What Is ChatGPT and How Does It Work?

Don't have time to read the entire article? Jump to the quick read summary.

©Irissca / Adobe Stock

Additionally, anyone can interact with ChatGPT (GPT stands for generative pre-trained transformer) in their own words. A natural, humanlike dialog ensues.

ChatGPT is often directly accessed online by users, but it is also being integrated with several existing applications, such as Microsoft Office apps (Word, Excel, and PowerPoint) and the Bing search engine. The number of app integrations seems to grow every day as existing software providers hurry to capitalize on ChatGPT’s popularity.

What is ChatGPT used for?

In general, people tend to use ChatGPT to guide or explain something, as if the bot were a fancier version of a search engine. Nothing is wrong with that use, but ChatGPT can do so much more.

How much more depends on how well you write the prompt. If you write a basic prompt, you’ll get a bare-bones answer that you could have found using a search engine such as Google or Bing. That’s the most common reason why people abandon ChatGPT after a few uses. They erroneously believe it has nothing new to offer. But this particular failing is the user’s fault, not ChatGPT’s.

What can ChatGPT do?

• Conduct an interview with a long-dead legendary figure regarding their views of contemporary topics.
• Recommend colors and color combinations for logos, fashion designs, and interior decorating designs.
• Generate original works such as articles, e-books, and ad copy.
• Predict the outcome of a business scenario.
• Develop an investment strategy based on stock market history and current economic conditions.
• Make a diagnosis based on a patient’s real-world test results.
• Write computer code to make a new computer game from scratch.
• Leverage sales leads.
• Inspire ideas for a variety of things from A/B testing to podcasts, webinars, and full-feature films.
• Check computer code for errors.
• Summarize legalese in software agreements, contracts, and other forms into simple laymen language.
• Calculate the terms of an agreement into total costs.
• Teach a skill or get instructions for a complex task.
• Find an error in their logic before implementing their decision in the real world.

Unfortunately, that’s true for bad guys too. For example, they can prompt ChatGPT to find vulnerabilities in computer code or a computer system; steal your identity by writing a document in your style, tone, and word choices; or edit an audio clip or a video clip to fool your biometric security measures or make it say something you didn’t actually say. Only their imagination limits the possibilities for harm and chaos.

Unwrapping ChatGPT fears

There are definitely legitimate reasons for caution and concern. Lawsuits have been launched against generative AI programs for copyright and other intellectual property infringements.

OpenAI and other AI companies and partners stand accused of illegally using copyrighted photos, text, and other intellectual property without permission or payment to train their AI models. These charges generally spring from copyrighted content getting caught up in the scraping of the Internet to create massive training datasets.

In general, legal defense teams are arguing the inevitability and unsustainability of such charges in the age of AI and requesting that charges be dropped.

The lawsuits regarding who owns the content generated by ChatGPT and its ilk lurk somewhere in the future. However, the U.S. Copyright Office has already ruled that AI-generated content, be it writing, images, or music, is not protected by copyright law. In the U.S., at least for now, the government will not protect anything generated by AI in terms of rights, licensing, or payment.

Meanwhile, realistic concerns exist over other types of potential liabilities. ChatGPT and ChatGPT alternatives are known to sometimes deliver incorrect information to users and other machines. Who is liable when things go wrong, particularly in a life-threatening scenario?

Even if a business’s bottom line is at stake and not someone's life, risks can run high and the outcome can be disastrous. Inevitably, someone will suffer and likely some person or organization will eventually be held accountable for it.

Then, there are the magnifications of earlier concerns, such as data privacy, biases, unfair treatment of individuals and groups through AI actions, identity theft, deep fakes, security issues, and reality apathy, which is when the public can no longer tell what is true and what isn’t and thinks the effort to sort it all out is too difficult to pursue.

In short, all of this probably has you wondering: Is ChatGPT safe? The potential to misuse it accelerates and intensifies the need for the rules and standards currently being studied, pursued, and developed by organizations and governments seeking to establish guardrails aimed at ensuring responsible AI. The big question is whether they’ll succeed in time, given ChatGPT’s incredibly fast adoption rate worldwide.

Examples of groups working on guidelines, ethics, standards, and responsible AI frameworks include the following:

• ACM US Technology Committee’s Subcommittee on AI & Algorithms
• World Economic Forum
• UK’s Centre for Data Ethics
• Government agencies and efforts such as the US AI Bill of Rights and the European Council of the European Union’s Artificial Intelligence Act.
• IEEE and its 7000 series of standards
• Universities such as New York University’s Stern School of Business
• The private sector, wherein companies make their own responsible AI policies and foundations

How does ChatGPT work?

You’re then presented with a list of search results ranked in order of relevance as determined by the search engine’s algorithm. From there, the user is free to consider the sources of each option and click a selection to do a deeper dive for more details from that source.

By comparison, ChatGPT generates its own unified answer to your prompt. It doesn't offer citations or note its sources. You ask; it answers. Easy-peasey, right? No. That task is incredibly hard for AI to do, which is why generative AI is so impressive.

Generating an original result in response to a prompt is achieved by using either the GPT-3 (Generative Pre-trained Transformer 3) or GPT-4 model to analyze the prompt with context and predict the words that are likely to follow. Both GPT models are extremely powerful large language models capable of processing billions of words per second.

In short, transformers enable ChatGPT to generate coherent, humanlike text as a response to a prompt. ChatGPT creates a response by considering context and assigning weight (values) to words that are likely to follow the words in the prompt to predict which words would be an appropriate response.

Some ChatGPT basics here: User input is called a prompt rather than a command or a query, although it can take either form. You are, in effect, prompting AI to predict and complete a pattern that you initiated by entering the prompt.

If you'd like a comprehensive ChatGPT guide, including more detail on how it works and how to use it, check out my book ChatGPT For Dummies.

Peeking at the ChatGPT architecture

ChatGPT is now also available on GPT-4, which is a multimodal model, meaning it accepts both image and text inputs although its outputs are text only. It's now the largest LLM to date, although GPT-4’s exact number of parameters has yet to be disclosed.

Parameters are numerical values that weigh and define connections between nodes and layers in the neural network architecture. The more parameters a model has, the more complex its internal representations and weighting. In general, more parameters lead to better performance on specific tasks.

ChatGPT for beginners

Writing effective ChatGPT prompts

©John Wiley & Sons, Inc.
The ChatGPT screen on the free, research preview version

The prompt bar, which resembles a search bar, runs across the bottom of the page. Just enter a question or a command to prompt ChatGPT to produce results immediately.

If you enter a basic prompt, you’ll get a bare-bones, encyclopedic-like answer, as shown in the figure below. Do that enough times and you’ll convince yourself that this is just a toy and you can get better results from an Internet search engine. This is a typical novice’s mistake and a primary reason why beginners give up before they fully grasp what ChatGPT is and can do.

©John Wiley & Sons, Inc.
ChatGPT responds with a general outline for a business plan.

Understand that your previous experience with keywords and search engines does not apply here. You must think of and use ChatGPT in a different way. Think hard about how you’re going to word your prompt. You have many options to consider. You can assign ChatGPT a role or a persona, or several personas and roles if you decide it should respond as a team, as illustrated in the figure below.

©John Wiley & Sons, Inc.
ChatGPT responds as a team of doctors.

You can assign yourself a new role or persona as well. Or tell it to address any type of audience — such as a high school graduating class, a surgical team, or attendees at a concert or a technology conference.

You can set the stage or situation in great or minimum detail. You can ask a question, give it a command, or require specific behaviors.

A prompt, as you can see now, is much more than a question or a command. Your success with ChatGPT hinges on your ability to master crafting a prompt in such a way as to trigger the precise response you seek.

Ask yourself these questions as you are writing or evaluating your prompt. Who do you want ChatGPT to be? Where, when, and what is the situation or circumstances you want ChatGPT’s response framed within? Is the question you're entering in the prompt the real question you want it to answer, or were you trying to ask something else? Is the command you're prompting complete enough for ChatGPT to draw from sufficient context to give you a fuller, more complete, and richly nuanced response?

And the ultimate question for you to consider: Is your prompt specific and detailed, or vague and meandering? Whichever is the case, that’s what ChatGPT will mirror in its response.

ChatGPT’s responses are only as good as your prompt. That’s because the prompt starts a pattern that ChatGPT must then complete. Be intentional and concise about how you present that pattern starter — the prompt.

Starting a chat

©John Wiley & Sons, Inc.
ChatGPT user interface

Following, are some things you can do in a prompt that may not be readily evident:

• Add data in the prompt along with your question or command regarding what to do with this data. Adding data directly in the prompt enables you to add more current info as well as make ChatGPT responses more customizable and on point. You can use the Browsing plug-in to connect ChatGPT to the live Internet, which will give it access to current information. However, you may want to add data to the prompt anyway to better focus its attention on the problem or task at hand. However, there are limits on prompting and response sizes, so make your prompt as concise as possible.
• Direct the style, tone, vocabulary level, and other factors to shape ChatGPT's response.
• Command ChatGPT to assume a specific persona, job role, or authority level in its response.

When you’ve finished chatting on a particular topic or task, it’s wise to start a new chat (by clicking or tapping the New Chat button in the upper left). Starting a new dialogue prevents confusing ChatGPT, which would otherwise treat subsequent prompts as part of a single conversational thread. On the other hand, starting too many new chats on the same topic or related topics can lead the AI to use repetitious phrasing and outputs, whether or not they apply to the new chat’s prompt.

To recap: Don't confuse ChatGPT by chatting in one long continuous thread with a lot of topic changes or by opening too many new chats on the same topic. Otherwise, ChatGPT will probably say something offensive or make up random and wrong answers.

When writing prompts, think of the topic or task in narrow terms. For example, don't have a long chat on car racing, repairs, and maintenance. To keep ChatGPT more intently focused, narrow your prompt to a single topic, such as determining when the vehicle will be at top trade-in value so you can best offset a new car price. Your responses will be of much higher quality.

For example, after ChatGPT responses to Bing users became unhinged and argumentative, Microsoft limited conversations with it to 5 prompts in a row, for a total of 50 conversations a day per user. But a few days later, it increased the limit to 6 prompts per conversation and a total of 60 conversations per day per user. The limits will probably increase when AI researchers can figure out how to tame the machine to an acceptable — or at least a less offensive — level.