You have an idea. You dictate it in three clumsy sentences. A machine turns it into a clear paragraph. You change two words, you cut a sentence, you publish. Is this text yours? Our instinct hesitates, and that hesitation is interesting: it reveals that we have never defined what it means for a text to belong to someone.
We covered the legal question in our article on ownership of generated content. Here is the prior question, which has no legal answer.
The three criteria we mix up
When we say a text is someone's, we are actually invoking three different things, and they do not always coincide.
Execution. They are the one who typed the words. This is the most superficial criterion, and the only one AI clearly removes. No one considers that a text dictated to a secretary ceases to belong to the person who dictated it.
Intention. They are the one who decided what the text should say, to whom, and why. This is a more solid criterion, and it withstands assistance: the machine does not decide what you want to express.
Responsibility. They are the one who signs, who stands behind it, who will answer if the text is false or hurtful. This is the most demanding criterion, and the most interesting.
Our discomfort with assisted text comes from the fact that these three criteria have come apart. Previously, the one who executed was generally the one who decided and the one who took responsibility. The question did not arise.
What is particular about style
There is, however, a serious objection to the idea that only intention matters, and it must be taken seriously.
In writing, form is not a dressing for thought. It is part of it. A writer does not know exactly what they want to say before having written it: the wording refines the thought, forces it to choose, reveals a contradiction one had not seen.
This is what we noted in our article on time saved: certain efforts produce the result and transform the one who performs them. Delegating the shaping can therefore amount to delegating part of the thinking itself, without realising it.
A text for which you did not search for the words is not necessarily less true. It is perhaps less thought through, because the thinking stopped at intention instead of going all the way to formulation.
One question is enough to settle most cases: could you defend every sentence? If you can explain why this one rather than another, what it claims and why you stand by it, the text is yours no matter how it was produced. If you discover its content upon rereading, it is not, even if you typed every word. This is also the criterion we proposed for school assignments: the right question is not how you did it, but can you explain what you handed in.
What this changes depending on context
The answer is not the same everywhere, and that is normal.
For a professional document, the question of authenticity is secondary. What matters is accuracy and usefulness. No one asks who drafted a manual.
For a work of art, it is central. We read an author for their perspective, and a text whose form they did not seek no longer quite offers that perspective. This is what we explored regarding art in the age of AI.
For a personal message, it is decisive and of a different nature. What a loved one expects from a letter is not that it be well written, but that you took the time. A condolence perfectly turned by a machine misses exactly what it claims to offer, because the value of the gesture lay in the cost of the gesture.
What to take away
There is no percentage of intervention that would decide it. The question is not quantitative; it concerns what you are able to stand behind.
What this technology forces us to clarify, at bottom, is a distinction we never needed to make before: between having produced a text and answering for it. The former has become easy. The latter is no easier than before, and it is probably what defined authorship all along.