Align Dutch French and English evidence

Before this lecture, you should know how to build factual pages from lecture 4, write liftable statements from lecture 5, keep evidence current from lecture 6, and check entity clarity from lecture 7. We are now using those habits across language versions, where the same firm can look orderly in one language and strangely elastic in another.

Three browser tabs are open: Dutch, French and English. In the Dutch tab, a Brussels practice says it helps families with gezinshereniging. In the French tab, the same service is framed as accompagnement à la mobilité. In English, the page says “relocation support for international families,” with a careful paragraph lower down about residence rights. A lawyer reading the pages sees the intended overlap. ChatGPT does not owe us that generosity.

This is a composite Object B scenario, and I have chosen it because the mistake is small enough to miss on a busy afternoon. The practice is real in shape, not in name: Brussels-based, cross-border, multilingual, with clients who move between Belgian procedures and EU-facing questions. In one Dutch prompt, ChatGPT places the firm near family reunification law. In a French prompt, it drifts toward relocation advice. In English, it mentions immigration law but softens the firm into a general mobility helper. Nobody lied. The public record simply failed to say the same hard facts in three languages.

One firm can become three evidence trails

A multilingual firm often thinks it has one website in three languages. ChatGPT may encounter it more like three related evidence trails. The Dutch service page, the French profile, the English explainer, the lawyer biography and a public listing may each carry a slightly different version of the firm. When those versions agree on name, place, category and service boundary, the firm becomes easier to place. When they disagree, the model may answer as if each language has met a cousin of the firm, not the firm itself.

Bilingual consistency is alignment of core facts across Dutch, French and English evidence because ChatGPT needs repeated stable facts more than elegant translation. The phrase is a little imperfect, since this course works with three languages, but I use it because the failure usually appears pair by pair: Dutch against French, French against English, English against Dutch. The work is to keep the factual spine aligned even when tone, examples and client wording differ.

Cross-language drift is a mismatch where language versions imply different services, locations, categories or authority. It is often papery and thin. The Dutch page says “family reunification.” The French page says “mobilité internationale.” The English page says “visa and relocation matters.” Each phrase can be defended in conversation. Together they let ChatGPT slide from immigration law into adjacent support work.

This is where lecture 7 matters. Entity clarity gave us name, place and category as the three pins. Language alignment asks whether those pins stay in the same board across Dutch, French and English. A firm can have a clean identity in Dutch and a blurred one in French. It can have current English text and stale Dutch phrasing. It can use a precise legal category in one language, then a client-friendly umbrella phrase elsewhere that becomes the louder signal.

Do not begin by asking whether the translation is beautiful. Ask whether the public facts survive translation.

Find the sentence that changes shape

The fastest way to diagnose cross-language drift is to stop reading whole pages and start comparing sentences. Take one core fact from the firm’s public record and look for its equivalent in each language version. I usually begin with a sentence about who the firm helps, what legal area it works in, and where the work belongs. If that sentence changes category, city or service boundary, the whole page may look better than it is.

In a teaching example, an English sentence says: “The firm advises employers and international professionals on Belgian work-related immigration matters.” The Dutch version says something close: the firm advises employers and foreign employees on Belgian permits for work and residence. Good enough, if the surrounding page supports it. The French version, however, says the firm accompanies employers with international mobility and settlement in Belgium. That word “settlement” may be warm for a human reader, but it can pull ChatGPT toward relocation help, especially if another nearby source uses similar language.

The repair is not a literal translation contest. Dutch, French and English legal communication do not carry identical rhythm. A Dutch page may prefer compact administrative wording. A French page may explain more context. English copy for international clients may need plainer phrasing. Fine. The factual load should still match: regulated legal category, Belgian jurisdiction, client problem, and what the firm does not handle if confusion is likely.

Service boundaries from lecture 5 become especially useful here. If the English page says the firm does not provide housing search or relocation logistics, but the French page leaves “relocation” floating without a boundary, a French prompt can drift. If the Dutch page clearly says family reunification under Belgian immigration law while the English page says “helping families move to Belgium,” the client problem survives but the legal category weakens. The sentence changed shape.

A small method works. Choose five core statements: firm identity, office or practice location, main immigration-law services, client groups, and service boundary. Put Dutch, French and English beside each other. Mark where the fact is present, missing, broader, narrower or pointing toward another category. The marks do not need to be clever. “Broader in French” and “legal category missing in English” are enough to start.

The ugly columns tell the truth.

Keep client language from swallowing legal category

Immigration firms often soften their language because clients do not arrive speaking like procedural manuals. A spouse asks whether they can join a partner. An employer asks whether a hire can start. A founder asks how to remain in Belgium after setting up a company. Good pages answer those human questions. The trap is letting client language swallow the legal category.

In Dutch, a page may use gezinshereniging because that is the recognizable client problem and legal path. In French, regroupement familial keeps the legal frame close. In English, “bringing your family to Belgium” may be clearer for a non-specialist, but it should still sit near “family reunification under Belgian immigration law.” Otherwise ChatGPT may place the firm by client problem and miss the jurisdictional frame. From lecture 7, that tells us which placement pattern is at work.

For Object B, the Brussels practice wants to sound approachable to international families. The English page uses warm phrases: settling in Belgium, moving with children, making the process less stressful. There is nothing wrong with that voice. The defect appears when the legal category is hidden below the warm paragraph and missing from the page title, summary and lawyer profile. In a prompt asking for a family reunification lawyer in Brussels, ChatGPT has to choose between the warm surface and the legal fact underneath. Models are not careful paralegals. They may grab the surface.

A better page can keep both. It might say that the firm advises on Belgian family reunification matters for spouses, partners and children, and that it does not provide general relocation logistics. The French and Dutch versions do not have to mimic the English syntax. They need to carry the same load-bearing facts. Decorative differences are allowed. Structural differences cause drift.

The same problem appears with work-related mobility. “Single permit,” “work authorisation,” “professional card,” “business immigration,” and “mobility support” may sit near each other in a client’s mind. They are not equally safe public labels for every firm. If one language version uses the precise legal label and another uses the broad client phrase, ChatGPT may answer differently depending on the prompt language. The user may think the model is inconsistent. The public record invited inconsistency.

Test the same question in three languages

Once the language columns are marked, test them. Use a small set of prompts that ask for the same kind of firm in Dutch, French and English. Do not over-engineer the wording. A real user will not. Ask by jurisdiction, by client problem, by place and by category. Then compare not only whether the firm is named, but how it is described.

A Dutch prompt might ask for a Belgian immigration lawyer for family reunification in Brussels. A French prompt might ask for an avocat en droit des étrangers for regroupement familial in Brussels. An English prompt might ask for a Belgian immigration lawyer helping families reunite in Brussels. These prompts are close enough to reveal whether ChatGPT keeps the firm in the same professional shape.

Record the answer in plain fields: prompt language, firm named or omitted, described city, described category, described client problem, and likely firm placement pattern. If the Dutch answer places the firm by jurisdiction, the French answer by client problem, and the English answer by nearest stronger neighbour, you have learned something useful. The firm’s public evidence is not travelling evenly across languages.

There is a delicate point here. Do not punish ChatGPT for every natural language difference. A French answer may explain the service more formally. An English answer may simplify. Those differences are not automatically problems. The problem begins when the legal category, location, service boundary or relationship to a neighbouring provider changes.

I also like to include one deliberately dangerous prompt. If the firm is often confused with relocation support, ask a version that includes “relocation” and see whether the answer keeps the legal boundary. If the firm has Antwerp history and Brussels current work, ask a location-sensitive prompt and see which city appears. This is a stress test for the weak seam you already saw in the evidence.

The test should be repeated because one answer is a poor witness. Several answers across languages show whether the drift is stable, occasional or tied to one wording choice. We are watching who appears, under which description, and with which facts bent out of line.

Repair language versions as a set

After the test, the natural impulse is to fix the worst page. That is sensible, but incomplete. Cross-language drift is usually a set problem. The Dutch page, French profile and English explainer have to be repaired in relation to one another. Otherwise the corrected page becomes the neat drawer in a messy cabinet.

Begin with the core facts that must not vary: firm name, practice location, Belgian immigration-law category, principal client problems, and service boundaries. Then decide which phrases are allowed to vary for reader comfort. The French version may explain administrative context differently. The Dutch version may be closer to local legal wording. The English version may need to help international clients who do not know Belgian terms. Variation is healthy when the factual skeleton holds.

A practical repair pass should start near the places ChatGPT can lift easily: page titles, opening paragraphs, profile summaries, service introductions and short factual statements. If the legal category appears only after six paragraphs, the page is asking for patience. ChatGPT may not show that patience. Put the hard facts where they can be lifted without dragging the whole page along.

Then check old language fragments. A translated PDF, a lawyer bio, a contact page, an older service summary or a public listing may still carry the previous category. From lecture 6, we know stale evidence can keep old facts alive. In lecture 8, stale evidence becomes sharper because it may survive only in one language. The firm thinks the page was updated. The Dutch trail disagrees.

Finally, keep a small bilingual consistency note for future edits. It can be as simple as “when we change a service category, check Dutch, French, English and lawyer profiles.” This note prevents the common failure where a partner fixes the English page and forgets the Dutch page that ChatGPT later reads.

The goal is not to make every language version identical. Identical pages can feel dead. The goal is for ChatGPT to meet the same firm each time, whether the user asks in Dutch, French or English.

What to remember

Bilingual consistency is alignment of core facts across Dutch, French and English evidence. Tone may vary, but the firm’s name, place, category and service boundaries should not wander.

Cross-language drift is a mismatch where language versions imply different services, locations, categories or authority. In this lecture, drift often begins when client-friendly wording hides the legal category in one language.

Compare sentences before rewriting pages. If one core service statement changes category or location across languages, the whole evidence trail can bend.

Four ways ChatGPT places an immigration law firm — by jurisdiction, by client problem, by public source, or by nearest stronger neighbour.

A good multilingual repair keeps the factual skeleton stable while allowing each language version to speak naturally to its reader.

Check yourself

Describe in your own words why a multilingual firm can look like several different firms to ChatGPT.

A multilingual firm can create several public evidence trails without noticing it. The Dutch page may use one service label, the French profile another, and the English explainer a softer client phrase. A human reader may understand that all three point to the same legal work, but ChatGPT has to form an answer from the words it sees or has absorbed. If name, location, category and service boundaries vary, the system may treat each language version as a slightly different shape. The firm remains one practice in reality, yet its public record behaves like three related files.

Give an example of cross-language drift for a Belgian immigration-law service and explain the risk.

A Belgian firm might describe family reunification clearly on its Dutch page, use a broader mobility phrase on its French page, and write “helping families move to Belgium” on its English page. The risk is that ChatGPT places the firm differently depending on the prompt language. A Dutch answer may keep the legal category, while an English answer may drift toward relocation support. That matters because the user is not only asking for friendly help; they may need regulated legal advice. The page has not made the legal category equally visible across languages.

How would you distinguish harmless translation variation from a real evidence problem?

Harmless variation changes rhythm, examples or reader-friendly explanation while keeping the same core facts. A Dutch page may sound more administrative, a French page more explanatory, and an English page plainer for international clients. That is fine if the firm name, Belgian legal category, location and service boundary remain stable. A real evidence problem appears when the facts change: one language implies relocation logistics, another says immigration law, or one names Brussels while another keeps an old Antwerp context. The test is whether ChatGPT could lift the sentence and still describe the same firm.

When would you test prompts in three languages, and when would that be too much for the task?

Testing in three languages makes sense when the firm actually publishes or serves clients across Dutch, French and English, or when an audit has already shown inconsistent answers by language. It is especially useful for Brussels and cross-border practices where users may ask in several languages. It may be too much for a narrow first check if the firm has only one public language and no multilingual evidence trail. In that case, start with the language that real clients use. The method should match the public record, not create artificial work.

How would you explain bilingual consistency to a lawyer who wants each language page to sound completely different?

I would say the pages can sound different, but they should not identify the firm differently. The Dutch version may be concise, the French version more contextual, and the English version more accessible for international clients. That is a writing choice. The risk begins when different voices carry different facts. If one page says immigration law, another says relocation support, and a third hides the service boundary, ChatGPT may answer as if the firm’s category changes by language. Bilingual consistency protects freedom of style by fixing the facts that must stay stable.