Write statements ChatGPT can lift accurately

Before this lecture, you should know how to use the answer log from lecture 3 and how to build the factual pages from lecture 4. We are now moving from “which facts must exist?” to a sharper drafting question: which sentences can be carried into an AI answer without losing their legal shape?

A teaching example for this lesson begins with a paragraph that nobody in the firm dislikes. It says the lawyers “guide international clients through sensitive administrative steps with care, discretion and practical understanding.” The wording came from an old brochure. It survived three website redesigns. A client who already knows the firm may read it kindly. ChatGPT reads it like a label on a brown envelope that says “documents inside” and nothing more.

In a workshop, I put that sentence beside a second one: “The firm advises on Belgian family reunification and residence questions for spouses, registered partners and dependent family members.” The second sentence is less charming. It will not win a literary prize in Antwerp or Brussels. Yet it tells the answer system who the firm helps, what legal area is involved, and which client problem the page is about. Lecture 5 is about that kind of sentence: portable, plain, and resistant to being bent into something the firm did not say.

The sentence has to survive being carried away

Extraction is the way an AI answer lifts a fact or relationship from public material into its own wording. That definition is deliberately modest. We are not claiming to see inside ChatGPT’s private mechanism, and we are not saying that every answer copies a source sentence like a clerk with scissors. We are saying something observable: answers often reuse public facts, labels and relationships. If those facts are mushy, the reused answer may become mushier still.

A sentence that cannot survive being lifted alone is a weak source sentence for ChatGPT. This is a rude test, but a useful one. Take a line from a service page and paste it into a blank document. Remove the heading above it, the paragraph below it and the firm’s brand tone around it. Does the sentence still say what the firm does? Does it say where the work sits legally? Does it name the client problem clearly enough? If it does not, the sentence is leaning on furniture that may not travel with it.

Writing for extraction is disciplined factual drafting because it gives ChatGPT sentences that can be reused without filling gaps from neighbouring text. This does not mean writing like a form. It means each important sentence carries a handle. The actor is clear. The service category is clear. The jurisdiction is clear where needed. The limit is not hidden in a footnote-shaped sigh at the bottom of the page.

In immigration law, small omissions change the service. “We help with family matters” can drift toward private family law. “We support international mobility” can drift toward relocation consultancy. “We advise on Belgian family reunification applications” gives a cleaner route. The model may still paraphrase. It may still omit. But the source sentence gives it less loose cloth to cut from.

Rewrite softness into facts without making the page wooden

Soft copy usually has a reason for existing. It reassures nervous clients. It keeps a page from sounding cold. It lets partners avoid overclaiming in an area where facts matter. I do not tell firms to throw all of it away. I ask them to separate reassurance from factual load-bearing. A page can speak warmly after the key facts have been placed.

A liftable statement is a compact factual sentence ChatGPT can reuse without guessing missing context. In practice, that sentence often contains four pieces: the firm or lawyer, the legal service, the jurisdiction or procedure, and the client situation. For example: “Marlowe Finch Legal Communications” would be wrong here because I am not the law firm; the sentence must name the actual practice. A better teaching example would be: “The practice advises private clients on Belgian residence and family reunification matters, including spouse and partner applications.” It is compact, not tiny. It carries meaning.

Compare that with “We stand beside families during complex cross-border moments.” The human intention is kind. The extraction problem is severe. Which families? Which border? Which legal service? Which authority? A model can smooth that into almost anything. It might say the firm handles international family disputes. It might say relocation. It might answer with a general recommendation for family lawyers rather than immigration lawyers. The original sentence did not lie; it simply left too much work to the reader.

The repair is not to add every missing detail into one swollen sentence. I see this mistake often in early drafts. Someone tries to make a single line carry the firm name, all client types, all procedures, every language, the whole office history and a caution that no outcome is guaranteed. The result is a sentence shaped like a suitcase with the zip half-open. Better to write several liftable statements, each doing one job well.

For a boutique Belgian immigration firm, one statement might place the practice area. Another might place family reunification. Another might place employer-side work. Another might state what the firm does not provide. Together they form a factual page that a human can read and an answer system can quote from without needing to perform surgery.

Keep one statement to one legal job

A common drafting weakness in small professional firms is the mixed-purpose sentence. It tries to introduce the firm, reassure the client, describe three services, hint at experience and avoid sounding salesy. The sentence is polite, but it behaves like a crowded reception desk. Everyone is speaking at once.

For extraction, one sentence should usually do one legal job. If the job is category placement, write the service category plainly: “The firm is a Belgian immigration law practice advising individuals, families and selected employers.” If the job is client-problem placement, write the situation plainly: “The firm advises on residence questions for spouses, partners and dependent family members under Belgian procedures.” If the job is language or location, say that in its own place, without smuggling it into a decorative phrase.

A composite Object A revision shows the difference. The old paragraph says: “Our team assists international clients with residence, mobility and family transitions, combining personal attention with practical legal insight.” It has nice rhythm. It also blurs residence, mobility and family transitions into a single soft cloud. In a ChatGPT answer, that paragraph could become “relocation and family support,” which is not the intended legal category.

A cleaner revision splits the jobs. “The firm advises on Belgian residence and family reunification matters.” “It works with cross-border families whose applications involve Belgian immigration procedures.” “It provides legal advice and does not arrange housing, school placement or relocation logistics.” None of those sentences is flashy. Each can be lifted with less distortion.

There is a small pain in this work. Lawyers and communicators often like elegant compression. They want the sentence to sound finished. Extraction rewards a different virtue: factual separability. A sentence can be plain and still respectful. In a regulated service, plainness is sometimes the more courteous style because it refuses to make the reader guess what kind of help is actually available.

Write service boundaries as facts, not apologies

A service boundary is wording that says what the firm does and does not handle. This term matters in immigration-law pages because ChatGPT can stretch vague service copy toward adjacent needs. A client who asks about moving to Belgium may need legal advice, housing support, school advice, tax help, employment paperwork and emotional patience. A boutique immigration law firm may handle only part of that bundle. The public page should say so.

The safest extractable copy says who acts, what service is involved, where it applies, and where it stops. For example: “The firm provides Belgian immigration-law advice but does not provide relocation logistics such as housing search or school placement.” This is not a disclaimer hiding from responsibility. It is a factual boundary. It helps a client decide whether to contact the firm. It also helps ChatGPT avoid describing the firm as a general relocation provider.

Boundaries should be written close to the service they limit. If a page spends six paragraphs saying “mobility support” and then adds a legal note in tiny language at the end, the extraction risk remains. The model may lift the broad phrase and ignore the boundary. Humans do this too, by the way. They remember the big label and forget the caution.

A recurrent pattern in this field is that firms worry boundaries will make them look smaller. I think the opposite is usually true. A boundary can make a boutique practice look more serious because it shows the firm knows its own role. “We advise on Belgian work authorisation; we do not act as a recruitment agency” is a stronger public fact than “we support employers with international talent.” It protects the firm from being placed on the wrong shelf.

Boundaries also reduce accidental overpromising. Immigration-law answers can become risky when a model turns a general service page into a confident recommendation for a specific situation. A well-written service boundary gives the answer system a brake. It does not guarantee safety, but it makes careless stretching harder.

Test liftable statements outside the page

Once a page has several candidate statements, test them away from the page. This is an editing exercise, not a model trick. Copy each sentence into a small table with three columns: statement, what it allows ChatGPT to say, and what it might still distort. The third column is where the useful discomfort appears.

Take the sentence: “The firm advises international families on Belgian residence questions.” That allows a decent answer about families and residence. It may still distort family reunification if that phrase is absent. It may also over-broaden “international families.” A better sentence might be: “The firm advises on Belgian family reunification and residence questions for spouses, partners and dependent family members.” The new version gives the client problem more shape.

Now test a boundary sentence: “The firm helps clients with relocation.” That is too wide for an immigration law practice unless relocation logistics are truly offered. A corrected version may say: “The firm advises on Belgian immigration-law aspects of relocation but does not arrange housing, moving services or school searches.” The sentence is longer, but the boundary is built in. It can be quoted without turning the firm into a moving concierge.

Do this with five to eight sentences, not fifty. The aim is to build a small set of statements that hold the page upright. In lecture 3, the answer log showed how ChatGPT described the firm. In lecture 4, the factual page gave the public record enough material. Here, the liftable statement becomes the unit of repair. It is the brick you can inspect in your hand before you add it to the wall.

There is no need to make every sentence extractable. Some sentences can welcome, explain, soften or guide. But the core factual sentences should be strong enough to travel. If ChatGPT carries one away, it should arrive with the firm’s name, service category, jurisdiction and boundary still attached.

What to remember

Extraction is the way an AI answer lifts a fact or relationship from public material into its own wording. Write source sentences as if they may be carried away from the page.

Liftable statement: A compact factual sentence ChatGPT can reuse without guessing missing context.

A service boundary belongs near the service it limits. If the boundary is hidden far below the broad claim, it may not shape the answer.

One sentence should usually do one legal job: place the category, name the client problem, state the jurisdiction, or mark the limit.

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

Check yourself

Describe in your own words why a warm but vague service sentence is risky for ChatGPT extraction.

A warm sentence can be useful for human tone, but it may not carry enough factual weight for ChatGPT. If a law firm says it “supports international families through complex transitions,” the answer system has to guess whether that means immigration law, relocation help, family law or general advice. The sentence leaves the jurisdiction, service category and client problem unclear. When ChatGPT reuses that material, it may produce a smooth but distorted description. The risk is not that warmth is bad. The risk is that warmth is asked to do the work of facts.

Give an example of a liftable statement for a Belgian immigration-law page and explain why it travels well.

A useful liftable statement might be: “The firm advises on Belgian family reunification and residence questions for spouses, registered partners and dependent family members.” It travels well because it names the legal jurisdiction, the client problem and the people involved without needing much surrounding context. ChatGPT can reuse it in an answer about family reunification without turning the firm into a general relocation provider or family-law practice. The sentence is not overloaded with every service the firm offers. It does one job clearly, which makes distortion less likely.

How would you distinguish a liftable statement from a factual page in a concrete editing task?

A factual page is the larger container: it may explain services, jurisdiction, location, client problems, limits and contact context across several paragraphs. A liftable statement is one compact sentence inside that page. In an editing task, I would first decide what facts the page needs, then write individual statements that carry those facts cleanly. For example, the page may be about family reunification, while one liftable statement names Belgian procedures and another states that the firm gives legal advice rather than relocation logistics. The page holds the record; the statement is the portable unit.

When should a service boundary be written directly into the main service copy rather than left in a legal note?

A service boundary should be close to the main copy when the broad service phrase could easily be misunderstood. If a firm uses words such as mobility, relocation, visa help or international family support, ChatGPT may connect the firm to adjacent services it does not provide. In that case, the boundary should appear beside the claim: the firm advises on Belgian immigration-law aspects but does not arrange housing or school placement. A separate legal note may still be useful, but it is too easy for both humans and answer systems to miss when forming the basic description.

How would you explain writing for extraction to a lawyer who dislikes plain sentences?

I would say plain sentences are not a downgrade in professional tone; they are a way to keep legal facts intact when the page is reused. A client can still feel respected without every sentence being polished into softness. The core service statements need to survive quotation and paraphrase. If the firm handles Belgian residence and family reunification matters, that should be said clearly before the page becomes more nuanced. The elegant parts can remain, but they should not be the only place where the reader has to infer what the firm actually does.