Paradox of AI

“A good fiction will blend a lot of reality and facts into its narration and storyline. The blend would be so fine, it would be difficult for one to differentiate fact from fiction.”

A famous Tamil writer made this remark while commenting on the popularity of Dan Brown’s The Da Vinci Code.

If one were to paraphrase, the current state of AI is just like that. It’s extremely good at certain things. It’s also extremely not so good at certain others. The challenge lies in figuring out when it’s good—and when it’s not.

Some time back, my dad wasn’t well, and I had to take him to the doctor at night. An ECG was recommended. It was taking some time to meet the doctor, so I uploaded the ECG. Voila! I had the results and a detailed reading almost instantly.

Recently, my uncle had to visit the doctor alone and came back with an X-ray. Since none of us had accompanied him, we wanted better insight into the diagnosis. I uploaded the X-ray, provided some context about him, and again—boom. It gave all the details and even asked, “Would you like me to explain this in simple terms for family members?”

Immediately, my aunt asked: “What will happen in the future? Will any of us even have jobs?”

Now, let’s switch to a different example.

I often give the New York Times Connections Puzzle to the leading AI models. It has 16 words, and the task is to group them into four correct clusters of four related words. In a sense, it’s a very simple puzzle. (Well, not that simple—at least to me—since I only have a 40% win rate.)

Nowadays, I regularly feed these puzzles to AI, and to my surprise, the win rate is often worse than mine. To be fair, the way humans solve it is quite different from how AI tackles it. But still, ideally, you’d expect AI to crack it easily.

So, it can decode an ECG, analyze an X-ray… but fails consistently at a Connections puzzle?

Herein lies the rub.

If you give AI enough sample puzzles along with the correct solutions and then ask it to generate new puzzles, it can churn out hundreds of them in no time. But even then, there’s no guarantee it would be able to solve new ones correctly 100% of the time.

So, where and how should one use AI?

Let me walk through another example.

I was once tasked with writing a report on the evolution of a domain/industry: its origins, development, current state, challenges, and future trends. (I’m very familiar with the domain and could speak about it without much prep.)

My traditional workflow would have looked like this:

1. Understand the problem.
2. Do a deep Google search.
3. Find and read at least 10–15 relevant reports or sources.
4. Highlight important points, copy-paste excerpts, consolidate notes.
5. Build a point of view.
6. Draft the narrative and create a document.

This would usually take 2–3 days and might even need two people, depending on complexity. The final report would be 5–20 pages long.

Now, with LLMs:

1. I think more deeply about the problem statement.
2. I write a good prompt, provide relevant context, and use a few prompt hacks.
3. Within 2–3 minutes, I get a 10-page draft.

I usually run the same prompt across 2–3 LLMs and get multiple versions. I then feed them all back into an LLM and ask it to consolidate.

Now, in less than 30 minutes, I’ve reached the “consolidated draft” stage.

All that’s left to do is read through it, check for hallucinations, customize the tone and structure as needed, and send it off. If I want to do a really good job, it takes just 2–3 hours. At most, 4–6 hours.

That’s the productivity gain.

Now, if you reflect on this, here are some key takeaways

1. You need to write a good prompt.
   It’s garbage in, garbage out. Writing a good prompt is easier said than done. Beyond prompt hacks, you need to really understand the problem, the workflow, and the expected output.
   You can’t automate something you don’t understand. You can’t use AI effectively if you don’t know what you want—or if you can’t instruct the AI clearly.
2. It can boost productivity like crazy.
   But again, only if you know what you’re doing and what you want.
3. Watch out for hallucinations.
   Unless you know the content cold, you might not realize where the AI has gone off the rails. Sniffing out those hallucinations requires real subject matter expertise.
4. Long story short:
   AI is a great mimicking engine.
   It is not a substitute for original thinking.
   (One could argue it helps in idea generation or acts as a sparring partner for brainstorming—and yes, it does. But even then, you won’t get the best out of it unless you know how to evaluate and refine the options. It aids thinking; it doesn’t think on its own—yet.)

As one leader put it: “It can generate 90% of an investment prospectus.”

But the crux—and the criticality—lies in the remaining 10%.

And that’s where the human element remains irreplaceable.

So, to answer my aunt:

Yes, humans will still have their jobs—at least for now.

P.S.

These days, after I write a post like this, I feed it into AI, ask it to fix grammar, smoothen the flow, and polish it while keeping the tone intact.

One of my editor friends—the kind who fixes your grammatical errors on WhatsApp—recently told me:

“Your writing has become more polished lately.”

I don’t think the writing would be this clean without AI’s help.