I grew up in Jhansi, a small city in the middle of India that does not show up on many maps of the future. My first computer was a second-hand machine running DOS. When the internet finally arrived, it felt like a door opening. A kid a long way from anywhere could suddenly read the same things, and learn the same things, as a kid in a much luckier place. That door changed my life. I have never quite gotten over it.
I think about that door a lot now, because AI is going to either widen it or quietly close it. And I am not sure which, yet.
Here is what I keep turning over. Every powerful technology arrives carrying two possible stories. In the first, it spreads outward — it hands ordinary people abilities that used to belong only to a few. In the second, it concentrates — it makes the people who already have an advantage even harder to catch. The technology itself does not pick a story. We do, through a thousand small decisions about how it gets built, priced, and pointed.
We talk a great deal about whether AI is biased, and we should. But there is a fairness question that sits underneath that one and gets far less attention: who can afford to use it at all? A model that is perfectly fair and costs more than a small shopkeeper earns in a month is not, in any way that matters to her, fair. The most important access question is not only what is inside the model. It is who is standing outside the room.
So I try to hold a simple test in my head when I build things. Does this give a capability to someone who did not have it before — or does it mostly make the strong stronger? A tailor who can now answer customers in three languages he does not speak. A farmer who can finally make sense of a government scheme written for lawyers. A teacher in a town with no specialists who can prepare a better lesson. Those are the doors. When AI does that, I get genuinely excited, because it rhymes with the thing that happened to me.
And then there is the other direction, which is quieter and easier to drift into. Tools priced so that only large companies can really run them. Knowledge locked behind a paywall it took the open internet thirty years to pull down. A handful of places deciding what the whole world's machines believe and refuse. None of this requires villains. It just requires everyone optimising for the obvious thing and nobody asking the inconvenient question.
I do not think the answer is to slow down or to pretend the risks are not real. I think the answer is to keep asking, out loud, who each thing is for — and to notice when the honest answer is "people who were already fine." That noticing is most of the work.
I am writing this from the Netherlands now, a long way from that DOS machine, and I am aware that I am one of the lucky ones who walked through the door. I do not have a tidy conclusion. What I have is a bias, formed early and hard to shake: that the point of a tool is to give more people more options, especially the people who had the fewest. I would like AI to be remembered as one of those tools. Whether it is depends less on how clever the models get and more on a question we can choose to keep asking.
Who is this for? And if the answer is not "more people than before," then what are we actually doing?