What 99% Don't Know (But Should Know) About How Generative AI Answers

A big thing I notice in my work is the complete lack of understanding of how Generative AI technology works. Specifically, the science behind why AI makes up stuff.

I'm not shocked by this.

After a decade plus in the learning tech industry, it's a common problem. Customers can't use tools and tech to its full potential because they don't understand the basics of how they work. While it might not be essential with all technology, Generative AI is one where it does count.

For those who’ve taken my?crash course, you know I focus on this as an essential block of knowledge.

The lack of understanding of why this happens and its unintended consequences leads to a lotta mistakes when applying tools at work.

Let’s fix that ↓

Why AI makes up stiff (aka hallucinations)

Not enough people understand that generative AI is a probabilistic system.

What does that mean?

Your favourite conversational AI tool is a probability engine.

Generative AI systems are prime examples of probabilistic models. When prompted, these models do not simply retrieve a pre-written response.

A probabilistic system uses data and patterns to make guesses about what might happen next.

It doesn’t always give the same result because it considers different possibilities and chooses one based on how likely it thinks that outcome is based on its training data.

How Generative AI creates answers

? Here’s how it works:

1. Learning from data: The model looks at a lot of examples (like reading many books or looking at many pictures) to understand patterns and relationships.
2. Making predictions:?When you ask ChatGPT to do something, it uses what it learned to guess the best answer. It doesn’t follow a strict recipe. It considers different possibilities and picks one based on what it thinks is most likely to be correct.

For example, if you ask a generative AI tool to write a story, it will use patterns from all the stories it has read to create a new one.

The story might be different each time you ask because the AI is making educated guesses based on probabilities.

This probabilistic sampling allows generative AI to produce varied, open-ended outputs like paragraphs, images, or code.

Yet,?it also means the outputs can be inconsistent?or contain hallucinations since the model is essentially “guessing” the most probable output.

This is both a strength and a limitation of current generative AI technology.

→ It’s not a bug, it’s a feature.

Historically, search engines like Google have been deterministic systems. They use methods to find relevant information and give you the same results for that query.

You can’t use generative AI tools intelligently if you don’t understand this.

There’s a reason companies are worried about full-scale deployment of generative-powered assistants.

In sum:?All these trendy AI tools have limitations. Be clear on these so you can use them intelligently.

Why this is important for L&D solutions and products ‘powered by AI’

Now I’ve pulled back the curtain on the matrix, you are better equipped to navigate how to use these tools in the L&D space.

Before you ask…no. It doesn’t matter how good your ’prompting’ skills are.

The capability for generative AI models to make up stuff is hard coded. As we’ve covered, it makes them both great and bad. If you go into using and building tools with AI with this at the top of your framework of?‘to use or not to use AI for this task’?you’ll do well.

A common example of mishaps can be seen with the thousands of chatbot solutions which have flooded the market.

Some of the worse offenders, imo, are the ‘coaching’ and specifically mental/health and wellbeing assistants. With what you know now:

1. Do you believe these areas are suitable for Gen AI support?
2. Are you comfortable with the 10% potential of bad advice?

Like any technology, Gen AI has it’s time and place for use.

Clarity on this = better solutions = improved performance

Final thoughts: TL;DR???

• What:?Gen AI systems are probabilistic models that generate varied outputs based on learned data patterns and probabilities rather than fixed responses.
• Why: The ability to produce different results each time can lead to inconsistent outputs. Knowing this equips you to pick the right time to use Gen AI tools.
• How:?You need to know these limitations to use AI intelligently.

?? Learn how to maximise tech for learning and performance

→ Over 3,000 L&D pros enjoy my Steal These Thoughts! newsletter every week.

Join us for the best breakdowns on leveraging tech for learning and performance with in-depth, research led insights. This is for the long form 'show me' fans.