Memory and context adaptation
GenAI remember the conversation, calibrate the tone, adapt to the style. From this comes the expectation that every system recognizes the user, does not start from zero, and adapts in real time to location, moment, preferences, skill level: the model called experience-led design.
However, a study of 937 ChatGPT users (Scientific Reports, March 2026) reports that greater adaptability does not automatically mean a better experience. It was found that, in some conditions of strong model adaptation, users report more negative perceptions — excessive adaptability can feel unsettling or manipulative. Memory and personalization have an optimum point, beyond which the system stops seeming attentive and starts seeming intrusive.
The decision this observation informs
The decision is how much the system should adapt to the user, and where the upper limit sits. It falls under the ASSIST position. Without an explicit choice, it gets pushed to the maximum, because every increment of personalisation is measurable as a local improvement and nobody measures the optimum point: the study on 937 users cited above shows that beyond that point users report more negative perceptions, and excessive adaptivity comes across as unsettling or manipulative. An upper limit nobody sets does not exist.
Setting a ceiling means giving up demonstrable performance in the A/B test of the individual feature, and defending in a meeting a no that the local data contradicts. Not setting it means crossing the optimum point without noticing, because the signal is not a drop in conversion but a shift in perception — and because, as the Personalization / CRO Specialist entry below puts it, “you do not A/B test infinite variants, you test generation policies”: without a declared policy there is not even an object to test. The choice belongs to Product / UX Designer for the perceived ceiling and to CDP / Martech Ops for the data ceiling, two distinct decisions to be made together: the CDP / Martech Ops entry already names the two tensions to design for, consent and minimisation against the expectation that the system already knows who you are.
Effects and role responsibilities
- Product / UX Designer — A system that remembers the user is a baseline expectation, not a premium feature: the “new user”/“returning user” states must be rethought with more granularity. The interface can compose itself in response to intent (a chart for an analysis, a table for a comparison, suggestions for those exploring): systems more complex than a static interface, which risk promising more than the technology delivers. Design also the upper limit of adaptability, not just the lower one. → Useful resources: AI Design — NN/g; Just Say It Simply: How Prompting Shapes AI Responses for Older Adults
- CRM / Email & Lifecycle — Communications identical for everyone lose effectiveness compared to a personalized conversational benchmark. The lifecycle must be rethought assuming that a growing share of messages is read, summarized, and filtered by an AI before the user (see The customer journey shortens).
- CDP / Martech Ops — Experience-led design has an infrastructural prerequisite: a unified, real-time user profile, activatable from every touchpoint. The CDP becomes the substrate of the whole product’s adaptability; the market moves in this direction with the “agentic CDP”. Two tensions to design: consent and data minimization against the expectation that “the system already knows who you are”; growing value of zero-party data, because what the user declares while conversing is richer than any behavioral inference. → Useful resource: Ready or not: the era of the agentic CDP — MarTech
- Personalization / CRO Specialist — Generative personalization (content composed on the fly for the single user) breaks the QA and experimentation model: there is no longer a finite number of variants to approve and test. You need brand guardrails, sample QA on output, and new methodologies: you do not A/B test infinite variants, you test generation policies.
- Front-end Developer — You need architectures for dynamic and contextual components, with non-negligible costs: shared design-development evaluation of what is desirable, what is feasible, what is feasible now.