageRank did not examine words alone; it used patterns of linking to estimate the relative importance of pages. The system was imperfect and vulnerable to manipulation, but its central insight was sound. A document’s place in a field of references reveals something that the document cannot say about itself.
This relational information is difficult to preserve in a compact answer. A summary tends to organize material by topic: definitions first, major points next, qualifications last. A network organizes material by connection. It shows that an obscure paper influenced a popular article, that a critic challenged a foundational assumption, or that several communities use the same term differently. These are not optional details around the knowledge. They are part of its structure.
We often speak as if the content of a document could be poured into a new container without changing. Sometimes it can. A weather forecast remains useful when reduced to temperature and rain probability. But many works derive their meaning from arrangement, sequence, tone, and response. An essay may delay its conclusion because the reader needs to experience the uncertainty that precedes it. A scholarly paper may separate results from interpretation because that boundary matters. A debate makes sense through the order of claims and replies. Compression can preserve propositions while removing the form that teaches us how to evaluate them.
What gets lost inside a synthesis
Every summary selects. It decides which facts are central, which distinctions can be omitted, and which disagreements deserve space. Human writers make these decisions too, but they can explain their standpoint and cite their sources. AI systems often present selection as if no selecting intelligence were involved. The prose sounds detached from any particular place, even when its underlying materials come from people with incompatible aims.
The first loss is voice. A historian reconstructing an archive, a participant describing an event, and a journalist reporting an investigation may all contribute relevant information, but they do not occupy interchangeable positions. Their vocabulary, access, incentives, and uncertainty differ. When their statements merge into one neutral register, readers may retain the facts while losing the reason each speaker was able to make a claim.
The second loss is scale. A summary can place a small observational study beside a large review and give each one sentence. It can mention a preliminary finding and a long established conclusion in parallel bullet points. Unless the interface carries information about methods, sample size, date, and disciplinary reception, the visual equality of the statements implies an evidential equality that may not exist.
The third loss is productive disagreement. A good research process does not merely count opinions and average them. It identifies why sources diverge. They may define a term differently, study different populations, rely on competing models, or value different outcomes. A synthesis that reports a balanced range of views without reconstructing those differences can make conflict look like noise. Yet the conflict may contain the most important knowledge available.
The fourth loss is the boundary of a claim. On a source page, a reader can inspect the paragraph before and after a quotation, examine a chart, and see the conditions under which an author drew a conclusion. In a generated answer, the same conclusion may appear without its local restraints. A citation helps only if the reader follows it and can determine exactly what it supports. If the answer has already satisfied the immediate question, many readers will not make that additional trip.
None of these losses requires deliberate deception. They follow from the purpose of synthesis itself. A useful summary must reduce variation. The danger comes from treating that reduction as a transparent window onto the source material. It is closer to a map: valuable because it leaves things out, and trustworthy only when users understand its scale, purpose, and omissions.
Memory needs routes
Information systems do not merely answer present questions. They determine what can be remembered later. A note with a link preserves more than a conclusion: it preserves the occasion for the conclusion and a route back to the evidence. Months afterward, the writer can recover not only what seemed important but why it seemed important at the time.
Generated summaries are harder to place within this kind of memory. They often arrive as temporary responses to prompts, separated from the documents that gave the inquiry its direction. A user may save the answer, but the saved prose does not necessarily preserve the search path, rejected alternatives, or changes in vocabulary that made the final question possible. The product is retained while the process disappears.
At a collective scale, this difference matters even more. Archives depend on identifiable objects and durable relations among them. Scholars reconstruct debates by tracing editions, citations, correspondence, reviews, and revisions. If public knowledge increasingly circulates through personalized answers that are neither stable publications nor transparent records of derivation, future readers may inherit conclusions without a usable history of how those conclusions formed.
An AI system could help solve this problem if it treated the research session as a trail worth preserving. It could record which sources changed the direction of the inquiry, distinguish materials the user read from those the model merely consulted, and export a navigable history alongside the final synthesis. Such a record would not reproduce the user’s mind. It would do something more practical: preserve enough of the route for another person to question, continue, or revise the work.
Friction is part of understanding
Anyone who has researched a difficult subject knows the irritation of the open web. Search results repeat each other. Promising links lead to paywalls or dead pages. Terminology shifts between disciplines. A source that appears decisive turns out to cite another source that makes a narrower claim. It is tempting to view all of this as inefficiency waiting to be automated away.
Some of it should be removed. Better tools can detect duplicate reporting, expose citation chains, translate jargon, compare versions, and warn when a claim has travelled far from its origin. But uncertainty itself is not a user interface bug. The pauses and reversals of research train judgment. They force a reader to decide what counts as evidence and to notice when a question has several legitimate forms.
Consider the difference between receiving a list of reasons and constructing one. In the first case, each reason appears as a completed item. In the second, the reader encounters material that does not initially fit. One source changes the period under discussion; another introduces a causal factor that makes the original question too simple. The eventual argument is stronger because the researcher has revised the frame rather than merely filled it.
Links support this kind of revision. They allow a reader to interrupt the author’s sequence and test a dependency. A footnote can be followed before the paragraph is finished. A term can lead to a history the author did not explain. A citation can expose selective quotation. The act of leaving one page for another is a modest exercise of intellectual independence.
Answer interfaces reverse that relationship. They usually ask the user to remain inside a conversation while the system fetches, condenses, and explains. Follow up questions deepen the exchange, but they also keep the model in the role of mediator. The user can request sources, objections, or greater detail, yet each request returns another synthesis. The conversation may become more elaborate while the reader’s contact with the underlying documents remains thin.
This is why the debate cannot be reduced to accuracy. An answer engine might achieve an excellent factual record and still narrow the practice of inquiry. Education is not only the transfer of correct statements. It includes learning how to encounter incomplete evidence, trace a concept across contexts, and recognize when apparent consensus has been manufactured by the format of presentation.
The ecology beneath every answer
There is also a material problem. AI answers depend on a world of documents produced by researchers, public institutions, journalists, specialists, enthusiasts, and communities. Those documents require time, money, maintenance, and an expectation of being read. When an interface extracts their usable conclusions while reducing visits to the original pages, it changes the incentives that keep the information supply alive.
The web has never offered a stable or fair economy for publishing. Advertising rewarded scale, subscription models created access barriers, and platforms captured much of the value produced by others. AI summaries enter this already damaged environment. Their novelty is that they can replace the visit itself. A user may receive the recipe without meeting the cook, the explanation without seeing the teacher’s other work, or the investigative finding without entering the publication that funded it.
The immediate effect is easiest to measure as traffic, but the deeper effect concerns reciprocity. A link acknowledges that an answer came from somewhere and gives the reader a way to return value, whether through attention, payment, citation, correction, or participation. A summary that hides its dependencies turns this exchange into a one way intake. The source remains necessary to the system but less visible to the person benefiting from it.
Over time, this can produce a feedback problem. If fewer independent sites can sustain careful work, future systems will have a thinner body of material to retrieve or learn from. Widely repeated summaries will become easier to find than the reporting, experiments, and local knowledge from which they originated. The information environment may appear abundant while becoming less diverse underneath.
Small sources are especially vulnerable. A major institution may retain direct audiences, licensing power, and brand recognition. A technical blogger, regional newspaper, patient community, or volunteer archive often depends on discovery through links. These sources may contain precisely the unusual observation that improves a synthesis. If answer systems reward only material already prominent enough to be recognized, they will reproduce a narrower world.
The question, then, is not whether technology companies can attach citations to generated text. It is whether the interface preserves a functioning public relationship between answers and the people who make knowledge available. Attribution must be legible, specific, and useful enough to send readers outward. Compensation and licensing matter, but no private contract can replace the reader’s ability to inspect and participate in the network.
Interfaces can preserve intellectual routes
AI assistance does not have to erase the web’s connections. The design choice is not between an unfiltered list of blue links and a sealed paragraph. Answer systems could use synthesis to expose relationships rather than conceal them.
The first requirement is claim level provenance. A reader should be able to see which source supports each factual statement and whether that source is primary reporting, analysis, opinion, or a later repetition. If several sources support a claim, the interface should show their independence or shared dependence. Ten articles repeating one press release do not amount to ten separate confirmations.
The second requirement is visible disagreement. When credible sources reach different conclusions, the answer should not smooth them into a vague sentence about debate. It should identify the point of divergence and give the reader direct paths into each argument. Uncertainty needs structure. A map of the dispute is often more valuable than a compromise sentence.
The third requirement is an exploratory mode that changes the unit of presentation. Instead of producing a longer summary when the user asks for detail, the system could display a source landscape: foundational documents, recent developments, major critiques, and unresolved questions. The reader could then choose a route. AI would still reduce the cost of orientation, but it would return agency before the inquiry was complete.
The fourth requirement is durable citation. Generated answers used in research or public discussion should have stable versions that can be revisited. Readers need to know when an answer was produced, which model and retrieval process shaped it, and whether later edits changed its claims. Without a durable object, correction becomes difficult and citation becomes ceremonial.
The fifth requirement is meaningful exit. Links should appear where curiosity is strongest, not hidden in a collapsed panel after the conclusion. A source preview can show the relevant passage in context while inviting the reader to open the full document. Good design would treat leaving the AI interface as successful use rather than lost engagement.
These features would add complexity. That is appropriate. Questions differ in their epistemic demands, and an interface should signal that difference. A query about a unit conversion may deserve one clean line. A question about the causes of a war, the safety of a treatment, or the interpretation of a law should make its sources and uncertainty impossible to mistake for decoration.
A better division of labor
The strongest case for AI in research is not that it can finish thinking on our behalf. It is that it can help us enter bodies of knowledge that would otherwise be difficult to approach. It can suggest vocabulary, reveal that a question spans several disciplines, compare definitions, translate passages, or locate a likely primary source. These are acts of guidance.
Used this way, an AI system resembles a skilled librarian more than an oracle. A librarian does not prove usefulness by preventing readers from touching books. The work lies in understanding the request, identifying promising materials, explaining how collections are organized, and helping the reader refine the search. The destination remains a meeting between reader and source.
AI can also make the link more expressive. It can explain why a source is connected, identify the claim that travels across documents, and visualize a chain of citation. It can warn when a widely shared statistic lacks an identifiable origin or when a modern article relies on an outdated study. Such tools would extend the web’s relational intelligence rather than replacing it with a single anonymous voice.
This division of labor accepts that synthesis is valuable while refusing to confuse it with knowledge as a whole. The machine is well suited to scanning, grouping, translating, and proposing connections. The reader remains responsible for deciding what those connections mean. The source remains available to resist both of them.
Designers often describe frictionless access as an unquestioned good. For routine transactions, that aim makes sense. For inquiry, the better goal is selective friction: remove obstacles that do not teach, preserve encounters that do. Repeated advertisements, inaccessible formats, and needless jargon can go. Contradiction, provenance, context, and the possibility of surprise must remain.
Keeping the question open
The web’s great achievement was not simply that it placed enormous quantities of text within reach. Libraries had long gathered texts, and broadcast systems had long distributed information widely. The web allowed documents to form a public system of references that readers could traverse for themselves. It made connection operational.
AI summaries offer another remarkable capability: they can transform a dispersed body of language into an immediate response. We should use that capability. But immediacy has a cost when it hides the relations that made the response possible. An answer detached from its route can be accurate yet intellectually incomplete.
The most important design question for the next generation of information tools is therefore not how much material a model can compress. It is how well the system can help a reader move between overview and origin, agreement and dispute, explanation and evidence. A useful answer should reduce confusion without pretending that the world has become simple.
We will know that these tools serve inquiry when they do more than end searches. They should create informed departures: moments when a reader understands enough to choose the next source, challenge the current frame, or follow an unexpected connection. Intelligence in an information system is not measured only by the speed with which it closes a question. It is also measured by the quality of the paths it leaves open.
Sources and further reading
Tim Berners-Lee, Information Management A Proposal, CERN, 1989
https://www.w3.org/History/1989/proposal.html
Vannevar Bush, As We May Think, The Atlantic, July 1945
https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881
Sergey Brin and Lawrence Page, The Anatomy of a Large-Scale Hypertextual Web Search Engine, 1998
https://research.google/pubs/the-anatomy-of-a-large-scale-hypertextual-web-search-engine
