I kept asking AI to make my writing sound more human.
The first version would be competent but strangely anonymous, so I would ask for more warmth, less formality and fewer polished transitions. The next version would use contractions, soften its claims and add something resembling personality, yet it still did not sound like me.
The problem was not that the model had failed to learn human language. It had been trained on an enormous record of human expression and could reproduce the vocabulary of a novelist, the structure of a strategy memo or the tone of a supportive friend.
What it did not have was access to the life behind any particular voice.
It did not know the memories that made one reference meaningful, the mistakes that created a hesitation, the relationships that changed a person’s vocabulary or the thousands of small decisions through which a style became recognisable. It could imitate the visible result, but it could not reconstruct the continuous experience that produced it.
That limitation may point towards one of the largest commercial opportunities in AI: giving people ownership of the personal context that models cannot create for themselves, then allowing that context to be used across applications, shared with others or licensed for particular purposes.
The data that makes a person specific
AI systems are built from fragments of human activity: books, photographs, conversations, articles, code and other records of what people have produced. By identifying patterns across those fragments, a model can generate convincing approximations of human communication.
But an approximation assembled from millions of people is different from an understanding developed through sustained access to one person.
A person’s identity is not contained in their published work or their preferred sentence length. It includes the experiences they remember, the ideas they abandoned, the advice they ignored, the people who influenced them and the contradictions they have never resolved. It also includes choices that appear insignificant in isolation but become revealing when observed over time.
The paragraph someone deletes says something about them. So does the version they select, the suggestion they ignore and the sentence they rewrite manually. A prompt records what they believed they wanted, while a correction often reveals what they actually preferred.
Most AI products treat these interactions as temporary instructions rather than as evidence of judgement. The result is a system that can learn almost anything about the world while repeatedly forgetting the person using it.
From a researcher’s work to a researcher’s context
Suppose you wanted an AI to approach a problem like a particular researcher. You could give it their papers, books, lectures and interviews, and the model could retrieve their arguments, reproduce their terminology and answer questions using the concepts most closely associated with their work.
That would give you access to their output, but not necessarily to the way they think.
Published work shows what survived the research process. It rarely reveals every source the researcher distrusted, every hypothesis they abandoned, every connection they noticed too early to prove or every question they decided was not worth pursuing. Their value lies partly in what they know, but also in how they decide what deserves attention.
A richer context could include annotations, rejected ideas, preferred sources, unpublished questions, changes between drafts and examples of how the researcher responds when the evidence is incomplete. The goal would not be to claim that the resulting AI was the researcher, but to allow their methods and judgement to shape how a general model approaches a new problem.
That context could remain private, be shared with students or be licensed to an institution for a defined purpose. The researcher would not simply publish more content; they would make a carefully controlled version of their intellectual context available as a new kind of product.
The same model applies elsewhere. A designer could make the references and decisions behind their visual taste available to a creative system. A founder could preserve the reasoning behind years of company decisions. A specialist could package professional judgement that would otherwise disappear when they retired or changed roles.
A public figure could create an authorised AI experience based on information they had deliberately supplied, rather than allowing their identity to be approximated from interviews, public appearances and material collected without their involvement.
In each case, the model provides the general capability, while the person provides the context that makes the capability distinctive.
A new form of commercial identity
People already commercialise their knowledge through books, courses, consulting, memberships and access to their time. Those formats are useful, but they are fixed. A book answers the questions its author anticipated, while a course follows a sequence prepared in advance.
Context can respond to questions that were never anticipated. It can reorganise itself around a new problem, connect information across an archive and apply a person’s standards or preferences in a situation they have not encountered before.
This does not mean selling access to an entire person. A useful context system would need to be selective by design.
A researcher might share their published annotations and methods while excluding personal correspondence. A public figure might provide one context for an educational experience and another for commercial collaborators. A consultant might license their professional framework without exposing the clients, relationships or private experiences that helped form it.
The commercial unit would therefore not be the person themselves, but a specific, permissioned context they had chosen to create.
This distinction separates the idea from impersonation. A digital replica attempts to convince the user that a simulated person is present. A context layer allows someone’s knowledge, judgement and perspective to influence the work while remaining clear that the underlying response is generated by a model.
The opportunity is not to manufacture artificial people. It is to give real people more control over the value created from what they know and how they think.
The ownership problem
The same idea creates an obvious risk.
The more personal context an AI company accumulates, the harder it becomes for the user to leave. A system that remembers someone’s work, preferences, relationships and history may become more useful over time, but that usefulness can also become a form of dependence if the accumulated context cannot move elsewhere.
Personalisation without portability is lock-in.
There is also a more fundamental question about consent. Personal experience is not ordinary product data. Memories, correspondence, private decisions and patterns of behaviour may reveal more about a person than they intended to share, particularly when separate fragments are combined.
A credible context layer would therefore need clear boundaries around ownership, access, duration and purpose. People would need to know what had been retained, remove information, revoke access and decide whether a context could be used only for retrieval, for generation or for further training.
Without those controls, the commercial opportunity becomes another system for extracting value from people while calling the result convenience.
With them, the balance of power changes. The user can move between models while retaining the context that makes those models useful, and the people whose knowledge or identity creates value can decide when, where and how that value is used.
What becomes valuable when models converge
Models will continue to improve, and the companies building them will remain important. But as capable intelligence becomes cheaper and more widely available, the differences between models may matter less for many everyday tasks.
The harder asset to reproduce will be the accumulated context around the model: the person’s memories, preferences, knowledge, corrections and history of decisions.
I began thinking about this while building Tansei, a visual workspace for the fragments behind creative work. I had noticed that my process no longer lived inside a single document. It lived between references, copied passages, screenshots, unfinished ideas and the changes I made while deciding what belonged together.
Those fragments were not clutter surrounding the final work. They contained evidence of the judgement that produced it.
As AI makes finished outputs easier to generate, the context behind them becomes more important. The output may be reproducible, but the lifetime of experiences, memories and decisions that shaped it is not.
The AI industry has spent years trying to place more human knowledge inside models. Its larger opportunity may be giving people the ability to own, use and commercialise the personal context that no model can create for itself.
If a lifetime of human context becomes AI’s most valuable input, who should own it?
Back to top