OpenAI released ChatGPT on November 30 as a free "research preview." On Monday, Sam Altman announced that it had surpassed one million users. One week later, it is the main topic of conversation across the technology sector, and among many people who had never previously taken an interest in technology.
Having been among the early testers of OpenAI's GPT-3, we noted the situation with some interest: the same model family, a new interface, and suddenly the general public is engaging with it.
That new interface matters more than we anticipated.
Why does the chat interface make such a difference?
Because it removes the technical barrier. GPT-3 has been impressive for about two years. However, using it required a developer tool, adjustments to a "temperature" setting and an understanding that the model was continuing the user's text rather than answering a question. Users had to learn how to guide it toward useful output. Most people never tried, which is understandable.
ChatGPT takes a model from the GPT-3.5 series, trains it further to follow instructions and hold a conversation, and places it behind a text box resembling a messaging app. The user asks a question and receives an answer. The user requests a shorter version, and the model shortens it.
From the public's perspective, this is the entire change. No new science is visible; there is simply an entry point accessible to everyone.
This pattern is common in technological progress and frequently underestimated. The capability already existed. What was missing was an access point that did not require reading documentation.
What is ChatGPT good at today?
Language tasks, primarily. The uses that work well so far include rewriting an awkward paragraph, summarizing a long document, drafting a first version of an email that has been postponed for a week, explaining a concept at three different levels until one is understood, and converting a bulleted list into readable text and back again.
It is also a surprisingly capable thinking partner. It is frequently wrong, but it will continue the discussion for as long as needed, without fatigue, and it occasionally identifies an overlooked angle.
For small business owners managing every aspect of their operations, this is significant. Writing represents a larger share of running a business than is generally recognized: emails, quotes, posts, descriptions and repeated answers to the same questions. Even this early, unrefined version can save meaningful time on these tasks.
Can the information ChatGPT provides be trusted?
Not without verification. This is a serious concern, and one widely shared.
ChatGPT fabricates information, and not only occasionally or in obvious ways. It can provide a confident, well-written answer containing a fictitious statistic, a non-existent book or an invented quotation, in exactly the same tone as the accurate content. OpenAI's own page states that its knowledge of the world after 2021 is limited, which means recent information is either missing or invented.
On Monday, Stack Overflow introduced a temporary ban on answers generated by ChatGPT, because users were filling the site with answers that appeared correct but were not. This decision illustrates the issue clearly.
Our rule is therefore straightforward: any fact that matters must be verified elsewhere, including names, figures, dates, laws, prices and anything related to health. The model is an excellent writer but an unreliable source.
Recommended practices
- Provide the facts yourself and ask the model to work with them, rather than asking it to supply them.
- Ask it to indicate when it is uncertain. It will not always do so, but the request helps.
- Do not enter any private information you would not want stored on a third-party server. This is a research preview, and OpenAI states that conversations may be reviewed.
- Review every word before anything is published under your name.
How should ChatGPT be used to get the best results?
As support for thinking, not as a replacement for it. A greater concern than the errors is the speed with which users are delegating entire tasks: writing an essay, a blog post or a reply to a customer, then moving on.
The output produced this way sounds acceptable but conveys little substance, and readers notice, if not immediately, then over time.
In our view, the most effective use is as a sparring partner. The user provides the knowledge, the perspective and the experience; the tool helps shape them, identify weaknesses and express them more clearly. Years of observing small businesses write their own websites have shown us that the difficulty rarely lies in grammar. It lies in determining what the business actually wants to say. ChatGPT can help with this when it is also allowed to ask questions.
Most people need neither excessive enthusiasm nor alarm about this tool. They need guidance on how to use it while continuing to exercise their own judgment. We anticipate that this will represent a substantial need in the coming period.
Who is likely to benefit most from ChatGPT?
Not those asking broad questions about the future of humanity, but those with specific, routine tasks: a weekly newsletter, product descriptions, replies to the same five customer questions, or any weekly task that is rarely enjoyable.
For small business owners who wish to try it, we recommend starting there: one task, your own work, over a few weeks. Determine what the tool is genuinely useful for before purchasing or changing anything. Learning it properly, as with any tool, is currently worth more than any software subscription, and for those who prefer not to learn it alone, practical AI training provides more value than the many quick opinions circulating online.
We also recommend reading OpenAI's announcement, Introducing ChatGPT. It is short and notably transparent about the model's limitations. OpenAI does not conceal the problems; most of those sharing screenshots have simply not read that far.