Multi exposure of work table with computer and brain hologram. Brainstorm concept.
ATD Blog

ChatGPT-3 Has Everyone Talking

Wednesday, March 1, 2023

In January, OpenAI’s conversational platform, ChatGPT-3, became the fastest-growing app in internet history. This is just one indication of growing curiosity about GPT (generative pre-trained transformer) and other chatbots in the past few weeks. As one of the first learning professionals to see the potential of conversational AI and one of the few designing the applications for years, I’m excited to see the recent attention paid to this engaging technology.

In my book, AI in Talent Development, I outline several ways chatbots can be used as part of a robust learning experience, and ChatGPT-3, the third version in OpenAI’s ChatGPT series, demonstrates most of them. When ATD asked me to write a post on this subject, I struggled to find the value I could add to the exploding conversation. I decided to do what people say I do best: Give you the bottom line and point the way towards practical applications. Let’s start with the basics.

How ChatGPT-3 Works

ChatGPT operates on an autoregressive model. It’s not as complicated as it may sound. A transformer is a program, model, or algorithm that converts code into something else—in this case, human language. The autoregressive part means the software uses prior events to predict the most likely future events, which is what your brain does all the time. It allows ChatGPT and similar applications to simulate a conversation between two humans. Learn the two keys to making an application like this work:

1. Natural Language Programming
Natural Language Processing (NLP) allows the bot to recognize input through human speech or writing, eliminating the need for menus or code to engage with it. This same ability enables the bot to reply through text. Before the development of NLP, most chatbots were menu-driven, and the bot responded by selecting the best fit from a database of pre-written options. Conversational AI represents a new way for humans to use computers; it’s a user interface requiring less hardware and prior knowledge than any previous UIs.

!CAUTION! Don’t confuse Natural Language Processing with the pseudo-science term using the same acronym. Neural Linguistic Programming is widely disproved and an invalid practice.

2. Massive Training Datasets
Like any neural network (including your own), AI models must be effectively trained using a technique called deep learning. Quite frankly, the stunning capabilities of ChatGPT-3 are nothing new. As David Chalmers points out, there is no new technology here. There are many bots with similar abilities, and I described them in detail in 2020. How this bot is trained and how well this approach works is innovative. First, Open AI used the largest training dataset, combining a database containing much of the internet, a huge library of books, and Wikipedia.

Then, after refining the bot in-house, Open AI released it free to the world. As we all started to play with it, we were training it further for them. Think of it as a crowd-sourced training effort. The sheer size of the training database resulted in what everyone always expected: Impressive results that can mimic human expression so well it is hard to tell a machine generated it. However, there continues to be an underlying problem: Chatbots are only as good as their training (like us humans). If you consider the source of GPT’s training, you can see why it sometimes makes mistakes.

What Can We Do With ChatGPT-3?

Because this product is still being explored, there may be new use cases by the time this post publishes. Below are a few applications that can save time and make your work as a learning professional more efficient and productive.

Write Content
If you ask the bot to write a paragraph introducing a subject, it comes back almost instantly with plausible, grammatically correct text. Because its training included unvalidated sources like the internet and Wikipedia, some of the content it produces must be corrected. Still, you can get excellent first drafts and refine your content from there.


Summarize Documents
Looking for a “Reader’s Digest” version of a book or article? Paste the text into a prompt to ChatGPT, and it returns a nice summary that should hit most of the highlights.

Write and Troubleshoot Computer Code
Because ChatGPT was built to analyze and produce output in human language, it is also great at computer languages. Programmers are using it to troubleshoot code. It can even generate new code if you ask it to develop a new capability.

Translate Documents
Far more effective than other translators out there, ChatGPT can take a website or article and translate it into another language.

Write and Grade Quizzes
You can feed your content into GPT and ask it to generate a quiz with answers. Or, give it a general subject, like world history, and it will do the same. Feed your student’s quiz responses into GPT, and it will grade them (even essays!).

Create Games
I asked GPT to create a game teaching sixth-grade students about the solar system, and the bot came back with a fun adventure through space.

Improve Your Writing
I didn’t do this for this article, but I could have asked GPT to review my draft and suggest improvements.

Conduct Research
I asked the bot to provide a list of current peer-reviewed studies on chatbot use in education using APA style and received a properly formatted list.


Why Does ChatGPT Have Some People Worried?

While most of the comments about GPT are positive, you will find a few educators deeply concerned about this new tool’s capabilities.

Mimicking Human Writing
Teachers are worried about GPT’s ability to write plausible text reading like a human created it and write in a known author’s style. A student could ask the bot to write a paper in the same style as one of their previous works, and it would be hard to tell the GPT paper wasn’t an original student effort. But, again, this mimicking ability is nothing new. AI has generated art in the style of great artists for years. ChatGPT applies the same predictive approach to language.

ChatGPT Makes Mistakes
Another concern is the bot’s ability to present false information so convincingly that some users forget to validate the response with research. For example, I asked the bot to calculate the mean of five numbers, and the result was interesting. I received a detailed explanation of the statistical definition of “mean” and how to calculate it, which I did NOT request. However, the bot added the five numbers incorrectly.

To be fair, ChatGPT wasn’t built to perform math problems. But it didn’t tell me that or report the result might be incorrect. Like any willing assistant, it did its best to provide an answer. If I had been a student relying on GPT’s help, I would have been wrong. The reason for most of these mistakes is the same reason that it is such a stunning improvement in the chatbot experience: It’s primary sources of information are the internet and Wikipedia.
I asked Vince Han, one of the pioneers in building chatbots for learning, to sum up the challenges presented by GPT’s new capabilities, and here’s his response:

One of the effects of ChatGPT & increased quality in Artificial Intelligence in general is shifting the skills users need. For example, most users have developed a skill for how to search which is knowing how to craft keywords and how to sift through search results. Now, people will need the skill of discernment, or learning how to decipher truth and views from a chatbot that provides what reads like a definitive answer. So, from a performance support perspective, yes, one can now get an immediate response to a question but how accurate and relevant that response is will be up to them to determine.

As educators, we must now confront a problem that has been boiling for decades—the decline in critical thinking across most domains, demographics, and workgroups. In my opinion, ChatGPT makes this gap more visible and increases the urgency for closing it quickly.

Closing Thoughts

Because Microsoft owns the license for the exclusive use of the model, we’ve all been working for them, revealing exciting use cases that may find their way into AI-enabled Microsoft products. For now, the platform consists of both free and professional versions. It’s an excellent way to begin learning about the promise of AI and our place alongside it. As machine learning expert Santiago Pino says, “AI will not replace you. A person using AI will.”

Where Can You Learn More?
There are so many articles about this subject right now that I’ve limited myself to a few favorites:

About the Author

Margie Meacham, “The Brain Lady,” is a scholar-practitioner in the field of education and learning and president of LearningToGo. She specializes in practical applications for neuroscience to enhance learning and performance. Meacham’s clients include businesses, schools, and universities. She writes a popular blog for the Association of Talent Development and has published two books, Brain Matters: How to Help Anyone Learn Anything Using Neuroscience and The Genius Button: Using Neuroscience to Bring Out Your Inner Genius.

She first became interested in the brain when she went with undiagnosed dyslexia as a child. Although she struggled in the early grades, she eventually taught herself how to overcome the challenge of a slight learning disability and became her high school valedictorian, graduated magna cum laude from Centenary University, and earned her master’s degree in education from Capella University with a 4.0.

Meacham started her professional career in high-tech sales, and when she was promoted to director of training, she discovered her passion for teaching and helping people learn. She became one of the first corporate trainers to use video conferencing and e-learning and started her own consulting company from there. Today she consults for many organizations, helping them design learning experiences that will form new neural connections and marry neuroscience theory with practice.

“I believe we are on the verge of so many wonderful discoveries about how we learn. Understanding what happens in the brain is making us better leaders, teachers, parents, and employees. We have no limits to what we can accomplish with our wonderful brains— the best survival machines ever built.”
—Margie Meacham

Sign In to Post a Comment
Hi, Margie.
After reading this, I wondered if there are other AI bots that have been trained using a better or more "valid" source of truth? As you said, we are training this particular bot for Microsoft and Wikipedia/ internet is part of its source. Are there other options (now or in the works) with which we should/ could spend our time?
Good point and thank you, Margie. Now to ask the bot for more time:) ...
Sorry! Something went wrong on our end. Please try again later.
Absolutely, Brian! I've been building custom, rules-based bots for several years now. The key is to use the bot to make your own, validated content more searchable and digestible for learners. For example, you might want to have a product training bot or a new employee onboarding bot. Because you are restricting responses to your own, vetted content, you can rely on the results.
Sorry! Something went wrong on our end. Please try again later.
Sorry! Something went wrong on our end. Please try again later.
Good article. It's not necessarily new technology, as you stated. Both individuals and marketers have been tinkering with similar prototypes[ for a while, so it's not a surprise that one of these tools would be wildly successful. There is a lot of content in the marketplace pulled from tools without a fact-check or validation, which bothers me. I think it's our job to find new ways to verify critical thinking in learners that recognizes that AI maybe part of the process, but they own the outcome
I share your concerns, Elizabeth. As I mention in this post, one way to address this issue may be to heighten our efforts to teach critical thinking in business and education curricula. Critical thinkers are needed to evaluate the outputs of these "black boxes" so that they can use them responsibly.
Sorry! Something went wrong on our end. Please try again later.
Sorry! Something went wrong on our end. Please try again later.
Very well done Margie, excellent and highly useful summary. And no, I didn’t have my comment created by a bot 😄
LOL Bob! The truth is, we may need to start labeling our work as either human-produced, bot produced, or bot-assisted. Wired magazine just published the policy on how they will use AI, and they have pledged not to use them to write articles in the magazine, although they will be using them for research, proofreading, and other tasks. The era of having an AI co-worker is definitely here!
Sorry! Something went wrong on our end. Please try again later.
Sorry! Something went wrong on our end. Please try again later.
Sorry! Something went wrong on our end. Please try again later.