One of the things I find most interesting about personalization in AI products is how much a person’s standard only becomes visible once the system has failed to meet it. Someone can describe their role, their goals, and the kind of output they want, and still find themselves making corrections they would never have thought to specify beforehand. Seeing an answer gives them something to respond to. It makes an expectation concrete enough to recognize.
A person might ask for a concise recommendation and then restore a qualification the system removed. They might reject a chart because it directs attention toward the wrong comparison, or rewrite a paragraph that is accurate but gives the wrong impression of the work. These are all moments in which the product could learn something useful. What makes them interesting is that the user is expressing a standard through the work itself, rather than stopping to describe a general preference.
Onboarding can provide a starting point for that understanding. It can establish context and save someone from repeating information the product reasonably needs. But I think there is a limit to treating personalization as something that can be completed by asking better questions at the beginning. Many professional standards depend on the audience, the stakes, and what someone is trying to accomplish. The user may understand those distinctions perfectly well in practice without having a ready explanation for every circumstance in which they apply.
Being concise is a good example. For one piece of work, it might mean removing repetition. For another, it might mean leading with the decision while keeping the reasoning available underneath it. A qualification that feels unnecessary in an internal note may be essential in a recommendation someone else will act on. Learning that a user prefers brevity does not establish which parts of their thinking can safely be compressed.
The TAHI study offers a useful way into this question. Across 600 writing and visualization tasks involving 30 professionals, adapted agents improved solo task success by 4.5–20.9%. Direct edits, plan changes, inline comments, and revised criteria provided useful learning signals beyond text feedback alone. Yet the agents captured only around 42–43% of the expertise expressed in contextual judgment and problem framing. Explicit conventions were easier to learn than the situational understanding that makes a convention appropriate.
What interests me about that gap is the difference between noticing a correction and understanding what it means. A system can observe that someone removed a recommendation without knowing whether they disagreed with it, thought the evidence was insufficient, or considered it wrong for that particular audience. The action is visible. The reason may still need to be understood. Treating every intervention as a preference risks making the product more consistent in an interpretation the user never intended.
That becomes more consequential when the interpretation persists. If removing a recommendation teaches the system to stop suggesting it, the person may have fewer opportunities to encounter it again and clarify what they meant. A lesson drawn from one situation can quietly influence another. Over time, the product could feel increasingly familiar while becoming less responsive to the circumstances in which the person is actually working.
I think this is where personalization becomes a question of authority as well as learning. The system is forming an understanding of someone’s standards, and that understanding influences what it presents, omits, or chooses to do. The person needs a way to see the consequential parts of that interpretation and change them. They also need to be able to say that something applies to this project, or this audience, without accepting it as a permanent description of how they work.
A second paper, Transfiver, proposes a shared editable state that both the human and the AI can update. Its central requirement is that the information a person inspects and edits is the information subsequent computation actually uses. The work includes a small implementation, with richer natural-language applications at scale still open. As a product question, though, the connection is immediate: a correction needs to change future behavior, including after the current conversation has ended.
Making remembered information visible is useful, but visibility alone does not establish that kind of control. Someone might be able to change an entry and still encounter the old assumption later. They might be able to remove a preference without knowing which other decisions depended on it. What matters is whether the product gives the person an effective way to revise its understanding, limit where a lesson applies, and reverse a change that turned out to be wrong.
This also changes how I would evaluate an AI product. The first interaction can show whether the system understands an initial request. Repeated collaboration shows whether it learns the right things from correction and whether those lessons survive changes in context. I would want to see what happens when a user makes an exception, corrects an inference, moves to another project, and returns later. A product can respond well to each individual instruction while still carrying an inaccurate understanding across the sequence.
For Studio Marelle, that makes behavioral learning an important part of thinking about AI adoption. The question is what the product learns from use, how it interprets that evidence, and how much influence the person retains over the interpretation. Those things affect whether repeated collaboration reduces the burden of explaining oneself or creates a new burden of managing what the system has misunderstood.
A product may remember a great deal about someone before it understands what makes its work useful to them. I think meaningful personalization depends on whether that understanding can become more precise through collaboration while remaining open to revision. People need room to make exceptions, change their minds, and recognize standards they could not articulate at the beginning. The product’s understanding of them needs room to change with it.
