AI does not need to become smarter than everyone to help millions of people do things they previously couldn’t.
Much of the AI debate seems caught between two questions: are we approaching superintelligence, or are we heading towards an expensive disappointment?
We argue about scaling, scientific breakthroughs, the spending of hyperscalers and whether the economics will ever justify the investment.
These are legitimate questions. But I think they leave something important out.
The economic value of AI does not depend only on how intelligent it becomes. It also depends on which human limitations it helps us overcome—and how many people it helps.
Communication might be one of the most consequential. And one of the most undersold.
The world is not an intelligence leaderboard
Intelligence matters. Scientific discovery matters. I am not arguing otherwise.
But the world does not distribute wealth, influence or opportunity according to a ranking of intellectual ability. Clever people do not automatically run businesses, win elections or get their ideas adopted.
Knowledge has to connect with trust, organisation, resources and other people before it becomes useful at scale.
That is why I find the fixation on AI’s intellectual achievements incomplete. Solving a difficult mathematical problem demonstrates a capability. It does not, by itself, tell us where the technology’s broadest economic impact will come from.
A system does not have to outperform the world’s best scientist to make an ordinary person’s day substantially easier.
AI has already made money by understanding us
We should also remember that AI’s commercial story did not begin with chatbots.
Long before the current language-model wave, machine learning was helping platforms select recommendations, target advertisements and predict behaviour. Meta, for example, describes AI as foundational to its discovery engine and advertising business, predating its generative AI products. Source
Those systems did not need to hold an impressive conversation. They needed to improve a commercially important decision, repeatedly and at scale.
In a sense, they were already “hacking” us: learning what caught our attention and using that understanding to influence what we watched or bought.
What interests me now is whether we can turn that relationship around.
Instead of AI helping a platform understand us better, can it help us understand one another better?
Beyond people who already work with words
The current generation of AI is an obvious fit for programmers, analysts, writers and people whose working day already happens inside documents and software.
But that is not the whole world.
Adoption is broader than that stereotype suggests. A recent NBER working paper describes workplace generative AI use as “widespread but shallow”: it appears across many occupations and tasks, but adoption varies substantially even among people doing similar work. Source
So I would not say the technology has failed to reach everyone else. Rather, I do not think we have fully realised what it could do for them.
More generated content is not the same as more people gaining useful capabilities.
What about a tradesperson explaining a complicated repair? A parent communicating with a school in an unfamiliar language? A small-business owner trying to resolve a misunderstanding with a customer?
These are not necessarily problems of insufficient intelligence.
They are problems of expression, context and understanding.
My agent talks to your agent
In “10 Bits Per Second,” I wrote about the bottleneck between fast computers and our limited ability to think, type, speak and act. But speed is only part of the problem. Getting words out does not guarantee that the intended meaning gets across. Source
I often find it easier to explain myself to a computer than to another person. I can give it a messy brain dump without managing someone else’s patience or anticipating every possible interpretation.
My recent interview experience reminded me how different human communication can be. Working through a problem and making my thinking understandable to someone else require different kinds of effort. Source
Now consider a workflow where my agent helps me communicate with yours.
I explain what I mean naturally. My agent helps organise it and supply relevant context. Your agent helps present it in a way that fits what you need to understand.
Perhaps I start with implementation details, but you need the business consequence first. Perhaps we speak different languages. Perhaps I have assumed something you do not know.
The agents help bridge those gaps.
Not by changing my intention or manipulating your response, but by helping us make the meaning accessible.
That is the hack: I gain some of the benefits of being a better communicator without having to master every communication skill myself.
Useful does not have to mean superhuman
This does not require a model capable of making a Nobel-worthy discovery.
Many parts of the job—restructuring an explanation, translating a message, identifying ambiguity—can be tackled with capabilities already available. The challenge is making them reliable, affordable and easy enough to use in ordinary life.
There are risks. An agent might alter my meaning, conceal uncertainty or make a weak argument sound convincing. We still need control over what is communicated, and opportunities to correct misunderstandings.
But those are reasons to design this carefully, not to overlook it.
I am not dismissing the pursuit of more capable AI. I am questioning the assumption that the most intellectually impressive capability must also be the most broadly valuable.
We could build an extraordinarily intelligent machine that relatively few people know how to use well.
Or we could make everyday communication substantially easier for millions of people.
The second possibility might attract fewer spectacular headlines. It could still change who gets heard, who can participate and who can get things done.
AI does not have to become superhuman before it can give us abilities we did not previously have.
Leave a Comment