Action: Generative AI and Its Mistakes

So far, the AI we have looked at makes predictions and sorts things into groups. There is another kind of AI that has been in the news a lot: AI that creates.

Activity 8: Is it generative AI? (3 minutes)

Tools that write text, make images from a description, or compose music are all generative AI. But not every AI is generative.

How does generative AI create text?

Generative AI uses large language models (LLMs) to predict what text should come next. LLMs analyze vast amounts of text, such as books, news articles, song lyrics, poems, conversations on messaging apps, and emails, to find patterns.

Here is the surprising part. When you ask generative AI "What is 2 × 2?", it says 4 because it has seen lots of examples of equations, and most of the time when it saw 2 × 2 it saw the answer 4. A calculator actually does the math. Generative AI just says "4 is what I usually see, so I will say 4 because it matches the usual pattern."

Does generative AI make mistakes?

Yes! Imagine your teacher asked generative AI to write your report card. The AI was trained by analyzing lots of other report cards, including some from other countries such as the United States. Students in the United States usually have American History on their report card, so the AI might add American History to yours, even if you never took that subject.

And just like the cat AI that could not tell dogs from cats, an LLM can make predictions we don't expect because we don't know exactly what it was trained on. An LLM trained only on Shakespeare's plays would suggest very different words than one trained only on text messages.

Activity 9: Spot the hallucination (3 minutes)

Read this made-up exchange with a chatbot.

Question: Tell me about the first Canadian to walk on the Moon.

Chatbot: Marc Garneau became the first Canadian to walk on the Moon in 1984 during his mission aboard the space shuttle Challenger.