Understanding What AI Can and Can't Do

Generative AI like the ChatGPT chatbot and DALL-E image generator don’t create brand new content, but merely repackages and blends existing content that they have been trained on.

Having led data science teams at LinkedIn for 12 years and built countless data science predictive analytics models, I’ve seen the successes and limitations of AI.

I want to help you understand the capabilities and limits of the AI technologies so that you can be better equipped in your career.

The Truth About Generative AI

First, let’s clarify a common misconception about AI, particularly generative AI. Contrary to popular belief, generative AI doesn’t truly create new text or image content from scratch. Instead, it creates outputs which best fit existing patterns from the training set data.

In other words, it rehashes existing content per detected patterns rather than totally innovates from scratch.

Generative AI tools like ChatGPT were never designed to create brand new, never seen before, patterns. Gen AI is always trained to form conventional results. In other words, the AI tries to shape the outputs to fit toward the “average” instead of toward the extremes.

But creative innovations don’t come from pursuing the average.

Can Generative AI Innovate?

To illustrate this, I conducted a simple experiment comparing the creativity of ChatGPT with that of my son.

I asked both to invent "ten new words not found in any language or name." This type of task is sometimes needed to create new product names and website domains, like with drug company product marketers.

I tested the proposed new words generated by both my son and by ChatGPT to see if they exist in Google's massive search index. I used double quotes in my search term, with a logged-out private-mode web browser with VPN. If Google finds zero results, then it's likely that this word doesn't exist anywhere on the internet.

Surprisingly, my son’s creations were remarkably uniquely original, with 9 out of 10 words yielding zero results from Google’s massive search engine—demonstrating a 90% success rate in complying with my instructions.

In contrast, ChatGPT struggled, with only 2 out of 10 words being outside of Google’s search index, for 20% compliance. Some of the "invented" words even had thousands of hits.

I did the same experiment with ChatGPT (both v4o and v3.5), Microsoft Copilot (both GPT4 and GPT3.5), Google Gemini, and Anthropic Claude. All of them failed miserably against my instructions.

This was an opportunity for intentional AI hallucinations to shine, but the AI was unable to exercise such creativity on demand.

Why did my son vastly outperform ChatGPT and other LLMs in this seemingly simple task?

As intelligent as he is, I don’t think my son’s brain is filled with the trillions of parameters that GPT-4 was trained on.

It comes down to how AI operates—it identifies and reproduces patterns from its training data, whereas human creativity transcends these patterns, often producing genuinely novel ideas.

Classical AI Models Also Aim For the Average

It’s not just new Generative AI that suffers from pursuing average results. What I call “Classical AI” also has these limitations.

Let’s explore an example of Classical AI.

Imagine you’re a data scientist analyzing customer behavior in a cafeteria setting. You want to determine which beverage to cross-sell to each customer who orders a main entree.

To train the predictive model, the data scientist first gathers extensive historical transaction data, including details like the customer’s entree choice (“Meat & Potatoes” or “Poke Bowl”), the day of the week, the month, and even augmented data such as weather conditions at the time of purchase. Additional insights from loyalty membership data provide demographics like age, gender, smartphone brand, and postal zip code—all which can be used to infer customer preferences.

We may infer that customers selecting the hearty “Meat & Potatoes” entree often pair it with a “Vanilla Milk Shake,” whereas those opting for the lighter “Poke Bowl” tend to favor “Matcha Green Tea.”

This isn’t deep magic—it’s pattern recognition at work. AI excels at spotting these patterns and making probabilistic predictions based on them.

Although this trivial case here with only two entrees is simplified for conceptual understanding, AI machine learning techniques would be hugely beneficial to scale the analysis and pattern detection with real-life data.

Example Support Vector Machine (SVM) in Python

The data scientist can write a few simple lines of code to generate the predictions or classifications, of which beverage a customer would likely purchase, given the entree order.

Here’s a Python code example using SVM, or Support Vector Machine, a popular machine learning algorithm.

SVM is used here to classify which beverage a customer is likely to choose based on their entree selection and other relevant data points. SVM finds the optimal hyperplane that best separates different classes of data (e.g. vanilla milkshake customers vs matcha ice tea customers) as in the below illustration.

You can see how the data scientist, with or without an AI “copilot” coding companion, can simply write a few lines of code to build this predictive classifier model. In practice, there's quite a bit more work needed to clean the data, evaluate the model, and tune it.

Predictive Models Are Limited to Stereotypical Averages

While AI excels at pattern recognition and making data-driven predictions, it’s essential to recognize its limitations. For instance, not all customer behaviors can be accurately predicted solely based on historical data patterns.

The above cafeteria beverage purchase predictive model would fail for the case where the customer recently got a doctor’s order to change dietary habits, that the model had no awareness of.

The predictive model would still pursue “average” results almost like “stereotyping,” rather than pursuing the outlier cases.

Sometimes, we want to focus on the extreme outliers for the most interesting discoveries and innovations.

ChatGPT Performs Autocomplete Towards Typical Responses

It’s important to note that new Gen AI tools like ChatGPT operate slightly differently.

Large Language Models look for patterns from its training data to essentially autocomplete the next word in a sentence, one token at a time. This process allows it to generate responses that are coherent and contextually relevant based on the input it receives.

However, like other forms of AI, ChatGPT is limited by the patterns it has learned and cannot truly create novel content or ideas beyond its training. If ChatGPT produces unique but truly random output, then it becomes useless. If it follows the conventions and expectations from the trainers, then it's unable to produce innovative content.

While my son can use his imagination to fabricate brand new words quite simply, a ChatGPT user would need to go through a chain of extra engineering steps to filter for the desired results, with a lot of pricey API calls going to waste.

We still need humans to prioritize what are truly innovative creations versus random junk.

Conclusion

While AI is a powerful technique to perform decision-making and generative tasks at scale, it lacks the true creativity and innovation that humans possess.

As we move forward, skills like creativity, critical thinking, and intuition will remain invaluable, complementing AI’s capabilities rather than competing with them.

In conclusion, AI offers tremendous potential to enhance our lives and careers by automating tasks, finding patterns, and providing insights based on data. By understanding its strengths and limitations, you can leverage AI to make informed decisions and drive innovation in your field.

Remember, AI is a sophisticated technique—but it’s our creativity and ingenuity that will continue to push the boundaries to drive progress and shape the future.

Let me know what you think... agree or disagree?

Best regards,
Jimmy Wong

Jimmy Wong Coach, speaker, and entrepreneur enabling people to thrive in the age of AI. Data science leader with 12 years experience at the LinkedIn company and 27 years in the industry. Visit aijimmy.com​ ​​​​​​

​