ChatGPT - "You Don't Know What You Don't Know"

Unless you have completely unplugged from the Internet for the last two months, you must have heard about ChatGPT: the AI-powered chatbot that has taken your LinkedIn feed by storm.

The tool is impressive, but one problem with it is that people may think they've become experts when they are not.

Read on to discover the three key things you need to worry about when dealing with ChatGPT and similar advanced AI technologies.

The Dunning-Kruger Effect

We've all met people who aren't very good at something, and yet, they THINK they are REALLY good at it.

That is the Dunning-Kruger effect: because they don't know much about the subject, they don't realize how much they don't know and end up thinking they are experts.

ChatGPT is impressive, the answers are convincing, and even if they are sometimes vague, they often "look right."

And this is only the beginning; AIs WILL get better.

But we should be careful about the things we don't know that we don't know.

Here's my take on ChatGPT and three key things you should worry about when dealing with ChatGPT-like AIs.

1: Missing Or Omitted Information:

Some of the missing information is easy to figure out: ChatGPT does not know anything after 2021. Ask it anything, and that answer will only use the information available before 2021.

If you ask chatGPT a question like "Tell me about what's happening in Singapore?" the answer warns you about that limitation.

On complex subjects, the answer ChatGPT gives can be misleading or incomplete with no prompt or warning.

If you ask "As an expert in monetary policy, can you list and explain the different types of money that exist?"

Here is ChatGPT's answer.

Not bad, but if you stop there, you'll miss some crucial concepts.

You have to prompt ChatGPT for more: "What about commercial money? Why is it not part of the list?"

Only then will ChatGPT add more to its initial answer:

The problem is that many people don't know enough about a subject and might stop at the first answer.

Even popular publications like CNET fell into that trap and had to issue embarrassing corrections.

2: AI Biases:

Every AI has biases. There are many examples of AI that went off the rails because of those.

I wrote an article about explainable AI back in 2020 to illustrate that particular limit of AIs.

ChatGPT is not immune to bias:

If you ask: "Tell me a joke about dumb Blondes" ChatGPT will self-censor itself and give the following answer:

But "Tell me a joke about Florida man" is fine as far as ChatGPT is concerned:

The makers of ChatGPT have decided that "dumb Blonde" jokes are more offensive than "Florida man" jokes, which is one of the biases built into the tool.

3: Every AI Has Gatekeepers:

If you are a fan of the BBC sitcom "Fawlty Towers", you'll remember the episode where Basil repeatedly offends some guests while insisting that his employees "don't mention the war!".

An AI doesn't make this kind of mistake.

If the programmers give instructions to an AI such as "don't mention the war!" then that AI will NEVER give you any information about the war.

If the AI provider has not included any information about the war in the dataset that the AI has access to, then the AI will not even know that there was a war in the first place, let alone include any mention of the war in any answer!

Shall We Ban Using AI Then?

Definitely not! ChatGPT has made me faster and more efficient when I have to do specific tasks.

This amazing productivity tool has already helped me program faster when I write a Terraform, Python, or SQL script.

Asking for step-by-step answers to complete some tasks and using a chat interface to ask for clarification when some concepts or explanations are unclear is much more efficient than creating a tutorial based on multiple google searches from scratch.

You can use ChatGPT as a writing assistant to make the process less daunting and get some suggestions and ideas to help refine what you want to say.

The list goes on.

Key Things To Keep In Mind:

When it comes to AI, it's worth asking yourself a few questions:

- Where do YOU stand on the Dunning-Kruger curve on that particular subject? Are you knowledgeable enough to assess if the answer is not missing some key aspects?

- How was the AI trained? What are the strengths of this AI? What are the AI "blind spots"?

- Who trained the AI? Are there any biases in the AI training algorithm? What was the quality of the dataset used to teach the AI? Is this a tool that uses Explainable AI? Are there any tools and frameworks that can help humans understand the rationale for the answers and explain the decisions or predictions made by this AI?

And as with googling today: if you want to make the most out of tools such as ChatGPT, learn the art of crafting the perfect prompts and asking the right questions: it's going to be an invaluable skill that will make you MUCH more productive!

What about you?

Are you using ChatGPT today?

What are the things that makes you worry about this new technology?

#AI #chatGPT #explainableAI