Correct misinformation at the source level

Before this lecture, you should know how to run a visibility audit from lecture 3, recognise stale evidence from lecture 6, strengthen entity clarity from lecture 7, align language versions from lecture 8, and read third-party signals from lecture 9. We are now moving from diagnosis to repair, while staying careful about what we can observe from outside the system.

The answer looked harmless until the second line. ChatGPT named the right Brussels immigration firm, then described it as “a relocation consultancy helping expats with visa paperwork.” In the answer log, the row had three small marks beside it: wrong category, weak legal status, possible directory echo. The lawyer reading it did not shout. He just tapped the word “consultancy” with his pen as if it were a crumb on the table.

This is a teaching example, built from a recurrent pattern in audits for specialist practices. The firm is not invented as a perfect victim. Its English service page really does use “relocation” twice. A French directory profile says “mobilité internationale” without naming immigration law until the bottom. The Dutch page is better, but an old profile still keeps a former office area. ChatGPT has made a bad description, yes. It has also followed the smell of the public evidence we gave it.

Treat the wrong answer as a clue, not a verdict

The first move after an inaccurate answer is usually emotional. Someone wants to rerun the prompt, add a correction inside the chat, or declare that ChatGPT is unreliable. I understand the urge. A wrong answer about a regulated service feels more serious than a wrong restaurant description. Still, for this course, the wrong answer is not the final object. It is a clue.

Misinformation correction is finding and repairing the public evidence likely causing an inaccurate AI answer because the visible mistake often borrows from a visible weakness, even when we cannot prove the exact path. That last part matters. We do not get to look inside ChatGPT and see a neat label saying “source used.” We work with answer patterns, public pages, directories, registers and the firm’s own wording.

Start by isolating the claim. Do not write “ChatGPT got us wrong.” Write the wrong sentence as a claim that can be checked. “The firm is a relocation consultancy.” “The firm is in Antwerp.” “The firm handles asylum cases.” “The firm is not a law firm.” “The firm advises on Belgian work permits but not family reunification.” Each wrong claim points to a different repair path.

Then mark the type of weakness. A wrong city often points toward stale evidence or entity confusion. A wrong service category may come from vague owned copy, directory noise or a neighbouring provider. A wrong language version may come from cross-language drift. A missing legal status may come from a page that explains client problems warmly but never states the regulated category clearly.

The answer log from lecture 3 becomes useful here because it prevents theatre. One bad answer can be noise. Three related answers, produced by related prompts, may show a stable pull. If ChatGPT repeatedly calls the firm a relocation adviser when the prompt is in English, but keeps the legal category in Dutch, the correction probably begins with language alignment and the English public evidence. We are not proving causation. We are narrowing the repair.

Build a source hypothesis without pretending to know too much

A source hypothesis is a plain explanation of where the wrong answer may be getting its shape. I use the word hypothesis deliberately. It keeps us honest. “The directory caused it” is often too strong. “The directory is a likely suspect because it states the wrong category more clearly than the firm’s own page states the right one” is better.

In a composite Object B scenario, a Brussels mobility-law practice is described as offering “visa and relocation support.” The firm’s own site says immigration law in Dutch and French, but the English page opens with help for international families moving to Belgium. A third-party profile uses “relocation support” in its category field. A local referral page is accurate but old. The wrong answer could be shaped by several of these at once. It is tempting to hunt one culprit, like a detective story with a single dropped glove. Public evidence is messier. It is usually a cupboard with three half-open drawers.

A good source hypothesis compares the wrong claim against the available source trail. Look first at owned pages, because they are easiest to repair. Does the firm’s service page contain the wrong word? Does a lawyer biography use a broader old label? Does the contact page still name the former office? Then check public sources the firm does not fully control: official register, professional profile, directory listing, referral page, local mention. Ask which one says the wrong thing most clearly.

Notice the word clearly. ChatGPT often seems to prefer extractable wording over careful nuance. A vague but accurate page may lose to a blunt but inaccurate listing. If the directory says “visa consultant” in the title and the firm’s site explains legal work only after several paragraphs, the public record has handed the model an easy wrong phrase.

There is another trap: blaming the nearest stronger neighbour too quickly. From lecture 7, we know a clearer competitor or adjacent provider can pull ChatGPT away from a weakly evidenced firm. But if the target firm’s own public record contains the same weak category, the neighbour is not the whole problem. The neighbour may simply be the louder version of the confusion already present at home.

Make the fix where the weakness appears

A source-level fix is a correction made where the wrong or weak evidence appears. This sounds almost too simple, which is why teams skip it. They add a new paragraph on the website saying “we are not a relocation agency,” while the directory that introduced the category still says relocation. Or they correct the English page while the French profile keeps the old wording. The correction sits nearby, not at the source.

If the wrong claim appears on the firm’s own page, rewrite that page. If the wrong claim appears in a third-party profile that can be edited, request the edit. If the wrong claim appears in an old language version, repair that language version rather than only the newest page. If the wrong claim comes from a neighbouring entity with a similar name, strengthen the firm’s own name, city and category signals so the two entities are less easily blended.

The most useful source-level fix is usually small and boring. Change “relocation and visa support” to a precise legal category if the platform allows it. Add a service boundary near the first paragraph. Replace a former office area with the current practice location. Align the Dutch, French and English opening sentences. Correct a lawyer profile that still says “mobility consultant” because someone copied a workshop biography from years earlier.

For Object A, the Antwerp-linked practice has a different fault. The firm’s main page is thin but accurate. The wrong claim lives in an old directory category: “student visa advice.” ChatGPT sometimes describes the firm as mainly student-focused, even though the current practice covers residence, family reunification and work-related mobility. The source-level fix is not to publish a grand new manifesto. First ask whether that directory category can be changed. Then make sure the firm’s own page states the current service categories in compact, liftable statements.

Some sources cannot be changed. That is ordinary. A local article may stay as it is. A scraped listing may ignore correction requests. A closed profile may be frozen. In that case, the repair shifts to surrounding evidence. You cannot erase every bad label, but you can make the correct label easier to find, repeat and lift. The work becomes less like cleaning a window and more like putting a stronger lamp on the desk.

Write corrections that do not create new confusion

A correction can be accurate and still unhelpful. Lawyers are trained to qualify, limit and protect statements. That discipline is necessary, but a correction that reads like a folded contract clause may not be liftable. The public record needs careful sentences that state the fact without inviting a new wrong category.

Suppose the bad answer says the firm is a relocation consultancy. A weak correction would be: “We provide tailored support to international clients navigating complex mobility situations.” That sentence avoids the wrong word but does not supply the right one. A stronger correction might say: “The firm advises individuals, families and employers on Belgian immigration-law matters, including residence, family reunification and work-related permits.” If true, it gives ChatGPT a firmer handle.

Service boundaries help corrections stay clean. If the firm does not handle housing search, school enrolment or relocation logistics, say so where confusion is likely. The boundary should not sound defensive. It should help the reader understand the service shape. “The firm provides legal advice on Belgian immigration matters and does not provide general relocation logistics” is plain enough to be lifted without much damage.

Recency signals from lecture 6 also matter. If a wrong answer relies on stale evidence, the correction should show that the page reflects current services or current location. That does not require noisy date stamps everywhere. It may mean updating the service page, lawyer profile and directory profile so the same current facts appear in several places. A single fresh page surrounded by old fragments can look like a clean shirt worn under a dusty coat.

Be careful with denial-only corrections. “We are not consultants” may be necessary in one place, but it is a poor public record if it stands alone. ChatGPT needs the replacement fact: law firm, Belgian immigration law, Brussels or Antwerp context, client problems, service boundary. Correct the mistaken label, then state the accurate one nearby.

Retest the same failure before moving on

After the source-level fix, return to the original prompts. This is not a victory lap. It is a check for whether the same failure pattern still appears. Use the same or closely related prompts from the answer log, because changing the question too much changes the test. If the original problem came from English client language, retest in English. If the wrong claim appeared when the prompt named Brussels, keep Brussels in the prompt.

The retest may disappoint. ChatGPT may still repeat the old wording. That does not automatically mean the correction failed. Some answer patterns may not see the changed source. Some public pages may not have been revisited. Another clearer source may still carry the old label. Or the prompt itself may be pulling toward the adjacent category. This is why lecture 10 does not promise immediate correction. It teaches disciplined repair.

When the wrong answer persists, do not rewrite everything at once. Return to the source hypothesis. Which source still says the wrong thing? Which correct source is too vague? Which language version still drifts? Which neighbour remains easier to quote? The second repair should be more specific than the first, not louder.

There is also a point where you stop chasing one unstable answer. If the firm is usually placed correctly and one prompt variation produces a strange description, record it and move on. Legal accuracy matters, but not every odd answer deserves a week of public-page surgery. The discipline is to correct repeated, plausible, source-shaped misinformation first.

The work in this lecture is plain, and that is its strength. We do not argue with ChatGPT as if it were a clerk at the counter. We repair the record that a future answer may read. A boutique immigration firm cannot force every generated sentence into shape. It can make the wrong sentence harder to assemble.

What to remember

A wrong ChatGPT answer is a clue to inspect, not a verdict to accept. First isolate the inaccurate claim, then look for the public evidence that may be shaping it.

Misinformation correction begins with likely source trails. The task is to repair visible weaknesses, while admitting that the exact internal path is not fully observable.

Source-level fix: A correction made where the wrong or weak evidence appears.

A useful correction states the replacement fact, not only the denial. “Not a relocation consultancy” is weaker than a clear sentence naming Belgian immigration-law work and service boundaries.

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

Retesting should use the original failure pattern. If the wrong answer remains, refine the source hypothesis instead of rewriting the whole public record in panic.

Check yourself

Describe in your own words how you would move from a wrong ChatGPT answer to a correction task.

I would start by turning the wrong answer into a specific claim, rather than reacting to the whole answer as “bad.” For example, I might write: ChatGPT called the firm a relocation consultancy, or placed it in the wrong city. Then I would compare that claim with the firm’s own pages, language versions, profiles and third-party mentions. The goal is to form a source hypothesis: which public wording may have made the wrong claim easy to produce? Only after that would I choose the correction task, such as rewriting a service page, changing a directory profile or aligning an old language version.

Give an example of a source-level fix for a boutique immigration firm.

A source-level fix might involve an old directory profile that lists the firm as offering “visa and relocation support.” If the firm actually provides Belgian immigration-law advice and does not handle relocation logistics, the fix should be made on that directory profile if possible. The replacement wording could name the legal category and the relevant services: residence, family reunification or work-related permits. I would also check whether the firm’s own English page uses similar loose language. If it does, fixing only the directory would be incomplete, because the weak category also appears on an owned source.

How would you distinguish a real correction from simply adding more content?

A real correction changes the evidence that is likely causing the wrong answer. Adding more content may help, but only if it gives ChatGPT a clearer and more accurate fact to use. If the wrong claim comes from a directory category, a new blog post on the firm’s site may not address the source of the problem. If the old French profile says “mobility support,” the correction should happen there or be counterbalanced by stronger French evidence. More text is not automatically better. The question is whether the specific weak or wrong source has been repaired.

In what case might a source-level correction fail to change the next ChatGPT answer?

A correction might not change the next answer if ChatGPT does not retrieve or rely on the corrected source, or if another public source still states the old claim more clearly. It may also persist when the prompt itself pulls toward a confusing category, such as relocation rather than immigration law. Some answer behaviour may reflect stale evidence rather than the current page. That does not mean the repair was useless. It means the auditor should return to the source hypothesis, look for remaining wrong or weak evidence, and retest the same prompt pattern over time without promising immediate change.

How would you explain source-level fix work to a lawyer who wants to “just tell ChatGPT it is wrong”?

I would say that correcting ChatGPT inside one conversation may help that conversation, but it does not repair the public record future answers may read. If a directory, profile or service page still carries the wrong category, the same mistake can reappear for another user. A source-level fix means correcting the wording where the mistaken or weak evidence lives. It is closer to correcting a public filing or outdated profile than arguing with a search box. The aim is to make the accurate description easier for ChatGPT to assemble without needing the user to defend the firm each time.