What generative AI can and cannot do

Rohan, a sales executive in Jaipur, asked a chat app to plan a three-day Goa trip for six college friends on a tight budget. In ten seconds he had a lovely plan: beaches by day, a spice farm, a cheap shack for dinner, a packing list and a cost split. Then he checked the ferry it recommended. The timing was wrong, and the shack had closed two years ago.

Both things are true at once. The plan saved Rohan an hour, and parts of it were confidently wrong. Knowing which parts to trust is the most useful skill in this whole course. This lesson gives you that map.

Imagine your office hires an intern named Asha. She has read an astonishing amount: textbooks, novels, manuals, forums, recipes, code. She writes beautifully in many languages and never gets tired. Ask her for a first draft and you get it before your chai cools.

But Asha has a few quirks. She stopped reading the news on a certain date and has not looked since. She does sums in her head and never uses a calculator. And she almost never says “I don’t know”; if she is unsure, she gives her best guess in the same calm voice. Generative AI is a lot like Asha.

Remember from lesson 1 that GenAI creates by following patterns it learned. So it does best when a good answer is mostly about language and patterns, and when “roughly right, then polished by you” is good enough.

• Drafting: a leave application, a cover letter, a reply to a tricky client email, a speech for your parents’ anniversary.
• Explaining: “Explain compound interest like I am fourteen,” then “now with a cricket example.”
• Summarising: a 30-page society meeting PDF turned into five bullet points.
• Changing style: make this message more polite, shorter, more formal, or turn it into Hinglish.
• Translating: a menu or a notice between English, Hindi, Tamil and many other languages, usually well for common languages.
• Brainstorming: twenty names for a home bakery, ten quiz questions for a school fest.
• Coding help: explaining an error, writing a small script, suggesting a fix you then test.

Each weakness has a plain reason behind it. Once you know the reason, you can predict the mistake before it happens.

Fresh facts

A model only knows what was in its training data, which stops at its knowledge cutoff. Ask about yesterday’s match score, today’s petrol price or a shop that closed last year, and it may answer from old patterns. Many apps now add a web search step to help, but the core model on its own does not know recent events.

Exact maths and counting

A language model writes numbers the way it writes words: by predicting what usually comes next. Small sums often come out right because they appear everywhere in its training text. A long bill split across six friends with GST and a discount is a different story. It can also struggle to count letters in a word, for a reason you will see in the tokens lesson. Good apps hand such work to a calculator or code tool.

Knowing when it is wrong

The model is built to produce text that looks like a good answer. When it lacks the facts, it can still produce something that looks right: a fake book title, a wrong date, a made-up court case. This is called a hallucination, and it has its own lesson in module 2.

Two more limits are worth knowing. The model does not know your private life or your company’s files unless you or the app provide them, which module 4 is all about. And it can repeat biases found in its training data, which module 8 covers.

People tend to either overtrust GenAI or dismiss it. Both cost you. Flip these cards to replace six common myths with what is actually true.