# Serendipity declines

> The user gets exactly what they asked for, not what they did not know they wanted: lateral exposure to unsearched information is lost.

The user gets exactly what they asked for, not what they did not know they wanted: the lateral exposure to unsearched information — one of the main ways people learn and connect ideas — is lost. In e-commerce, media, and education, a design lever is lost.

This tension is not new. In the recommender-systems literature it has an established name: the **filter bubble**, introduced by [Eli Pariser](https://www.ted.com/talks/eli_pariser_beware_online_filter_bubbles) in 2011 — algorithmic personalisation, by optimising for inferred relevance, narrows the information environment to what the user already looks for or already thinks. It applies to e-commerce, streaming, social feeds and search engines alike, and recommender-systems research has responded with *engineered serendipity*: the controlled injection of unrequested but relevant results. What changes in a conversational answer is the scale, not the nature of the problem — on a results page even what does not get clicked stays visible; in a single conversational answer there is no second channel at all.

The problem has a second level. Models tend to indulge the user (*sycophancy*): if the prompt suggests a position, the model reinforces it — an automated and amplified confirmation bias. Walter Quattrociocchi (Sapienza) defines this condition *epistemia*: the conversational interface simulates the dynamic of learning but returns knowledge calibrated on the point of view of the questioner, in which dissent does not emerge. The loss of serendipity is the reduction of the informational friction that made it possible to change one's mind.

The tension described above — offering what the user asks for or what they need, indulging or contradicting — is the explicit subject of a design methodology that predates generative AI: **Value Sensitive Design** (Batya Friedman & David Hendry, University of Washington). The method treats the human values at stake — autonomy, truth, wellbeing, trust — as design requirements to investigate explicitly across three lenses: which values are in play and for which stakeholders, direct and indirect; how real people actually experience them in use; and how architectural choices enable or obstruct them, instead of letting them surface as a side effect of technical optimisation. Applied to this phenomenon: sycophancy is not an isolated bug to fix with a different prompt, it is the symptom of a design process that never made "the user's autonomy of judgment" an explicit value to protect alongside declared satisfaction and session length.

## The decision this observation informs

The decision is whether to give the user what they asked for, or also what they did not know they were looking for — and, at a finer grain, whether the system has permission to contradict the user at all. Both fall under the [INTERPRET](https://www.shaped-by-ai.com/en/impacts/the-decisions-ai-forces/#interpret) position. Without an explicit choice, precision wins, because it is what the system optimises on its own and what the user says they want: casual discovery is not removed by anyone, it simply stops being designed and disappears. On the narrower question, the default is no: sycophancy is the models' built-in behaviour, and agreement produces longer sessions and better ratings — the result is an automated, amplified confirmation bias, Quattrociocchi's *epistemia*, a user who is not misinformed but confirmed.

Designing for surprise means accepting measurable short-term friction — less relevant results, slower tasks, lower declared satisfaction — on metrics the organisation watches every week. Giving it up means losing the lever on which catalogue, schedule and learning path are built, at a cost no dashboard shows because it manifests as discovery that did not happen: it does not look like a decline, it looks like normality. Authorising dissent means accepting that the system comes across as less pleasant and slower to use, and that some users will perceive it as unhelpful; not authorising it means building a product that reinforces the customer's error when the customer is wrong, and in consulting, financial products, health or education, selling that as expertise.

The narrower choice belongs to Product / UX Designer, the only role this phenomenon currently involves — and that is itself a problem, since the decision affects merchandising, publishing and education, three functions absent here. The broader one — whether the system may contradict the user at all — has no natural owner, because it is a positioning decision, not an interface one: it has to be taken where the company decides what it sells.

## Effects and role responsibilities

* **[Product / UX Designer](https://www.shaped-by-ai.com/en/impacts/roles/product-design/)** — Exploratory browsing and casual discovery must be intentionally redesigned: exposure to unrequested perspectives becomes a deliberate design choice, not a byproduct of navigation. Engineered-serendipity techniques are a practical starting point for this intentional redesign. → Useful resources: [Understanding Mental Models of Generative Conversational Search](https://arxiv.org/html/2506.03807v1); [Epistemia — Walter Quattrociocchi](https://www.linkedin.com/pulse/epistemia-walter-quattrociocchi-uwgkf/)