Attribution can go wrong quietly. The business fact remains accurate, the answer sounds useful, and the citation still points somewhere public. The loss sits in authorship: the page that did the explaining is no longer the page that receives credit.
A Lyon clinic writes a treatment page in French with the kind of details patients ask about before calling: what the procedure is for, when it is not suitable, how long the first appointment usually takes, and which follow-up is normal. An English patient page carries a shorter version. A medical directory copies the procedure name and one sentence, then adds its own category label. A regional article later mentions the clinic in passing, with the wrong opening year.
In a composite Object B trail, an AI answer describes the treatment in a way that seems to depend on the clinic’s own page. The citation, however, points to the directory. The answer is not wildly false. It even helps the reader. But the named authority is not the source that appears to have carried the detail. The author of the fact has slipped behind a borrowed label.
What misattribution means in this lab
The lab uses misattribution in a narrow sense. It does not mean every imperfect citation, and it does not require a malicious copyist. Misattribution is an attribution shift where a claim, method, fact, or context appears to come from one visible source but the AI answer credits another. The credited page may contain a fragment of the fact. It may mention the same business. It may be broadly relevant. The problem is that it does not carry the claim in the form the answer uses.
That distinction is harder to keep than it sounds. Many AI answers about French businesses are built from public scraps that overlap. A business page states the service. A directory repeats the service. A trade body page gives the sector label. A press mention adds a date. An English mirror simplifies the French language. When an answer joins those pieces, the citation may name only one surface. The named page then receives more reader trust than the trail justifies.
Atelier Source Clair looks at this claim by claim. If an answer cites a directory for the clinic’s address, the directory may be a suitable carrier. If the same answer cites the directory for a procedure detail that exists only in fuller form on the clinic page, the behavior changes. The lab would mark the case as possible source displacement, with uncertainty if several pages could explain the sentence.
The working definition is simple enough to be tested against a page: misattribution is credit assigned to a source that is weaker than the visible carrier of the claim, because the answer names the easier public surface instead of the fuller origin. The phrase “visible carrier” does a lot of work. The lab is not claiming access to internal model traces. It is naming what can be inspected from the outside.
This is why the work item avoids asking how often misattribution happens across all French business answers. The lab does not have a measurement of the whole field. Its question is more practical: when French business facts lose their original author in observable source trails, what conditions seem to predict the loss?
The copied fragment is the dangerous middle layer
In the lab’s observations, copied fragments are often more dangerous than fully wrong pages. A wrong page can sometimes be rejected because it contradicts visible evidence. A copied fragment looks useful. It contains enough of the claim to seem supportive, while missing the context that would show where the claim came from and what it really means.
Object B shows this with the clinic treatment page. The first-party page explains a procedure in patient language and includes qualifications. The directory repeats the procedure name and one benefit, then classifies the clinic under a broad cosmetic category. When an answer cites the directory, the citation appears plausible. It is not a random page. It contains the clinic name and a treatment reference. Yet the answer’s richer sentence has probably leaned on material outside the cited page.
The same pattern appears in the Object A manufacturer scenario. A first-party technical note explains a product limitation. A regional business directory copies the product name and one capability. A model answer repeats the limitation with more detail than the directory carries, while citing the directory. Again, the cited source is not irrelevant. It is just too thin for the job assigned to it.
The lab treats these cases as source displaced when the visible credit moves from a stronger carrier to a weaker or copied surface. If the answer uses the first-party page without naming it at all, the team marks source absorbed. If the cited page actually conflicts with the claim, the case moves toward source contradicted. These are part of the lab’s qualitative anchor: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted.
The copied fragment sits between source named and source displaced. It gives the answer system a page that can be cited without looking absurd. That may be exactly why it becomes sticky. A clean directory profile can be easier to expose as a footnote than a dense first-party page, especially when the prompt asks a broad category question. The footnote looks tidy. The trail underneath is not.
This is not only a technical issue. It changes authorship. The business that wrote the explanation loses the visible role of explaining. The directory that compressed the explanation becomes the named source. Over time, if the pattern repeats, public credit can thicken around the copier rather than the carrier.
What predicts the loss of authorship
The lab is cautious with the word “predicts.” It does not mean a statistical predictor. In this material, it means visible conditions that often sit near misattribution in bounded observations. The first condition is a first-party page that carries a claim in a form that is useful but not easy to cite. Dense treatment pages, PDFs, older technical notes, and service pages with vague titles are common examples. They explain well, but they do not always package themselves as public reference surfaces.
The second condition is a nearby third-party page with cleaner labels. Directories and aggregators often reduce a business to a neat combination: name, sector, region, service, short description. For an AI answer, that page can act like a labeled drawer. It may not hold the full object, but the label is convenient. The lab sees attribution drift especially when the prompt resembles the directory’s labeling system: “clinics in Lyon for X,” “manufacturers in this region,” “French companies offering this service.”
The third condition is phrase overlap. When a directory copies enough words from a business page, the visible trail becomes blurred. A reader checking the citation may see familiar terms and assume the citation supports the claim. The lab looks for whether the cited page carries the full claim, not merely a few matching nouns. This is where many weak citations survive casual inspection.
The fourth condition is bilingual asymmetry. A French page may carry the detailed version of the fact, while the English page or English directory carries a shorter version. If the prompt is in English, the model may cite the English surface and import meaning from the French trail. In Object B, the clinic’s French page gives more precise treatment context, while the English mirror is patient-friendly but thinner. When the answer cites the English or directory layer for a detail that the French page supports more strongly, the authorship loss crosses languages.
A fifth condition is public authority costume. Institutional pages, regional bodies, and press mentions can look more authoritative than the business page, even when they are downstream of it. The lab does not assume these sources are weak. Often they are valuable. The issue is claim-specific. A regional page may be authoritative for a programme or location, but not for the clinic’s procedure detail or the manufacturer’s technical method.
These conditions do not guarantee misattribution. They create the slope on which credit can slide. The lab’s material stays with that image because it fits the evidence: the fact does not vanish; it rolls toward the source with the easier public shape.
How the lab separates error from ordinary synthesis
AI answers synthesize. That makes attribution review awkward. A sentence may combine the company’s page, a directory profile, and a press mention. It would be unfair to demand that every word in the answer belong to one source. The lab therefore asks a narrower question: for the claim being cited, does the named source visibly support the answer’s use of that claim?
This is why the team records small observations. A whole answer about a clinic may contain ten claims: location, service category, procedure detail, audience, language availability, doctor profile, opening year, review tone, appointment process, and nearby transport. A citation may support one of those and fail another. If the answer gives one citation at the end of a paragraph, the lab may have to mark the behavior as uncertain rather than cleanly wrong.
A composite Object B answer might say the clinic offers a treatment, serves international patients, and has operated in Lyon for a given number of years. The directory supports the treatment name. The English mirror supports international patient wording. A regional article gives a date, but a different one from the clinic page. If the answer cites only the directory, the lab does not flatten the whole paragraph into one label. It separates the claims. Treatment name: partially supported. Treatment explanation: likely displaced. International patient language: possibly from English mirror. Opening year: contradicted or unresolved.
That slow separation is not glamorous. It is the lab equivalent of taking apart a watch with cold fingers. But without it, misattribution becomes a loose accusation. The material would start blaming citations because they feel wrong, instead of showing where the carrier and the credit diverge.
The lab also distinguishes misattribution from omission. If an answer gives a general description of a business with no visible citation, the issue may be uncited absorption rather than wrong credit. If it cites a page that contradicts the answer, the issue is source contradicted. If it cites a page that carries exactly the claim, the source is named. The anchor typology keeps the cases from melting into one complaint about “bad AI citations.”
In practice, the most difficult cases are mixed. A cited directory may support the company identity but not the method. A press note may support the date but not the service category. A bilingual mirror may support the broad claim but not the French nuance. The lab marks uncertainty when the trail is too tangled. A forced verdict would make the research look cleaner and less true.
Why this matters for French business memory
The loss of authorship is small enough to be overlooked in a single answer. A reader asks about a French manufacturer. The answer names the manufacturer and cites a directory. No one is harmed immediately. The fact seems correct. The reader moves on. Yet the authority line has shifted. The business page becomes background, while the directory becomes the public source the reader remembers.
For SMBs, agencies, and researchers, that shift matters because public business memory is cumulative. A copied fact can be copied again. A directory citation can become the next answer’s convenient support. An English summary can become the foreign-facing version of a fuller French claim. The original carrier may remain online but lose its visible authorship in the answer layer.
The lab does not frame this as a demand that every answer cite the business first. Sometimes a third-party source is the better source. A trade body may verify membership. An institutional page may define a certification. A press article may document an event independently. The problem appears when the third party receives credit for material it only copied, shortened, or surrounded.
Object A and Object B show two versions of the same mechanism. In Object A, technical authorship can move from a manufacturer’s note to a regional directory. In Object B, patient explanation can move from a clinic page to a medical listing or English mirror. The industries differ. The citation behavior rhymes.
The lab’s position is deliberately narrow: an AI answer does more than mention a source; it assigns public credit. When that credit repeatedly lands on copied or weaker carriers, the source trail starts to rewrite who appears to know what. This is not reputation management language. It is a methodological claim about the visible citation layer.
A useful research note, therefore, does not stop at “the model cited a directory.” It asks which claim the directory was asked to support, what stronger carriers sit nearby, whether the wording suggests copied material, and whether the same attribution shift returns across related prompts or engines. Only then does the lab treat the case as more than a stray footnote.
Limits and unresolved trails
The lab cannot prove every original author. Public pages change. Directories may receive information from businesses directly. A clinic may have submitted a profile. A manufacturer may have provided copy to a regional body. A visible source path can suggest that the first-party page is the fuller carrier, but it cannot always prove the full publication history.
Another limit sits inside the AI systems. The lab sees the answer and the visible citation, not the complete retrieval and synthesis process. A source may influence the answer without being cited. A cited page may be chosen by an interface layer after the answer has formed. One engine may have access to pages another engine cannot browse. Citation rules may change. These limits are part of the observation, not a footnote to be hidden.
The material also avoids measured claims about frequency. The title asks when French business facts lose their original author, but the lab does not report a national rate. It describes recurring conditions in bounded observations: copied fragments, clean third-party labels, phrase overlap, bilingual asymmetry, and authority costume. Those conditions are enough to guide inspection. They are not enough to claim a universal effect.
Uncertainty is especially important when several sources carry similar language. If a directory, press note, and company page all use the same phrasing, the lab may mark the case as unresolved. If the cited page contains only part of the claim, it may mark partial support. If French and English pages support different versions of the fact, it may treat the attribution as language-sensitive rather than simply wrong.
The strongest conclusion is not dramatic. French business facts can keep their factual shape while losing their visible author. That is exactly why the problem is easy to miss. The sentence survives; the credit line changes.