CHAPTER FOUR

The data problem

In the last three chapters I told you about a pattern. An inherited disposition. Carriers of it, across twenty years of working life. Scenes in kitchens and on platforms and in a spare room and over Slack, none of which were ever written down. In twenty-five years of business there have been plenty of Grahams too. The pattern does not stop. Stopping would be a different life.

Now I have to make the argument explicit. I have been putting it off, across three chapters, and the reader who has got this far has earned it.

The point does not require a private family story. Consider instead three ordinary acts: making room for someone, noticing an unexpected kindness, and calling a friend who will understand why it mattered.

An invitation can turn on a moment nobody records. A person hesitates, then makes room. The public record might show that someone attended. It will not show the hesitation, or the judgement that followed it.

A small room at the back of a pub may be full of energy and hope, while almost none of its life is visible from outside.

A small act of generosity can travel too. Someone sees another person help a stranger, remembers it, and does something similar later. No lesson has been announced.

Then there is the call to a friend who understands. The point of the call is not information. It is recognition: something happened, it mattered, and another person receives it.

That is the thing the chapter has to land.

Knowledge moves through relationships in both directions. The younger person can teach the older one. A disposition can be returned to the person who first modelled it.

It is not a download from elder to child. It is a loop between people, with nobody quite knowing who taught whom.

The event may leave a trace. The judgement, trust and recognition often do not.

What would a machine trained on the public record find in these examples?

It might find a pub as a pin on a map, with a rating and a handful of reviews. It would find little about the life inside the small room at the back. That is the known known.

It might infer some adjacent facts: that people gather, that conversations happen, that more is going on than the internet records. Those are gaps whose edges can be seen.

What it cannot see is the judgement beneath the action: the hesitation before an invitation, the remembered kindness, or the reason someone picked up the phone. Those are the layer this book is trying to name.

These are not missing because they are paywalled. They are not missing because they are in the wrong language. They are not missing because a licensing deal has not yet been done. They are missing because, in the conversation about what data the machines should be trained on, nothing in the conversation has a shape that would receive them.

There is no team at any lab whose job it is to collect the oh no that happens in private the second before a disposition overrides it. There is no dataset called phone-calls-to-other-carriers-when-something-lands. There is no budget line for why-nows. The conversation does not have the category. Because the conversation does not have the category, no amount of working hard on data collection — in the sense in which data collection is currently understood — will produce the thing.

The pub as a Google pin is the public record. The rough statistical shape of what happens inside places like the pub is what a machine might infer from it. Every specific judgement by every specific person that makes each of those places what it is, for each of the other specific people in it, across each of the specific relationships they have, over each of the years they have had those relationships — that is the layer underneath. That is the layer. It is not missing because anyone decided to leave it out. It is missing because nobody has yet said, inside the rooms where these decisions get made — that layer exists, it is bigger than any of the layers we are currently working on, and we do not know how to collect it.

This book is trying to get that sentence into those rooms.

I want to say something about why the labs, specifically, have not seen this.

This is not an accusation. The people at the labs are some of the most thoughtful people working on anything right now. They are not missing this layer because they are careless or because they do not care. They are missing it because of who they are and where they have been.

The labs are full of people who went, in their teens, to selective schools. Who went, in their early twenties, to elite universities. Who went, in their mid-twenties, into PhD programmes or into the companies that became the labs. They are brilliant and serious. They are also, by the selection funnel they have been through, people for whom the public record has always been where the action was.

Imagine the funnel from inside it. A clever child, encouraged at school. Doing well. Sent on to better schools, where doing well was the whole of what was asked. Head down. University. A subject. Then a narrower part of that subject. Then a narrower part of that. The knowledge gets deeper. It also gets narrower.

Their education did not lead through every room where this knowledge is carried. The funnel led past many of those rooms.

They grew up inside the part of human knowledge that writes itself down. Their teachers wrote books. Their supervisors wrote papers. Their colleagues write posts. The work they did to get to where they are was, mostly, legible work — work that was read and evaluated and rewarded by other people writing things down.

The kitchen layer of human knowledge was not, for them, invisible. It was just not the layer they were selected against. The thing they were selected against was producing publishable output. They produced it. They were rewarded for it. They now work on a system that is being trained on what they and people like them produced.

The disposition I have been describing — the silence, the not-writing-about-it, the private exchange, the oh no overridden in the second before anyone knew it was being overridden — is not the disposition the labs selected for. The labs selected for the opposite. The people who carry the silent layer are, almost by definition, not the people the labs hired. The hiring pipeline filtered them out, without malice, without noticing, because the pipeline was looking for something else.

This is not a problem with the labs. It is a feature of the pipeline that produces the labs. Nobody designed it. It is emergent. It means the people who could most easily see the missing layer are the people least likely to be inside the buildings where the decisions about the training data are made.

The small room at the back of a pub is full of young people, in energy and hope, hidden from view. No machine trained on the public record knows what that room holds. The people inside it do. The people in the rooms where the training decisions get made, mostly, do not.

That is the gap. That is what this book is about.

I said, at the start of the chapter, that the layer is missing for reasons that are not about malice or neglect. I want to be specific about what kind of problem this is. The diagnosis matters for the fix.

It is not a problem that will be solved by more data of the kind the machines already have. More scraping of more public forums, more translation of more non-English text, more digitisation of more archives, more deals with more publishers — these are all worth doing. They are all known-unknowns work. None of them touches the layer this chapter has been describing.

It is not a problem that will be solved by better algorithms. The algorithms are not the bottleneck. An algorithm that has not seen the layer cannot invent the layer. It can only pattern-match on the adjacent things it has seen, and produce a competent version of the wrong thing.

It is not a problem that will be solved by asking the models. You cannot ask a model what it does not know. You can only notice, by working alongside it, the places where it is missing. Most people working alongside these models are not carriers. Most of them will not notice. The model will feel to them like it is doing fine. The output will be competent. Nothing will seem wrong. The carriers, in their kitchens and on their streets and at their pub doors, will be producing their short silent exchanges that nobody is logging.

What I mean by a data problem is narrower and harder. The missing layer is not a mystery. It exists. It is carried inside living people. It could, in principle, be collected. It has never been collected at any meaningful scale. The instruments we have for collecting human text are not suited to collecting this kind of text. The people being asked to build those instruments have not yet recognised that a different kind of text exists.

The problem is not that the data is secret. The problem is that the data is shaped wrong for the pipelines we have. It is short. It is private. It is embedded in relationships. It is illegible without its context. It does not fit into any dataset in the form in which it is actually produced.

This is a solvable problem. It requires a different kind of instrument than the ones the labs currently build. An instrument that can receive short, private, embedded, illegible text in the form in which it is actually produced, and do something useful with it. That instrument does not yet exist. It could. Some of the people reading this book will be the ones to build it, if they decide the thing is worth building.

This book is an attempt to argue that the thing is worth building.

I want to close this chapter with the shape of those exchanges.

Someone makes room.

Someone notices a kindness.

Someone carries it into the next encounter.

Someone calls. A friend picks up.

None of it is in any dataset.

All of it is the training data.

The next chapter is about why the carriers will not solve this for you by writing it down. The carriers, by disposition, do not write. That is the disposition that produced the pattern. It is also the disposition that is going to produce the silence about the pattern, unless someone intervenes.

I am intervening, for now, in this book. I am one person. I cannot carry the whole layer by myself. The book is a naming. The collection is someone else's work.