What It Gets Confidently Wrong

The signal that has gone

People judge whether to check something partly on how it is written.

Hesitant phrasing, vagueness, or an obvious mistake elsewhere all prompt a second look, and that heuristic has worked reasonably well for a long time.

This produces uniformly confident, well-structured prose whether the content is right or wrong.

So the signal is gone, and the only remaining method is to check everything specific, which is considerably more work than people expect.

What it gets wrong most reliably

• Figures and statistics.
• Rules, regulations, and requirements.
• Anything local to a place.
• Recent events, since 2021.
• Names, dates, and sources.
• Anything specific to your trade.

The fifth is the one that surprises people most. It will produce a citation, an author, and a date that look entirely ordinary and refer to nothing that exists, because a plausible-looking reference is exactly what the pattern predicts.

Why it does this

Briefly, since understanding it changes how you use it.

It is producing likely text rather than retrieving stored facts, so there is no distinction inside it between a claim it can support and one it cannot.

A statistic that sounds like the kind of statistic that appears in such a paragraph is exactly what the method produces, whether or not the number is real.

It also has no connection to the web and no knowledge of anything after 2021, which its makers state plainly.

So anything about a current price, a rule changed last year, or a recent event is being produced without any source at all.

It is not lying, because it has no notion of the difference, which is why asking it whether it is sure does not help.

The specific risks for a small business

Since the general warning is easy to nod at and ignore.

A page stating a legal requirement incorrectly, which a customer may rely on.

A figure about typical costs or timescales that is wrong for your area and undercuts or oversells your own work.

A safety statement that is close to right and wrong in a way that matters.

An invented statistic that a competitor or a customer checks and cannot find.

Each of those is worse than the page not existing, and each arrives looking exactly like the correct version.

A worked example

A business asked it about the requirements for a common piece of work in their province.

The answer was clear, well organised, and cited a specific regulation with a number.

Two of the four requirements listed were correct, one applied in a different province, and the regulation number did not correspond to anything.

The owner only noticed because he had done this work for years and the third item looked wrong to him.

Somebody without that knowledge would have published all four, which is the actual danger.

His conclusion was that it was useful only on subjects where he could already tell, which considerably narrows what it is for.

The verification is the work

And it is frequently longer than writing the thing yourself.

Every figure has to be traced to a source you trust, which means looking it up properly rather than searching for confirmation.

Every rule has to be checked against the actual regulator or authority, which is where you should have started.

Every name, date, and reference has to be confirmed to exist.

Do that honestly on a generated page and the time saved usually disappears, which is the finding most people reach after a few weeks.

The saving is real where no verification is needed, which is a narrower category than it first appears.

Ask it only what you can check

The practical rule that follows from all of this.

Use it on subjects where you would immediately spot an error, which for most people means their own trade and little else.

Do not use it to learn something new, because you cannot evaluate what it tells you and confident wrong answers are the worst possible teacher.

Do not use it for anything you would be liable for: legal, safety, financial, or medical.

And do not ask it for sources, since it will provide convincing ones that do not exist.

That rule is restrictive and it is the one that prevents the expensive mistakes.

Tell whoever writes for you

Since the risk arrives through other people.

Anybody producing content for you has access to this and it is free, which means the first generated page may already be on your site.

Say plainly that factual claims must be verifiable and that you will ask where a figure came from.

That question, asked once about one statistic, establishes the standard more effectively than any policy document.

And check anything published recently that contains a number you cannot account for.

It agrees with you too readily

A failure mode worth knowing, because it makes checking harder rather than easier.

Push back on an answer and it will frequently apologise and produce a different one, whether or not the original was correct.

Which means you can talk it out of a right answer as easily as a wrong one, and its agreement is not evidence of anything.

The same applies to asking whether it is certain, which produces a confident reassurance rather than a genuine assessment.

So the checking has to happen outside the conversation entirely, against a real source, since nothing inside it can confirm anything.

The counter-case

The error rate is not uniform.

On general explanations of well-established subjects it is frequently accurate, and treating everything it says as suspect wastes something genuinely useful.

These tools are also improving, and the specific failure described here may be considerably reduced in versions that can look things up.

And a human writer produces errors too, which nobody proposes to solve by not employing writers.

Check every specific claim, use it only where you could spot an error yourself, and never ask it for a source.

What to do

1. Assume every figure is wrong until checked.
2. Trace claims to a real source.
3. Never accept a reference it supplies.
4. Use it only where you could spot an error.
5. Never use it for legal or safety content.
6. Ask writers where a figure came from.
7. Recheck anything published recently.

Step three catches the failure people are least prepared for, since an invented citation with an author and a date looks exactly like a real one.

Writing beyond your own knowledge is covered in writing about something you half understand.

Frequently asked questions

West Coast Media Solutions Inc. provides web design, web development, hosting, digital marketing, and business consulting to organisations across Canada, drawing on more than twenty-five years in the field.