Thinking out of the box with AI

Hi there,

I often see people complain about AI giving wrong answers. "How can we be sure it's not hallucinating?" is a common question. Well, as AI author Andriy Burkov rightly points out: You can't.

Turns out, if you're working on a use case that requires 100% reliability, it might not be a good use case for AI.

But are all AI use cases inherently flawed? No, in fact, the principle that AI doesn't know what's true can be one of its greatest assets.

Let's find out why!

New Book Announcement

Thrilled to announce the early release of my new book - Augmented Analytics! Together with my co-author Willi Weber, Head of Analytics at HDI Global SE, we'll explore how to cross the analytics chasm and use AI and automation to transform into an insight-driven company!

Read the early preview now.

There is no truth in AI yet - which is ok

Modern AIs struggle with the concept of truth:

• Gemini struggles to get historic facts "right"

• Microsoft Copilot hallucinating team members

• AI algorithms identified everything but COVID

While most of us would consider this a flaw, it's actually something deeply rooted in AI. AI systems aren't logical truth-finders.

AI systems are probabilistic machines.

And these probabilities are created by patterns in the data (which is often called "bias".) In the neural networks of an AI, the concept of "truth" doesn't exist. If there are many examples in your data that suggest X leads to Y, the AI will suggest X if you want Y as an outcome.

Simply put, AI tries to exploit existing patterns in data to optimize for a particular outcome.

• In ChatGPT, that outcome is predicting the next token (word).

• In time series forecasting, that's guessing the next time step.

• In churn modeling, it's flagging whether a customer will leave or not.

And on and on.

Ultimately, what all these systems do is give you probabilities for a desired outcome. And sometimes that outcome is clearly defined (e.g., whether or not a customer has churned in the past - it's measurable). And sometimes that outcome is fuzzy (whether we like the words ChatGPT predicted or not).

It's up to us to decide what's right or wrong, true or false, useful or useless.

When AlphaGo broke the rules

Here’s a little story to illustrate the point:

Back in 2015 (when Google was still leading the frontier of AI innovation), the search giant’s AI lab DeepMind developed AlphaGo, an AI designed to play the ancient board game Go.

The game of Go is known for its complexity and strategic depth, where it's virtually impossible to anticipate every next move (unlike less complex games like chess).

In 2016, AlphaGo faced off against Lee Sedol, one of the world's top Go players, in a landmark match. Unlike humans, AlphaGo didn't discern between traditional 'right' or 'wrong' strategies; it learned and applied moves based purely on data-driven insights.

In a stunning turn of events, AlphaGo defeated Lee Sedol, utilizing moves that puzzled and amazed the Go community. One move, in particular, Move 37 in the second game, was so unconventional that it shifted the understanding of viable strategies in Go. And let to AlphaGo winning the game.

This event underscores a crucial lesson: AI's 'ignorance' of right and wrong can lead to innovative breakthroughs.

Leverage AI randomness for your business

By analyzing patterns without human directions, AI can uncover solutions and strategies that revolutionize fields and challenge human preconceptions.

In business, if we embrace the fact that AI has different bias than us and is able to think out of the box, that’s actually something we can use for so many things.

Here are some examples:

Sales and Marketing

Content Marketing: Guide content strategy in directions that deviate from industry norms but are highly effective for the specific audience in question.

Product Recommendations: Uncover unique buying patterns to suggest products that customers didn't even know they wanted.

Lead Scoring: Find patterns of engagement that correlate strongly with conversion but are overlooked by human marketers.

Churn Prevention: Reveal counterintuitive insights, such as specific product features or interaction points that are more critical to retention than might be assumed based on conventional wisdom.

Customer Experience

Chatbots: Suggest resolutions that humans might not think of because they can't find connections between different pieces of information.

FAQs: Spot recurring questions or problems raised by customers that have previously been overlooked in your FAQs.

Product Development: Create novel features by analyzing customer feedback, revealing unexpected patterns of user needs and preferences.

Data Analytics

Anomaly Detection: Spot patterns that would be too subtle or complex for manual detection.

Forecasting: Identify trends before they become apparent through traditional analysis.

Optimization: Propose novel optimization algorithms that improve efficiency in ways humans hadn't even considered.

In all of the scenarios above, we can control the behavior and output of the AI by shifting the data.

That's why data is king in AI.

Conclusion

AI's inability to determine absolute truth is not necessarily a flaw, but can be a powerful feature to augment or even overcome human bias.

By operating on probabilities and patterns, it can unlock innovation and new solutions in areas ranging from sales and marketing to customer experience and data analytics.

This shows the enormous potential of using AI when you accept it as a probabilistic creature.

When managed with the right data, this will lead to better AI success.

See you next Friday!

Tobias