“So, Tell Me What Users Want, What They Really, Really Want!”

The question of how to bridge the potential gap between what users actually do and what they ‘really wanted’ to do has a relatively long history in human-computer interaction research. In the 1960’s, Warren Teitelman’s ‘Do What I Mean’ (DWIM) philosophy argued that systems should not just execute whatever potentially erroneous instructions users put into a terminal  [31]. Instead, they should try to interpret users’ true intentions and correct their errors (the implication being DWIM, Now What I Say (or Do)). In practice, Teitelman’s error-correction systems were critiqued as merely reflecting what their designer would have meant (‘do what Teitelman means’) [30].

The issue crops up in more fundamental ways in the domains of decision support and recommender systems, where the gap is not just between what the user typed and what they really intended, but between recorded interaction behaviour and what can be inferred about the user’s wants and needs. On those rare occasions where HCI researchers in such fields venture into the moral minefield of defining ’what users really want’, they often provide definitions which are intuitively reasonable yet dis-satisfyingly cursory; the related philosophical debate is then swiftly swept under the carpet.

For instance, in an otherwise enlightening chapter, Jameson et al. offer the following on what decision-support systems should optimise for: “… a ‘good outcome’ [is] one that the chooser is (or would be) satisfied with in retrospect, after having acquired the most relevant knowledge and experience. Admittedly, this assumption is subject to debate…”  [15, p. 35].

Similarly, Pommeranz et al. write in relation to methods for eliciting user preferences: “More research is needed to design preference elicitation interfaces that elicit correct preference information from the user” [24, p. 361]. Later, they consider what normative aspects might be required to determine such ‘correctness’ of a preference and find answers to this question in short supply: “There is much room for more explicit consideration of human preference construction also including values and affective aspects” [24, p. 365].

Parallel questions are arising in recent AI research. For instance, techniques like ‘inverse reinforcement learning’ (IRL) attempt to infer an underlying goal function from behavioural output [22]. It is implied that such goal functions are equivalent to the ‘true desires’ of the human from whom the machine is learning. Indeed, as Stuart Russel put it, a well-aligned AI ‘will watch all of us to learn more about what it is that we really want”.¹¹1Talk at TED2017 https://www.ted.com/talks/stuart_russell_how_ai_might_make_us_better_people/transcript A common assumption in these techniques is that humans make optimal decisions, with deviations from optimality reflecting ‘random noise’ in action selection [12]. However, real human decision-making deviates systematically from optimality, because of cognitive biases like asymmetric perception of losses and gains, which make people sensitive to how identical outcomes are framed, or hyperbolic discounting of future rewards, which often make people inconsistent over time [32, 16]. A rational model may wrongly assume that the true preference of, for example, a smoker who is trying (and failing) to quit is to smoke [12].

Indeed, some recent work in this space acknowledges the limitations of the pure behaviourist dream. Armstrong et al. argue that current IRL methods are “fundamentally and philosophically incapable of establishing a ‘reasonable’ reward for the human”, which can only be overcome, they argue, by building in “normative assumptions about the reward and/or planner”  [3, p. 2]. However, the authors stop short of articulating what those normative assumptions might be.

An alternative approach is to devolve responsibility for those normative assumptions back to the user. Numerous ways of explicitly eliciting users’ preferences have been explored in the literature, including user ratings and example-critiquing in recommender systems [25, 24] and absolute measurement and pairwise comparison in decision support systems [2, 7]. Moreover, pioneering work in this space explored ways to let users inspect and tweak a system’s model of them [8, 19, 25, 18].

However, explicit methods also run into problems. Foremost, the elicitation process itself greatly influences what users say they want. For example, users may prefer different options based on whether they are framed as losses or gains [24], or depending on the moment in time in which they are asked. What point in time reflects what the user ’really’ wants - the most recent, a weighted average over the last day / week / year, or something else? [17, 26].

Related to the question of timeframes is the tension between broad or narrow construals of the user’s context. One of the foundational tenets of user-centred design, as articulated by Ritter et al., is to consider the user context more broadly  [27]. That is, moving beyond the immediate, task-related issues pertaining to a specific product, where the user’s goals can be more easily approximated, and instead view applications of technology as "the development of permanent support systems and not one-off products that are complete once implemented and deployed"  [27, p. 44]. In other words, the designer should consider longer-term effects of systems on people’s lives, which in turn requires deeper insight in order to align systems with users’ more general goals, values, and life situation.

The question of how to elicit ’correct’ preferences therefore ends up connecting with the philosophical debate on the fundamental constituents of ‘what makes someone’s life go best’. While it’s possible that one’s best life could be at odds with one’s desires, it is often held that being able to satisfy ‘true desires’ are at least partly constitutive of the good life. Most philosophers’ answers tend to involve one or more of the following: pleasurable experiences or ‘hedonism’ (e.g. [5]), where a good life is one full of pleasurable experiences; desire-satisfaction (e.g. [23]), where a good life is one in which one’s desires are fulfilled; and objective list theories, where a good life is one in which certain objectively worthwhile things are experienced, achieved or engaged in (e.g.  [14]). Hedonists often concede that certain hedonic pleasures are more truly constitutive of a good life, and proponents of desire-fulfillment acknowledge that the fulfillment of misguided desires contributes little to one’s best interests. As we hope to show, these philosophical debates underlie some of the conceptual difficulties encountered by the more practically-oriented research within the HCI / AI research cited above.

To illustrate this, we now present an alternative yet very real take on the issue, via a fictional dialogue between senior executives in a global tech company whose product aims to help people live their best possible life. How can they settle on the metric their product should optimise so that they give their users what they really want? In addition to the literature reviewed so far, the positions of our fictive designers are inspired by common views within classical economics and rational actor models [20], behavioural economics [32, 16], and philosophical and psychological work on the ’good life’ [28, 9], as well as common ideological stances. We acknowledge that this is only a portion of the relevant literature, and we have taken artistic license in our formulations of each stance. As such, we cannot claim our fictive characters to be fully and fairly representative of the possible range of positions. Furthermore, any resemblance to real persons is probably not accidental.

SCENE: The Californian morning sun shines through the gleaming windows into a pristine meeting room at the headquarters of Gamaface, a global technology company. Gamaface is the industry leader in what they call ’algorithmic life services’. Their augmented reality platform - which combines decision support, persuasive computing, and ubiquitous personalised nudging - is used by 3 billion users 24 hours a day, 7 days a week.

CAST:

• •

  Sunny Zuckerbezos, CEO of Gamaface
  @Walkin_on_sunshine | Mission: connecting people and data to make life worth living.

• •

  Randy Na, Information Architect
  @Nozick_SoSick | Libertarian | Autono-me, autono-you | "I hold it to be the inalienable right of anybody to go to hell in his own way."

• •

  Harald Richter, User Researcher
  @WinkWinkNudgeNudge | Pavlov’s dog, striving to be Pavlov’s bell. #cognitivebias #positivepsychology

• •

  Nichola Machian, Lead Ethicist
  @eudaimonia_for_all | "Meaning arises when subjective attraction meets objective attractiveness"

Given the problems with equating users’ true needs and desires with simple behavioural measures of ‘engagement’, alternative metrics are needed. In this paper, we have discussed this issue in the context of a dialogue between designers in a fictive tech company searching for a metric with which to measure whether they are giving users what they ‘really’ want.

Despite the often glib treatment of these questions in HCI research, as seen in the examples in the introduction, there are some notable exceptions. The general critique of the behaviourist tendency in recommender systems is well-articulated by Ekstrand & Willemsen [11]. They suggest one promising corrective, which allows users to choose between different algorithms underlying their recommender systems [10]. In allowing some direct control of the inference process, this is reminiscent of earlier work which aimed to create ‘scrutable’ user models which are transparent and configurable  [18, 19].

Harald’s call for systems which help put users in a position from which they can reflect on themselves and their desires, has some precedent in Slovak et al.’s proposal to design for ‘reflective practicum’; a state in which someone can engage in transformative revision of their outlook or behaviour  [29]. In addition, we suggest that empirical findings from positive psychology and reflections from the philosophy of what makes life go well will prove important as a source of more opinionated takes on how we might design to support people’s ‘better selves’ and draw on life wisdom that does not originate with the designer or individual user [28, 9, 6].

Nichola’s call to consider the meaningfulness of user experiences is something explored by Mekler and Hornbæk, who argue that UX designers should consider eudaimonic (‘living the good and virtuous life’) as well as hedonic (‘pleasurable’) experiences [21]. Similarly, Zimmerman proposes ‘designing for the self’, such that products can help their users become the people they desire to be [33].

For those sympathetic to Randy’s position, the notion of an external force guiding users’ search for the good life may detract from the sense in which an individual ought to be responsible for their own life journey. However, behavioural economists have found that the crucial question for whether people are in favour of ‘nudging’ –in the sense of setting up the choice environments to support particular kinds of behaviour– is whether or not they agree with the vision of the good life behind those nudges [13]. We can easily imagine a future in which users expect the right to choose which normative assumptions should be embodied in the algorithms feeding their recommender systems.

A central goal of this paper has been to show that whereas it is easy to criticise tech companies for equating users’ true preferences with simple metrics of engagement, it is not obvious what good alternative metrics look like. Every metric implicitly embodies particular assumptions about human nature - and ultimately about the good life - with few decisive arguments for favoring one as the ’best’. It is a fallacy, however, to conclude that the question is therefore not worth bothering with and/or that all metrics are equally valid. Some ways of solving the problem are clearly worse - such as simply equating true preference with number of clicks or time spent using a service - even if there is no way to tell which of the better alternatives might be the global optimum. We believe that allowing users the to choose between different ways for a system to infer their preferences - for example, implementations of Randy’s, Harald’s, and Nichola’s positions - would be a big step forward. To reach such a future, we need to properly engage with the issue and explore, build, evaluate, and discuss what good alternatives look like.

Some readers may still feel that this discussion is akin to moral philosophers’ "trolley problem" - important in principle, but not terribly relevant in practice [4]. However, even though recommender and decision-support systems typically used today are limited in scope to specific task and interest contexts, more sophisticated systems could begin to model their longer term effects on user’s lives, and user’s preferences towards such influence. In fact, such a future might rapidly be approaching. Addressing recent criticisms of Facebook, CEO Mark Zuckerberg recently announced that his personal challenge for 2018 included "making sure that time spent on Facebook is time well spent" ²²2Facebook post on 4th January 2018 https://www.facebook.com/zuck?hc_ref=ARQqfRj278TDWekby2TLyI0A0meA4
-4PxqohaalwAfzCeAsMaft16fKBkDYiHEg4cQk&fref=nf. Following up a week later, Zuckerberg elaborated that his team felt "a responsibility to make sure our services aren’t just fun to use, but also good for people’s well-being" and that he would be "changing the goal [of] our product teams from focusing on helping you find relevant content to helping you have more meaningful social interactions"³³3Facebook post on 11th January 2018 https://www.facebook.com/zuck?hc_ref=ARTYPggwbi_dcIl5f8M8r1dGZYZGhpWmAXwj_9C6g6mmSCSxA0dpxqWqEaPojN1IWD0&fref=nf. Time will tell which, if any, of the paths suggested by Randy, Harald, and Nichola will be taken by world-leading social media platforms.