ChatGPT told me so.

"I asked but it said that there is no point venturing into apparel, the better suited option is to open a bakery," the young man behind me said into his phone. We were in a slow queue boarding a flight... the kind of slow that only IndiGo manages, where the first call is already the "last and final call" and you are trhough security and standing in the aerobridge wondering how it is possible to be both at the gate and nowhere near the aircraft.

The next few sentences were also about business ideas, and from his side of the conversation, it was clear that both of them trusted the advice they were discussing. There was a third party whose guidance they were weighing seriously. A mentor, I assumed. Someone who had been in business. Someone who knew the market.

It was almost a full minute before I realised the trusted party was ChatGPT.

The conversation continued for the 10 to 15 minutes it took us to board. And trhough those 15 minutes, I watched three things happen in real time that I had previously only described in theory. I had written earlier about what I called the three characters of AI: IKIA (It Knows It All), the Yes Man, and the Eloquent Speaker. What I saw in that aerobridge was all three, working together, on one person, in one conversation.

Actually, let me stop here and be more precise about what I mean. Because I am not saying the advice was necessarily wrong. A bakery might genuinely be a better business than an apparel store for certain cities, certain budgets, certain people. What I am saying is something different: the trust was disproportionate. The kind of trust you extend to someone who has skin in the game. The kind you reserve for a person who has failed at something similar and come back to tell you why.

The FirstCry franchise observation is reasonable, actually. Mother and child care spending is increasing. That is a real trend; India's organised baby care market crossed Rs 18,000 crore last year and is growing. But a car detailing store in a mall. In Delhi. I am not sure ChatGPT has ever spent time looking for parking in Saket or tried to get a vehicle into a basement in Vasant Kunj. Delhi does not have space even in the basement, and no user is particularly interested in exploring a basement to get their car cleaned. That is my personal opinion. But it is a personal opinion that comes from having lived here.

ChatGPT has not lived anywhere.

Here is what makes this combination: IKIA, the Yes Man, the Eloquent Speaker; particularly effective on a certain kind of person at a certain moment.

Put those three together and you have something that feels like a trusted advisor but has none of the properties that make advisors trustworthy. An advisor has a track record. An advisor has something at stake. An advisor will sometimes look you in the eye and say: this is a bad idea, I have seen this fail three times, do not do it. ChatGPT will not do that. It will offer you the balanced perspective and let you decide, which sounds respectful but is actually a form of abandonment at the moment you most need direction.

My concern is not that people are using AI to think through business ideas. That is fine. I use it to think through things too. My concern is the level of trust. The black box that says it understands you, never says no with conviction, and mesmerises you with the language... that is a particular kind of dangerous. Not dangerous in a dramatic way. Dangerous in the way that a compass is dangerous if you trust it in a city full of magnetic interference. The readings are not random. They are just slightly wrong in ways that accumulate over distance.

What we need to understand is what we understand and what we do not. We know that AI is trained on patterns. We know it does not know your city's basement parking situation. We know it has not failed at a business in Delhi in 2019 and had to tell told me the story of why later, quite casually, over tea. That failure, that story, that cup of tea, is the thing that would actually inform good advice. ChatGPT does not have it.

I am not saying do not use it. I am saying know what you are using. A bakery recommendation from a language model and a bakery recommendation from someone who ran one in Lajpat Nagar for three years and eventually closed it are not the same recommendation. They are the same sentence. They are not the same thing.

I keep coming back to that aerobridge conversation. Which usually means I have not resolved it. I think what unsettled me was not the young man's trust in the tool. It was that the trust looked exactly like the trust you place in a person. The nodding. The "yes, yes" of someone receiving confirmation. The way he told his friend on the other end of the call, with a kind of settled confidence: "It said the idea was nice."

It said the idea was nice.

That is probably the sentence I will be thinking about for a while.