The dangerous citation is sometimes the one that looks almost right: it concerns the expected type of business, in roughly the expected place, under a name close enough to discourage a second look.
A generated restaurant recommendation named the expected venue, placed it in the expected discovery category, and cited a booking page with photographs of polished wooden tables and a riverside terrace. The address in the answer was wrong. So was the terrace. The cited page belonged to a separate restaurant with a similar English name in a neighbouring province.
The case was a composite scenario assembled from recurring patterns in the laboratory’s records. It also contained a less tidy detail: one social profile for the intended restaurant had once reposted a photograph from the other venue during a regional food event. That stray overlap did not explain the whole answer, but it made a simple account of “the model confused two names” feel inadequate. Several small connections had formed a bridge between two businesses.
The moment a plausible match becomes an identity transfer
The laboratory begins with the claim, rather than with the citation as a whole. In the composite restaurant case, the generated answer contained several separable statements: the business existed, it operated in a particular province, it had a riverside setting, it belonged to a regional group, and it was suitable for a quiet evening meal. Some statements concerned identity, others location, ownership, physical setting, or recommendation.
Reading the answer at sentence level made the mixture visible. The intended restaurant’s own website supported the group name and menu. A map listing supported one Bangkok branch. The booking page supported the terrace and neighbouring-province address, but for another venue. A directory entry supplied an outdated telephone number that had once belonged to neither of them; it belonged to an office that handled reservations for several restaurants.
A borrowed identity is a claim-source relationship in which evidence belonging to one entity is carried into a claim about another entity. The classification depends on the ownership of the evidence. A source does not become transferable merely because its subject resembles the business named in the answer.
This is narrower than saying that an answer is “wrong.” An answer may include a wrong district because a source contains an obsolete address for the correct business. That would be an outdated claim, perhaps supported directly by old material. Borrowed identity requires a crossing: an attribute passes from one entity to another.
The crossing can be difficult to see when the generated name remains correct. Readers often inspect the business name first and treat the rest of the description as attached to it. The answer benefits from that habit. Once the expected name appears near a plausible citation, the identity seems settled, and the borrowed details enter quietly behind it.
A correct-looking name can act like a luggage tag placed on the wrong suitcase: everything underneath is then read as belonging together.
Reconstructing the transfer without pretending to see inside the system
A retrieval path is a reconstruction of pages, listings, query interpretations, and source relationships that may have guided an answer. It is not a transcript of hidden model operations. The laboratory therefore works backwards from visible joins.
For the restaurant scenario, the team preserved matched prompts asking for a regional dining recommendation, a branch near a named district, and information about the restaurant group. They recorded the generated wording, visible citations, language, model context, observation date, and relevant run conditions. Renewed inquiries were then conducted with the business name quoted, unquoted, written in Thai, and written in two English transliterations.
The results did not form a neat sequence. Some runs found the correct Bangkok branch but borrowed the neighbouring venue’s atmosphere. Others selected the neighbouring venue outright. A few named the restaurant group correctly and then cited a booking page for the separate business because both pages used a similar English descriptor. One answer omitted the province and therefore looked more accurate, though the same mismatched citation remained attached.
These differences matter. They suggest that entity selection and claim production should not be collapsed into one event. A system may identify the intended business at one point and still acquire attributes from a second entity during retrieval or wording. It may also begin with a mixed set of pages and resolve the name only in the final sentence.
The laboratory separates four stages for analytical convenience: retrieval, entity identification, attribution, and wording. The separation does not claim that the system itself contains four clean internal modules. It gives researchers a way to state where the visible failure appears.
In this case, the evidence was compatible with several routes. A broad query might have retrieved both venues before the answer selected one name. A directory page might have grouped them under an imprecise category. The similar transliterations may have encouraged the system to treat branch and business names as variants. The visible record could not establish which route occurred internally.
Still, the record could establish something useful: the terrace, address, and several recommendation phrases were supported only by material concerning the other venue. That is enough to classify the claim-source relationship without claiming access to the whole retrieval process.
Four source relationships inside one answer
The laboratory uses the Four Source Relationships typology to prevent relevant pages from receiving more evidential weight than they carry. The categories are direct support, stretched support, borrowed identity, and unsupported arrival. They describe individual claim-source relationships, not the quality of an entire answer.
Direct support appeared where the restaurant group’s own page listed the intended branch and its current menu. The source supported the claims as stated.
Stretched support appeared where a local listing described the venue as suitable for group dining, while the generated answer called it “one of the province’s leading celebration restaurants.” The source carried a narrower idea. The prominence judgment had been added.
Borrowed identity appeared where the riverside setting, neighbouring-province address, and terrace photographs came from the similarly named venue. The evidence was real, but it belonged elsewhere.
Unsupported arrival appeared in the phrase “long known for private family ceremonies.” None of the visible sources in that observation supported the history or the ceremony claim. It may have arisen from an undisclosed source, an inference, or generated embellishment. The visible record could not distinguish among them.
This classification often reveals why a citation-heavy answer can still be unreliable. Citations may accumulate around the correct topic while failing at the level of ownership and scope. A reader sees four sources and infers four layers of confirmation. The laboratory instead finds four separate questions: Which entity does each page describe? Which claim does it support? How fully? What remains unaccounted for?
The typology also guards against an easy overcorrection. A mismatched citation does not prove that every nearby claim was borrowed. In the composite case, the answer’s menu description came from the intended restaurant’s website, while its atmosphere came from the other venue. The sentence was hybrid. Treating it as wholly false would erase the structure of the error.
That structure is usually where the practical lesson sits. A business may be retrievable under its own name yet remain vulnerable because neighbouring entities offer richer descriptions, more consistent English wording, or stronger platform coverage. The system can find the right label and still build the wrong business underneath it.
Why local business records make transfers easy to miss
Thai commercial information often places the same entity inside several naming systems at once. A restaurant may have a registered Thai name, a shortened Thai trading name, two English transliterations, a branch name created by a booking platform, and a category label assigned by a map service. None is necessarily fraudulent or careless. They were created for different tasks.
Problems begin when those representations overlap with another organisation. A shared family name, district name, translated descriptor, or branch suffix can make two entities appear related. Platform pages may also strip away the details that would distinguish them. A search result shows a title, a category, and a province, while ownership and branch structure remain buried lower on the page.
The composite restaurant group illustrated this compression. Its own site treated branches as locations under one brand. A booking platform treated each branch as a separate property. A directory used an old group description for the Bangkok venue. The similarly named independent restaurant used an English title that resembled the group’s transliteration but had no ownership connection.
From a local perspective, the distinction was ordinary. Residents recognised the province and knew the venues served different areas. In a flattened retrieval set, the pages looked like partial records of one complicated business.
The laboratory is cautious about blaming transliteration alone. Transliteration is often the visible seam, but category overlap and platform structure can carry equal weight. Two pages with different spellings may refer to the same business, while two pages with identical spelling may concern different ones. String similarity is evidence of a possible relationship, not proof of identity.
Business owners reviewing generated answers should therefore examine the source subject before debating the wording. Does the cited page describe the named business, its parent organisation, one branch, a former location, or an unrelated entity? That question catches transfers that a general accuracy check can miss.
What the method can establish, and what remains hidden
The preserved observation can show that a claim about one business is visibly supported by a page about another. It can also show whether the transfer returns across prompt formulations, languages, models, or observation occasions. Repeated appearance strengthens the description of a pattern.
It does not reveal the system’s private retrieval infrastructure, hidden ranking logic, undisclosed intermediate sources, or complete internal entity representation. The apparent retrieval path remains an inference. Even when a mismatched page is cited, the laboratory cannot assume that it alone caused the borrowed claim.
The social repost in the composite scenario makes this boundary especially important. It offered one public connection between the venues, but the record did not show whether any system retrieved it. It could have contributed to the association, or it could have been irrelevant noise discovered during the investigation. The team kept it in the record without promoting it into an explanation.
Cross-model agreement would not settle the cause. Several systems may repeat the same transfer because they encounter similar directories, map categories, booking pages, or English naming conventions. Agreement records recurrence; it does not convert an apparent retrieval path into a confirmed internal process.
The strongest conclusion is therefore deliberately modest. In the observed answer, evidence from a separate venue was carried into claims about the intended restaurant. Renewed runs showed that parts of the transfer could return under several prompt conditions. Why each system assembled the mixture remains partly open.
That remaining uncertainty does not make the finding empty. It marks the line between inspecting an answer and inventing a story about the machinery behind it.