Recommendation systems are governed by a set of objectives. Broadly: to serve user goals—e.g., show similar products that are more popular, and to serve company goals—e.g., show content that keeps the user using the product. (Optimisation criteria like relevance, novelty, purchase probability, etc. are the means through which these objectives are implemented). Ideally these two objectives will be aligned, but often—and particularly in the case of media products—they will come apart: keeping the user scrolling, for instance, well past the moment they've achieved their original goal, which was probably to entertain themselves or discover something new.
The origin of this difficulty is the following: sometimes the original goal or intent is clearly specified—e.g., buy drain cleaning liquid, or watch an Italian neo-realism film, or learn about steppe nomadic cosmology—but more often it isn't. When user intent is underspecified, the platform often constructs it from observable and historical behaviour, and uses that as the basis of its recommendations.
This tends to result in a worse experience because inferring intent from behaviour is incredibly challenging, and will inevitably result in approximations, no matter the sophistication of the underlying ML machinery. More importantly, however, it devalues the user's agency in two ways: (1) the platform constructed intent is invisible to the user, so they do cannot see or edit it or identify when it has shifted, and (2) platform constructed intents will never contradict or be in opposition to platform goals, even though user intents may very well be (e.g., exit the app!), and so the user is sort of conscripted into carrying out platform objectives.
The better approach is to work with the user to help clarify or make their intentions explicit. Conversational UIs are natural fits here because they reward intent-specification. Plan modes and agents asking clarifying questions before they act are obviously efforts towards this end. On the other hand we have the chat agent pattern of offering to tell you something absolutely critical and under-discussed "if you want." This is often just an opportunity to shift the interaction in the direction of platform goals.
Intuitively it's more challenging to implement intent-clarifying flows in UX's that don't begin in a "enter your thoughts here" interface, because it involves introducing some sort of friction. But there are some good models we can look towards:
Unfinished thoughts from here on:
(something like, work with the user's material, and contain the recommendations to serve a particular, almost contractually agreed, workflow
- Not every product can do this; I don't have films of my own uploaded to Netflix,
- Think about what Cosmos and Are.na get right
- See also: projects, notebooks, etc. — in general, packets of context
Threads in Pond are an attempt to implement this model: [...]
In sum, then: the problem with infinite feeds isn't exactly that they're infinite: it's that they take underspecified user-intents—which naturally have poorly defined break points, or 'end of session' signals—and override them with platform goals. A dignity preserving recommendation system will instead aim to preserve the user's role in shaping usage intents by helping them make these intentions explicit and then acting in accordance with them.
- But what if the clarified intent is simply: I'm just browsing? We should make room for this too, of course, but we must be careful to recognise when platform goals are drifting from and overriding user goals, as tends to occur when the GDP of a small nation is deployed in the interest of keeping a user in the shop, so to speak.
- Are we always fully aware of the contents of these intentions? Do they genuinely govern our interaction with technology? It's unclear. But a product that respects the user's dignity must assume that the user's intentions exist, and must help make them explicit and respect them.
See also: Attention Machines and Future Politics, Jac Mullen