It is a free research preview that writes fluent text on request. It cannot look anything up, its knowledge stops in 2021, and it states incorrect things confidently.

What actually happened

OpenAI released a conversational tool called ChatGPT on Wednesday, free to use, described as a research preview.

You type a question or an instruction and it replies in fluent prose, and it remembers what was said earlier in the same conversation.

It has spread very quickly, which is unusual for something requiring an account and no promotion.

Which means you are going to be asked about it, by customers, by suppliers, and possibly by whoever handles your marketing.

What it does well

Worth stating first, because the capability is genuine and dismissing it would be wrong.

It produces grammatical, well-structured, confident prose on almost any subject, immediately.

It follows instructions about tone, length, and format reasonably well.

It will rewrite something you give it, summarise it, or produce several versions to choose between.

And it does all of that in seconds, at no cost during the preview, which is why it has spread the way it has.

What it cannot do

The first two are stated by the makers themselves: it has no connection to the web and its knowledge of the world stops in 2021. So anything about a current price, a recent rule, or a local detail is being produced without any source.

Fluent and correct are separate

The single most important thing to understand about it.

The tool produces text that reads as though it were written by somebody who knows, whether or not the content is right.

That is different from every previous tool, where an error usually looked like an error.

Here a wrong regulation, an invented figure, or a plausible but incorrect explanation arrives in the same confident register as a correct one.

The makers say as much: it may occasionally generate incorrect information, which in practice is more than occasional on anything specific.

Which means everything factual it produces has to be checked by somebody who already knows, and that check is the actual work.

It is not a search engine

Worth separating, because the two are already being confused.

A search engine finds pages that exist and shows you where the answer came from, so you can judge the source.

This produces an answer with no source at all, assembled from patterns rather than retrieved from anywhere, and there is nothing to click through to.

Which means it cannot be checked in the way a search result can, and a confident answer offers no way to find out whether it is right.

Anybody using it as a substitute for looking something up has removed the one step that made the answer verifiable.

Keeping it in proportion

Since the volume of commentary this week is out of proportion to what has changed for a small business today.

Nothing about your site, your customers, or your obligations changed on Wednesday.

Nobody is going to find your business through this, and no competitor has gained an advantage over you in three days.

What is worth doing this week is trying it, on something you already know the answer to, so your view is based on having used it.

That takes twenty minutes and puts you ahead of most opinions being offered about it.

A worked example

A business owner asked it to write a page about a service he had provided for eighteen years.

The result was well written, correctly structured, and read like a competent page.

It was also generic, contained two claims about the work that were wrong, and included a figure he could not account for.

His assessment was that it had produced in thirty seconds something it would have taken him an hour to write and that he would not publish either version.

He also noticed that it had produced the page a competitor would get, since neither of them had told it anything about their own work.

That last observation is the useful one and is the subject of the next few days.

Where it goes from here

Briefly, since forecasting is not the point of this.

Tools of this kind will improve, become available inside other software, and stop being remarkable.

What is unlikely to change quickly is that it does not know anything about your business, which is the limitation that matters most for a site that has to say something specific.

The right posture is neither dismissal nor rearrangement of your plans, but understanding what it is for.

The coming week covers what happens when you actually try to use it for a website, which is the question everybody is about to have.

Be careful what you type into it

A practical caution worth having before anybody in the business starts using it.

What you type goes to a company's servers, and the makers say conversations may be reviewed to improve the system.

Which means customer names, addresses, order details, contracts, and anything confidential should not be pasted into it.

That is easy to do accidentally when somebody asks it to rewrite an email or summarise a document that happens to contain a client's details.

Say so plainly to anybody who works with you, now, before it becomes a habit that has to be corrected later.

The counter-case

Both extremes are being argued this week and both are wrong.

Dismissing it as a toy underestimates something that genuinely writes well and will be used by a great many people regardless of anybody's opinion.

Treating it as a replacement for knowing your own business misunderstands what it is doing, which is producing likely text rather than accurate text.

And a great deal of confident commentary this week is being written by people who have used it for an hour.

Try it yourself, on something you can check, and form a view from that.

This week

  1. Try it on something you already know.
  2. Check every fact it produces.
  3. Notice what it invents.
  4. Ask it about your own trade.
  5. Change nothing on the strength of one week.
  6. Do not publish anything it wrote.
  7. Form your own view.

Step two is where the surprise is, since the errors do not look like errors and are visible only to somebody who already knew the answer.

The last announcement worth keeping in proportion is covered in Google started rewarding helpful content.


Frequently asked questions

What was released?

A conversational tool called ChatGPT, from OpenAI, free during a research preview. You type an instruction and it replies in fluent prose, remembering earlier turns in the conversation.

What does it do well?

Produces grammatical, well-structured, confident prose immediately, follows instructions about tone and length, and will rewrite or summarise something you give it.

What can it not do?

Look anything up, know about recent events, know your business, or tell you when it is wrong. Its makers state its knowledge of the world stops in 2021.

What is the main risk?

That fluent and correct are separate. A wrong regulation or an invented figure arrives in the same confident register as a correct answer, which is unlike previous tools.

Has anything changed for my business?

Not this week. Nothing about your site, customers, or obligations changed. Nobody will find you through it and no competitor has gained an advantage in three days.

What should I actually do?

Try it for twenty minutes on something you already know the answer to, check every fact it produces, and form a view from having used it.

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.

Being asked about the new chatbot?

Spend twenty minutes asking it about your own trade and check every claim. That is worth more than any commentary this week.

Start a Conversation