Sub-shift
“THE ASSIGNED SELF”
A system’s description of a person can start changing the choices they see.
Research a subject for someone else and a service may start treating it as your own interest. Explore an unfamiliar way of living and the next recommendation may assume you have already chosen it. These small mistakes reveal a larger problem: behaviour supplies clues about a person, not a complete account of them.
Over the next one to three years, we expect some consumers to encounter those interpretations more directly. An assistant might describe their priorities, temperament or likely future, rather than merely serve a related item. The description could become something they accept, argue with or avoid.
An inferred description has consequences beyond an occasionally irrelevant recommendation if it begins shaping how people are treated—or how they understand their own possibilities.
A guess can acquire a history
John Cheney-Lippold's work on algorithmic identity describes categories assembled from behaviour rather than simply copied from a person's declaration. His 2011 account predates today's assistants, but the distinction becomes more important when the input includes personal conversation. [62]
A question about a condition is not a diagnosis. Helping a relative prepare for retirement does not say how soon the person asking intends to retire. A system can make a persuasive interpretation from these details and still have the wrong subject, purpose or moment in someone's life.
Repeated use could give that interpretation a longer reach. If future responses draw on an early assumption, the person may keep seeing options suited to a role they never chose. They might eventually adopt the description, repeatedly correct it or stop using the service. Those are possible consumer responses, not an established population-wide pattern.
Recognition that can be revised
For businesses, useful personalization requires knowing when a description has stopped helping. A customer exploring a new field may want encouragement as a beginner rather than recommendations based entirely on past expertise. Someone changing a routine may need the old pattern to lose its influence.
The stakes vary. A misplaced entertainment suggestion is different from an assumption that affects access, price or consequential treatment. Ruha Benjamin's critique of discriminatory design is relevant here: apparently neutral technical categories can reproduce existing inequalities. More data and greater confidence do not, by themselves, resolve that problem. [65]
You could make a useful distinction between what customers told you, what your service inferred and what they are merely considering. Where an assumption matters, let them correct it in a way that changes the subsequent experience. Avoid asking for an ever more intimate history just to disprove an unnecessary label.
Not everyone wants to maintain a profile. The opportunity may be a service that asks one well-timed question, remembers the answer appropriately and lets a person move on. Customers should not have to keep explaining who they are not.
THE SOURCE NOTESFollow the thinking 2 sources · what they show and where their limits lie
Sources and further reading
These sources inform the article. A product announcement describes an offer, not how widely people use it. Arguments and imagined futures are kept separate from observed results.
Behavioral categorization and its consequences. Written before conversational AI; applying it to assistants is our interpretation.
Explains discriminatory design and apparent neutrality. Used as a conceptual challenge, not a current population estimate.
About this research edition
This article draws on selected sources and our interpretation of their implications for consumers. The wider collection contains 50 articles and original illustrations. We have not read everything by everyone on the reading list, calculated the probability of each future or established worldwide adoption. The collection is not exhaustive.
The illustrations are AI-generated, with a visual approach informed by TrendWatching’s 2024 and 2025 reports. They depict ideas rather than observed events. If an article’s central idea changes, its image needs another look. Small edits need not mean a new illustration; automatic replacement is not active.
Each innovation has a source, a review date and an explanation of its relevance and limits. Some are early signs of the change; others offer a useful comparison. None proves the forecast. Monthly monitoring and automatic selection of new examples are not running.